From 0cf7794b16fab8d4561bc5f6379f6d48bd59e101 Mon Sep 17 00:00:00 2001
From: Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Date: Wed, 3 Dec 2025 19:43:29 -0800
Subject: [PATCH] ggml update to b7108 (#12992)

* Revert "vulkan: temporary cary of vulkan fixes (#12971)"

This reverts commit 3a9e8e9fd42f32711b8aeea355e3ed5e155d49b2.

* ggml update to b7087

* fix argsort on metal

* update to b7108

* fix bakllava regression

This model lacks the metadata for the projector type.

* update to b7209

* fix TopK perf

* only build arm code on arm
---
 Makefile.sync                                 |     2 +-
 llama/build-info.cpp                          |     2 +-
 llama/llama.cpp/.rsync-filter                 |     3 +
 llama/llama.cpp/common/common.cpp             |   103 +-
 llama/llama.cpp/common/common.h               |    51 +-
 .../common/json-schema-to-grammar.cpp         |    28 +-
 .../llama.cpp/common/json-schema-to-grammar.h |     2 +
 llama/llama.cpp/common/log.cpp                |     6 +
 llama/llama.cpp/common/log.h                  |     2 +
 llama/llama.cpp/common/sampling.cpp           |    68 +-
 llama/llama.cpp/include/llama.h               |    28 +-
 llama/llama.cpp/src/llama-arch.cpp            |   251 +-
 llama/llama.cpp/src/llama-arch.h              |    28 +
 llama/llama.cpp/src/llama-batch.cpp           |    96 +-
 llama/llama.cpp/src/llama-batch.h             |    13 +-
 llama/llama.cpp/src/llama-chat.cpp            |    32 +
 llama/llama.cpp/src/llama-chat.h              |     1 +
 llama/llama.cpp/src/llama-context.cpp         |    70 +-
 llama/llama.cpp/src/llama-context.h           |    10 +-
 llama/llama.cpp/src/llama-cparams.h           |     1 +
 llama/llama.cpp/src/llama-grammar.cpp         |    26 +-
 llama/llama.cpp/src/llama-graph.cpp           |    28 +-
 llama/llama.cpp/src/llama-hparams.cpp         |    12 +-
 llama/llama.cpp/src/llama-hparams.h           |     8 +-
 llama/llama.cpp/src/llama-impl.cpp            |     6 +-
 llama/llama.cpp/src/llama-kv-cache-iswa.cpp   |     4 +-
 llama/llama.cpp/src/llama-kv-cache.cpp        |    77 +-
 llama/llama.cpp/src/llama-kv-cache.h          |     6 +-
 llama/llama.cpp/src/llama-kv-cells.h          |    46 +-
 .../llama.cpp/src/llama-memory-recurrent.cpp  |    39 +-
 llama/llama.cpp/src/llama-memory-recurrent.h  |     4 +-
 llama/llama.cpp/src/llama-model.cpp           | 13733 +---------------
 llama/llama.cpp/src/llama-model.h             |    17 +-
 llama/llama.cpp/src/llama-quant.cpp           |    20 +-
 llama/llama.cpp/src/llama-sampling.cpp        |    24 +-
 llama/llama.cpp/src/llama-vocab.cpp           |    23 +-
 llama/llama.cpp/src/llama-vocab.h             |     2 +
 llama/llama.cpp/src/llama.go                  |     6 +-
 llama/llama.cpp/src/models/afmoe.cpp          |   187 +
 llama/llama.cpp/src/models/apertus.cpp        |   125 +
 llama/llama.cpp/src/models/arcee.cpp          |   135 +
 llama/llama.cpp/src/models/arctic.cpp         |   138 +
 llama/llama.cpp/src/models/arwkv7.cpp         |    86 +
 llama/llama.cpp/src/models/baichuan.cpp       |   122 +
 llama/llama.cpp/src/models/bailingmoe.cpp     |   144 +
 llama/llama.cpp/src/models/bailingmoe2.cpp    |   135 +
 llama/llama.cpp/src/models/bert.cpp           |   176 +
 llama/llama.cpp/src/models/bitnet.cpp         |   160 +
 llama/llama.cpp/src/models/bloom.cpp          |   101 +
 llama/llama.cpp/src/models/chameleon.cpp      |   178 +
 llama/llama.cpp/src/models/chatglm.cpp        |   132 +
 llama/llama.cpp/src/models/codeshell.cpp      |   111 +
 llama/llama.cpp/src/models/cogvlm.cpp         |   100 +
 llama/llama.cpp/src/models/cohere2-iswa.cpp   |   131 +
 llama/llama.cpp/src/models/command-r.cpp      |   122 +
 llama/llama.cpp/src/models/dbrx.cpp           |   123 +
 llama/llama.cpp/src/models/deci.cpp           |   135 +
 llama/llama.cpp/src/models/deepseek.cpp       |   144 +
 llama/llama.cpp/src/models/deepseek2.cpp      |   237 +
 llama/llama.cpp/src/models/dots1.cpp          |   134 +
 llama/llama.cpp/src/models/dream.cpp          |   105 +
 llama/llama.cpp/src/models/ernie4-5-moe.cpp   |   150 +
 llama/llama.cpp/src/models/ernie4-5.cpp       |   110 +
 llama/llama.cpp/src/models/exaone.cpp         |   114 +
 llama/llama.cpp/src/models/exaone4.cpp        |   123 +
 llama/llama.cpp/src/models/falcon-h1.cpp      |   113 +
 llama/llama.cpp/src/models/falcon.cpp         |   120 +
 .../llama.cpp/src/models/gemma-embedding.cpp  |   120 +
 llama/llama.cpp/src/models/gemma.cpp          |   112 +
 llama/llama.cpp/src/models/gemma2-iswa.cpp    |   125 +
 llama/llama.cpp/src/models/gemma3-iswa.cpp    |   131 +
 llama/llama.cpp/src/models/gemma3n-iswa.cpp   |   377 +
 llama/llama.cpp/src/models/glm4-moe.cpp       |   153 +
 llama/llama.cpp/src/models/glm4.cpp           |   127 +
 llama/llama.cpp/src/models/gpt2.cpp           |   105 +
 llama/llama.cpp/src/models/gptneox.cpp        |   144 +
 llama/llama.cpp/src/models/granite-hybrid.cpp |   196 +
 llama/llama.cpp/src/models/granite.cpp        |   211 +
 .../src/models/graph-context-mamba.cpp        |   283 +
 llama/llama.cpp/src/models/grok.cpp           |   159 +
 llama/llama.cpp/src/models/grovemoe.cpp       |   141 +
 llama/llama.cpp/src/models/hunyuan-dense.cpp  |   132 +
 llama/llama.cpp/src/models/hunyuan-moe.cpp    |   154 +
 llama/llama.cpp/src/models/internlm2.cpp      |   120 +
 llama/llama.cpp/src/models/jais.cpp           |    86 +
 llama/llama.cpp/src/models/jamba.cpp          |   106 +
 llama/llama.cpp/src/models/lfm2.cpp           |   175 +
 llama/llama.cpp/src/models/llada-moe.cpp      |   122 +
 llama/llama.cpp/src/models/llada.cpp          |    99 +
 llama/llama.cpp/src/models/llama-iswa.cpp     |   174 +
 llama/llama.cpp/src/models/llama.cpp          |   155 +
 llama/llama.cpp/src/models/mamba.cpp          |    55 +
 llama/llama.cpp/src/models/minicpm3.cpp       |   199 +
 llama/llama.cpp/src/models/minimax-m2.cpp     |   124 +
 llama/llama.cpp/src/models/models.go          |     6 +
 llama/llama.cpp/src/models/models.h           |   544 +
 llama/llama.cpp/src/models/mpt.cpp            |   126 +
 llama/llama.cpp/src/models/nemotron-h.cpp     |   121 +
 llama/llama.cpp/src/models/nemotron.cpp       |   122 +
 llama/llama.cpp/src/models/neo-bert.cpp       |   104 +
 llama/llama.cpp/src/models/olmo.cpp           |   121 +
 llama/llama.cpp/src/models/olmo2.cpp          |   150 +
 llama/llama.cpp/src/models/olmoe.cpp          |   124 +
 .../llama.cpp/src/models/openai-moe-iswa.cpp  |   124 +
 llama/llama.cpp/src/models/openelm.cpp        |   124 +
 llama/llama.cpp/src/models/orion.cpp          |   123 +
 llama/llama.cpp/src/models/pangu-embedded.cpp |   121 +
 llama/llama.cpp/src/models/phi2.cpp           |   121 +
 llama/llama.cpp/src/models/phi3.cpp           |   152 +
 llama/llama.cpp/src/models/plamo.cpp          |   110 +
 llama/llama.cpp/src/models/plamo2.cpp         |   316 +
 llama/llama.cpp/src/models/plm.cpp            |   168 +
 llama/llama.cpp/src/models/qwen.cpp           |   108 +
 llama/llama.cpp/src/models/qwen2.cpp          |   117 +
 llama/llama.cpp/src/models/qwen2moe.cpp       |   151 +
 llama/llama.cpp/src/models/qwen2vl.cpp        |   117 +
 llama/llama.cpp/src/models/qwen3.cpp          |   117 +
 llama/llama.cpp/src/models/qwen3moe.cpp       |   124 +
 llama/llama.cpp/src/models/qwen3next.cpp      |  1042 ++
 llama/llama.cpp/src/models/qwen3vl-moe.cpp    |   149 +
 llama/llama.cpp/src/models/qwen3vl.cpp        |   141 +
 llama/llama.cpp/src/models/refact.cpp         |    94 +
 llama/llama.cpp/src/models/rnd1.cpp           |   126 +
 llama/llama.cpp/src/models/rwkv6-base.cpp     |   162 +
 llama/llama.cpp/src/models/rwkv6.cpp          |    94 +
 llama/llama.cpp/src/models/rwkv6qwen2.cpp     |    86 +
 llama/llama.cpp/src/models/rwkv7-base.cpp     |   135 +
 llama/llama.cpp/src/models/rwkv7.cpp          |    90 +
 llama/llama.cpp/src/models/seed-oss.cpp       |   124 +
 llama/llama.cpp/src/models/smallthinker.cpp   |   120 +
 llama/llama.cpp/src/models/smollm3.cpp        |   128 +
 llama/llama.cpp/src/models/solar.cpp          |   158 +
 llama/llama.cpp/src/models/stablelm.cpp       |   146 +
 llama/llama.cpp/src/models/starcoder.cpp      |   100 +
 llama/llama.cpp/src/models/starcoder2.cpp     |   121 +
 llama/llama.cpp/src/models/t5-dec.cpp         |   166 +
 llama/llama.cpp/src/models/t5-enc.cpp         |    96 +
 .../llama.cpp/src/models/wavtokenizer-dec.cpp |   149 +
 llama/llama.cpp/src/models/xverse.cpp         |   108 +
 llama/llama.cpp/src/unicode.cpp               |    77 +
 llama/llama.cpp/tools/mtmd/clip-impl.h        |    38 +-
 llama/llama.cpp/tools/mtmd/clip.cpp           |  1628 +-
 llama/llama.cpp/tools/mtmd/clip.h             |    11 +-
 llama/llama.cpp/tools/mtmd/mtmd-helper.cpp    |    65 +-
 llama/llama.cpp/tools/mtmd/mtmd-helper.h      |     5 +
 llama/llama.cpp/tools/mtmd/mtmd.cpp           |    68 +-
 llama/llama.cpp/tools/mtmd/mtmd.h             |    14 +-
 ...loc-and-free-using-the-same-compiler.patch |    44 +-
 llama/patches/0002-pretokenizer.patch         |     8 +-
 llama/patches/0003-clip-unicode.patch         |    12 +-
 llama/patches/0004-solar-pro.patch            |   172 +-
 .../0005-fix-deepseek-deseret-regex.patch     |     4 +-
 ...ntain-ordering-for-rules-for-grammar.patch |     4 +-
 ...target-ggml-cpu-for-all-cpu-variants.patch |     6 +-
 llama/patches/0009-remove-amx.patch           |     4 +-
 .../0010-fix-string-arr-kv-loading.patch      |     4 +-
 llama/patches/0011-ollama-debug-tensor.patch  |     4 +-
 ...add-ollama-vocab-for-grammar-support.patch |    49 +-
 ...13-add-argsort-and-cuda-copy-for-i32.patch |    99 +-
 ...14-graph-memory-reporting-on-failure.patch |    20 +-
 .../patches/0015-ggml-Export-GPU-UUIDs.patch  |    26 +-
 .../0016-add-C-API-for-mtmd_input_text.patch  |     6 +-
 ...-no-power-throttling-win32-with-gnuc.patch |     4 +-
 .../0018-ggml-Add-batch-size-hint.patch       |    30 +-
 llama/patches/0020-ggml-No-alloc-mode.patch   |   100 +-
 .../0021-decode-disable-output_all.patch      |     6 +-
 ...gml-Enable-resetting-backend-devices.patch |    14 +-
 .../0024-GPU-discovery-enhancements.patch     |   133 +-
 .../patches/0027-interleave-multi-rope.patch  |   173 +-
 ...-Add-memory-detection-using-DXGI-PDH.patch |    43 +-
 ...> 0029-ggml-cuda-skip-large-batches.patch} |     4 +-
 ..._vk_buffer_write_2d-from-ggml_vk_buf.patch |    32 -
 ...er-Dot-Refactor-and-K-Quant-support-.patch |  2140 ---
 ...h => 0030-win-exit-instead-of-abort.patch} |     2 +-
 .../0031-fix-bakllava-regression.patch        |    25 +
 ...pk_moe-fusion-to-handle-gpt-s-late-s.patch |   657 -
 ...0032-vulkan-Fuse-rope-set_rows-16769.patch |  1242 --
 ...gsort-with-a-large-number-of-rows-16.patch |    85 -
 ...shmem-overrun-in-mmq-id-shader-16873.patch |    77 -
 ...-when-FP16-mul_mat-accumulation-is-n.patch |    80 -
 ml/backend/ggml/ggml.go                       |     2 +-
 ml/backend/ggml/ggml/include/ggml-rpc.h       |     2 +-
 ml/backend/ggml/ggml/include/ggml.h           |    93 +-
 ml/backend/ggml/ggml/src/CMakeLists.txt       |    47 +-
 ml/backend/ggml/ggml/src/ggml-alloc.c         |    26 +-
 ml/backend/ggml/ggml/src/ggml-backend.cpp     |    14 +-
 .../ggml/ggml/src/ggml-cpu/CMakeLists.txt     |   165 +-
 .../ggml/ggml/src/ggml-cpu/arch-fallback.h    |    24 +-
 .../ggml/ggml/src/ggml-cpu/arch/arm/arm.go    |     2 +
 .../ggml/ggml/src/ggml-cpu/arch/arm/quants.c  |   454 +-
 .../ggml/src/ggml-cpu/arch/arm/repack.cpp     |   721 +
 .../ggml/src/ggml-cpu/arch/x86/repack.cpp     |    12 +-
 .../ggml/ggml/src/ggml-cpu/ggml-cpu-impl.h    |     4 +-
 ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c  |    80 +-
 ml/backend/ggml/ggml/src/ggml-cpu/ops.cpp     |   872 +-
 ml/backend/ggml/ggml/src/ggml-cpu/ops.h       |     9 +-
 ml/backend/ggml/ggml/src/ggml-cpu/repack.cpp  |   413 +-
 ml/backend/ggml/ggml/src/ggml-cpu/repack.h    |     6 +
 .../ggml/ggml/src/ggml-cpu/simd-mappings.h    |    80 +-
 .../ggml/ggml/src/ggml-cpu/unary-ops.cpp      |    16 +
 ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.h |     2 +
 ml/backend/ggml/ggml/src/ggml-cpu/vec.cpp     |    17 +
 ml/backend/ggml/ggml/src/ggml-cpu/vec.h       |   290 +-
 .../ggml/ggml/src/ggml-cuda/CMakeLists.txt    |     1 +
 ml/backend/ggml/ggml/src/ggml-cuda/argsort.cu |     6 +-
 ml/backend/ggml/ggml/src/ggml-cuda/common.cuh |   252 +-
 .../ggml/ggml/src/ggml-cuda/convert.cuh       |    10 +
 .../ggml/ggml/src/ggml-cuda/cpy-utils.cuh     |     2 +-
 ml/backend/ggml/ggml/src/ggml-cuda/cpy.cu     |   255 +-
 .../ggml/ggml/src/ggml-cuda/fattn-common.cuh  |     4 +-
 .../ggml/ggml/src/ggml-cuda/fattn-tile.cu     |     4 +
 .../ggml/ggml/src/ggml-cuda/fattn-tile.cuh    |    33 +-
 .../ggml/ggml/src/ggml-cuda/fattn-vec.cuh     |    36 +-
 ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu   |     5 +-
 .../ggml/ggml/src/ggml-cuda/ggml-cuda.cu      |   951 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mma.cuh    |   426 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmf.cu     |    22 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmf.cuh    |    94 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu     |     8 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh    |   441 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cu    |   393 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cuh   |     5 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu    |   336 +-
 ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cuh   |     2 +-
 ml/backend/ggml/ggml/src/ggml-cuda/rope.cu    |   265 +-
 ml/backend/ggml/ggml/src/ggml-cuda/rope.cuh   |     2 +
 .../ggml/ggml/src/ggml-cuda/set-rows.cu       |   148 +-
 ml/backend/ggml/ggml/src/ggml-cuda/set.cu     |    39 +
 ml/backend/ggml/ggml/src/ggml-cuda/set.cuh    |     7 +
 .../ggml/ggml/src/ggml-cuda/solve_tri.cu      |   203 +
 .../ggml/ggml/src/ggml-cuda/solve_tri.cuh     |     3 +
 .../fattn-tile-instance-dkq72-dv72.cu         |     5 +
 .../ggml/ggml/src/ggml-cuda/topk-moe.cu       |    66 +-
 .../ggml/ggml/src/ggml-cuda/topk-moe.cuh      |     5 +-
 ml/backend/ggml/ggml/src/ggml-cuda/unary.cu   |    62 +-
 ml/backend/ggml/ggml/src/ggml-cuda/unary.cuh  |    33 +
 ml/backend/ggml/ggml/src/ggml-cuda/upscale.cu |    93 +-
 .../ggml/ggml/src/ggml-cuda/vendors/hip.h     |     2 +-
 .../ggml/ggml/src/ggml-hip/CMakeLists.txt     |     5 +-
 ml/backend/ggml/ggml/src/ggml-impl.h          |     2 +-
 .../ggml/src/ggml-metal/ggml-metal-context.m  |    11 +-
 .../ggml/src/ggml-metal/ggml-metal-device.cpp |   161 +-
 .../ggml/src/ggml-metal/ggml-metal-device.h   |    11 +-
 .../ggml/src/ggml-metal/ggml-metal-device.m   |   225 +-
 .../src/ggml-metal/ggml-metal-embed.metal     |  1185 +-
 .../ggml/src/ggml-metal/ggml-metal-impl.h     |   102 +-
 .../ggml/src/ggml-metal/ggml-metal-ops.cpp    |   559 +-
 .../ggml/ggml/src/ggml-metal/ggml-metal-ops.h |     3 +
 .../ggml/ggml/src/ggml-metal/ggml-metal.cpp   |     9 +
 .../ggml/ggml/src/ggml-metal/ggml-metal.metal |  1083 +-
 .../ggml/ggml/src/ggml-vulkan/ggml-vulkan.cpp |  4859 +++---
 .../src/ggml-vulkan/vulkan-shaders/abs.comp   |    21 +
 .../src/ggml-vulkan/vulkan-shaders/add1.comp  |    28 +
 .../ggml-vulkan/vulkan-shaders/arange.comp    |    20 +
 .../ggml-vulkan/vulkan-shaders/argsort.comp   |    41 +-
 .../vulkan-shaders/argsort_large.comp         |   114 +
 .../src/ggml-vulkan/vulkan-shaders/ceil.comp  |    22 +
 .../ggml-vulkan/vulkan-shaders/conv2d_mm.comp |    75 +-
 .../vulkan-shaders/copy_transpose.comp        |    67 +
 .../ggml-vulkan/vulkan-shaders/cumsum.comp    |    69 +
 .../vulkan-shaders/dequant_funcs.glsl         |     7 -
 .../src/ggml-vulkan/vulkan-shaders/fill.comp  |    19 +
 .../vulkan-shaders/flash_attn.comp            |    48 +-
 .../vulkan-shaders/flash_attn_cm1.comp        |    35 +-
 .../vulkan-shaders/flash_attn_cm2.comp        |    68 +-
 .../src/ggml-vulkan/vulkan-shaders/floor.comp |    22 +
 .../vulkan-shaders/generic_binary_head.glsl   |    15 +
 .../vulkan-shaders/generic_unary_head.glsl    |     7 +
 .../src/ggml-vulkan/vulkan-shaders/log.comp   |    18 +
 .../vulkan-shaders/mul_mat_vec.comp           |     1 +
 .../vulkan-shaders/mul_mat_vec_base.glsl      |    91 +-
 .../vulkan-shaders/mul_mat_vec_iface.glsl     |    35 +
 .../vulkan-shaders/mul_mat_vec_nc.comp        |    16 +-
 .../vulkan-shaders/mul_mat_vec_p021.comp      |    16 +-
 .../vulkan-shaders/mul_mat_vecq.comp          |    62 +-
 .../vulkan-shaders/mul_mat_vecq_funcs.glsl    |   379 +
 .../ggml-vulkan/vulkan-shaders/mul_mm.comp    |    33 +-
 .../ggml-vulkan/vulkan-shaders/mul_mmq.comp   |     8 +-
 .../vulkan-shaders/mul_mmq_funcs.glsl         |   251 +-
 .../ggml-vulkan/vulkan-shaders/multi_add.comp |   104 +-
 .../src/ggml-vulkan/vulkan-shaders/neg.comp   |    20 +
 .../vulkan-shaders/quantize_q8_1.comp         |     2 +-
 .../ggml-vulkan/vulkan-shaders/rms_norm.comp  |    44 +-
 .../vulkan-shaders/rope_funcs.glsl            |   227 +
 .../ggml-vulkan/vulkan-shaders/rope_head.glsl |    55 +-
 .../vulkan-shaders/rope_multi.comp            |    49 +-
 .../ggml-vulkan/vulkan-shaders/rope_neox.comp |    45 +-
 .../ggml-vulkan/vulkan-shaders/rope_norm.comp |    45 +-
 .../vulkan-shaders/rope_params.glsl           |    27 +
 .../vulkan-shaders/rope_vision.comp           |    44 +-
 .../src/ggml-vulkan/vulkan-shaders/round.comp |    29 +
 .../ggml-vulkan/vulkan-shaders/softplus.comp  |    23 +
 .../ggml-vulkan/vulkan-shaders/solve_tri.comp |    72 +
 .../src/ggml-vulkan/vulkan-shaders/step.comp  |    22 +
 .../ggml-vulkan/vulkan-shaders/sum_rows.comp  |    25 +-
 .../ggml-vulkan/vulkan-shaders/sum_rows.glsl  |    25 +
 .../vulkan-shaders/topk_argsort.comp          |   113 +
 .../vulkan-shaders/topk_nary_search.comp      |   199 +
 .../src/ggml-vulkan/vulkan-shaders/tri.comp   |    43 +
 .../src/ggml-vulkan/vulkan-shaders/trunc.comp |    22 +
 .../ggml-vulkan/vulkan-shaders/upscale.comp   |    46 +-
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |   145 +-
 ml/backend/ggml/ggml/src/ggml.c               |   232 +-
 303 files changed, 32711 insertions(+), 23435 deletions(-)
 create mode 100644 llama/llama.cpp/src/models/afmoe.cpp
 create mode 100644 llama/llama.cpp/src/models/apertus.cpp
 create mode 100644 llama/llama.cpp/src/models/arcee.cpp
 create mode 100644 llama/llama.cpp/src/models/arctic.cpp
 create mode 100644 llama/llama.cpp/src/models/arwkv7.cpp
 create mode 100644 llama/llama.cpp/src/models/baichuan.cpp
 create mode 100644 llama/llama.cpp/src/models/bailingmoe.cpp
 create mode 100644 llama/llama.cpp/src/models/bailingmoe2.cpp
 create mode 100644 llama/llama.cpp/src/models/bert.cpp
 create mode 100644 llama/llama.cpp/src/models/bitnet.cpp
 create mode 100644 llama/llama.cpp/src/models/bloom.cpp
 create mode 100644 llama/llama.cpp/src/models/chameleon.cpp
 create mode 100644 llama/llama.cpp/src/models/chatglm.cpp
 create mode 100644 llama/llama.cpp/src/models/codeshell.cpp
 create mode 100644 llama/llama.cpp/src/models/cogvlm.cpp
 create mode 100644 llama/llama.cpp/src/models/cohere2-iswa.cpp
 create mode 100644 llama/llama.cpp/src/models/command-r.cpp
 create mode 100644 llama/llama.cpp/src/models/dbrx.cpp
 create mode 100644 llama/llama.cpp/src/models/deci.cpp
 create mode 100644 llama/llama.cpp/src/models/deepseek.cpp
 create mode 100644 llama/llama.cpp/src/models/deepseek2.cpp
 create mode 100644 llama/llama.cpp/src/models/dots1.cpp
 create mode 100644 llama/llama.cpp/src/models/dream.cpp
 create mode 100644 llama/llama.cpp/src/models/ernie4-5-moe.cpp
 create mode 100644 llama/llama.cpp/src/models/ernie4-5.cpp
 create mode 100644 llama/llama.cpp/src/models/exaone.cpp
 create mode 100644 llama/llama.cpp/src/models/exaone4.cpp
 create mode 100644 llama/llama.cpp/src/models/falcon-h1.cpp
 create mode 100644 llama/llama.cpp/src/models/falcon.cpp
 create mode 100644 llama/llama.cpp/src/models/gemma-embedding.cpp
 create mode 100644 llama/llama.cpp/src/models/gemma.cpp
 create mode 100644 llama/llama.cpp/src/models/gemma2-iswa.cpp
 create mode 100644 llama/llama.cpp/src/models/gemma3-iswa.cpp
 create mode 100644 llama/llama.cpp/src/models/gemma3n-iswa.cpp
 create mode 100644 llama/llama.cpp/src/models/glm4-moe.cpp
 create mode 100644 llama/llama.cpp/src/models/glm4.cpp
 create mode 100644 llama/llama.cpp/src/models/gpt2.cpp
 create mode 100644 llama/llama.cpp/src/models/gptneox.cpp
 create mode 100644 llama/llama.cpp/src/models/granite-hybrid.cpp
 create mode 100644 llama/llama.cpp/src/models/granite.cpp
 create mode 100644 llama/llama.cpp/src/models/graph-context-mamba.cpp
 create mode 100644 llama/llama.cpp/src/models/grok.cpp
 create mode 100644 llama/llama.cpp/src/models/grovemoe.cpp
 create mode 100644 llama/llama.cpp/src/models/hunyuan-dense.cpp
 create mode 100644 llama/llama.cpp/src/models/hunyuan-moe.cpp
 create mode 100644 llama/llama.cpp/src/models/internlm2.cpp
 create mode 100644 llama/llama.cpp/src/models/jais.cpp
 create mode 100644 llama/llama.cpp/src/models/jamba.cpp
 create mode 100644 llama/llama.cpp/src/models/lfm2.cpp
 create mode 100644 llama/llama.cpp/src/models/llada-moe.cpp
 create mode 100644 llama/llama.cpp/src/models/llada.cpp
 create mode 100644 llama/llama.cpp/src/models/llama-iswa.cpp
 create mode 100644 llama/llama.cpp/src/models/llama.cpp
 create mode 100644 llama/llama.cpp/src/models/mamba.cpp
 create mode 100644 llama/llama.cpp/src/models/minicpm3.cpp
 create mode 100644 llama/llama.cpp/src/models/minimax-m2.cpp
 create mode 100644 llama/llama.cpp/src/models/models.go
 create mode 100644 llama/llama.cpp/src/models/models.h
 create mode 100644 llama/llama.cpp/src/models/mpt.cpp
 create mode 100644 llama/llama.cpp/src/models/nemotron-h.cpp
 create mode 100644 llama/llama.cpp/src/models/nemotron.cpp
 create mode 100644 llama/llama.cpp/src/models/neo-bert.cpp
 create mode 100644 llama/llama.cpp/src/models/olmo.cpp
 create mode 100644 llama/llama.cpp/src/models/olmo2.cpp
 create mode 100644 llama/llama.cpp/src/models/olmoe.cpp
 create mode 100644 llama/llama.cpp/src/models/openai-moe-iswa.cpp
 create mode 100644 llama/llama.cpp/src/models/openelm.cpp
 create mode 100644 llama/llama.cpp/src/models/orion.cpp
 create mode 100644 llama/llama.cpp/src/models/pangu-embedded.cpp
 create mode 100644 llama/llama.cpp/src/models/phi2.cpp
 create mode 100644 llama/llama.cpp/src/models/phi3.cpp
 create mode 100644 llama/llama.cpp/src/models/plamo.cpp
 create mode 100644 llama/llama.cpp/src/models/plamo2.cpp
 create mode 100644 llama/llama.cpp/src/models/plm.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen2.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen2moe.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen2vl.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen3.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen3moe.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen3next.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen3vl-moe.cpp
 create mode 100644 llama/llama.cpp/src/models/qwen3vl.cpp
 create mode 100644 llama/llama.cpp/src/models/refact.cpp
 create mode 100644 llama/llama.cpp/src/models/rnd1.cpp
 create mode 100644 llama/llama.cpp/src/models/rwkv6-base.cpp
 create mode 100644 llama/llama.cpp/src/models/rwkv6.cpp
 create mode 100644 llama/llama.cpp/src/models/rwkv6qwen2.cpp
 create mode 100644 llama/llama.cpp/src/models/rwkv7-base.cpp
 create mode 100644 llama/llama.cpp/src/models/rwkv7.cpp
 create mode 100644 llama/llama.cpp/src/models/seed-oss.cpp
 create mode 100644 llama/llama.cpp/src/models/smallthinker.cpp
 create mode 100644 llama/llama.cpp/src/models/smollm3.cpp
 create mode 100644 llama/llama.cpp/src/models/solar.cpp
 create mode 100644 llama/llama.cpp/src/models/stablelm.cpp
 create mode 100644 llama/llama.cpp/src/models/starcoder.cpp
 create mode 100644 llama/llama.cpp/src/models/starcoder2.cpp
 create mode 100644 llama/llama.cpp/src/models/t5-dec.cpp
 create mode 100644 llama/llama.cpp/src/models/t5-enc.cpp
 create mode 100644 llama/llama.cpp/src/models/wavtokenizer-dec.cpp
 create mode 100644 llama/llama.cpp/src/models/xverse.cpp
 rename llama/patches/{0036-ggml-cuda-skip-large-batches.patch => 0029-ggml-cuda-skip-large-batches.patch} (91%)
 delete mode 100644 llama/patches/0029-vulkan-Call-ggml_vk_buffer_write_2d-from-ggml_vk_buf.patch
 delete mode 100644 llama/patches/0030-Vulkan-MMQ-Integer-Dot-Refactor-and-K-Quant-support-.patch
 rename llama/patches/{0036-win-exit-instead-of-abort.patch => 0030-win-exit-instead-of-abort.patch} (95%)
 create mode 100644 llama/patches/0031-fix-bakllava-regression.patch
 delete mode 100644 llama/patches/0031-vulkan-Update-topk_moe-fusion-to-handle-gpt-s-late-s.patch
 delete mode 100644 llama/patches/0032-vulkan-Fuse-rope-set_rows-16769.patch
 delete mode 100644 llama/patches/0033-vulkan-Handle-argsort-with-a-large-number-of-rows-16.patch
 delete mode 100644 llama/patches/0034-vulkan-fix-shmem-overrun-in-mmq-id-shader-16873.patch
 delete mode 100644 llama/patches/0035-vulkan-Fix-crash-when-FP16-mul_mat-accumulation-is-n.patch
 create mode 100644 ml/backend/ggml/ggml/src/ggml-cuda/set.cu
 create mode 100644 ml/backend/ggml/ggml/src/ggml-cuda/set.cuh
 create mode 100644 ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cu
 create mode 100644 ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cuh
 create mode 100644 ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/log.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/round.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/step.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp
 create mode 100644 ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp

diff --git a/Makefile.sync b/Makefile.sync
index b1fcde45..4991ad84 100644
--- a/Makefile.sync
+++ b/Makefile.sync
@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=3cfa9c3f125763305b4226bc032f1954f08990dc
+FETCH_HEAD=7f8ef50cce40e3e7e4526a3696cb45658190e69a
 
 .PHONY: help
 help:
diff --git a/llama/build-info.cpp b/llama/build-info.cpp
index 7f5e28c7..0122c7ed 100644
--- a/llama/build-info.cpp
+++ b/llama/build-info.cpp
@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "3cfa9c3f125763305b4226bc032f1954f08990dc";
+char const *LLAMA_COMMIT = "7f8ef50cce40e3e7e4526a3696cb45658190e69a";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";
diff --git a/llama/llama.cpp/.rsync-filter b/llama/llama.cpp/.rsync-filter
index 650d9463..df75ca65 100644
--- a/llama/llama.cpp/.rsync-filter
+++ b/llama/llama.cpp/.rsync-filter
@@ -22,6 +22,9 @@ include /src/llama.*
 include /src/llama-*.*
 include /src/unicode-data.*
 include /src/unicode.*
+include /src/models/
+include /src/models/*.h
+include /src/models/*.cpp
 include /vendor/
 include /vendor/miniaudio/
 include /vendor/miniaudio/*.h
diff --git a/llama/llama.cpp/common/common.cpp b/llama/llama.cpp/common/common.cpp
index b0591e84..0d7fd9a9 100644
--- a/llama/llama.cpp/common/common.cpp
+++ b/llama/llama.cpp/common/common.cpp
@@ -8,6 +8,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "sampling.h"
 
 #include <algorithm>
 #include <cinttypes>
@@ -26,7 +27,6 @@
 #include <sstream>
 #include <string>
 #include <thread>
-#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -60,6 +60,14 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+common_time_meas::common_time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+
+common_time_meas::~common_time_meas() {
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
+    }
+}
+
 //
 // CPU utils
 //
@@ -355,11 +363,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
 }
 
 void common_init() {
-    llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
-        if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
-            common_log_add(common_log_main(), level, "%s", text);
-        }
-    }, NULL);
+    llama_log_set(common_log_default_callback, NULL);
 
 #ifdef NDEBUG
     const char * build_type = "";
@@ -908,11 +912,96 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
+std::vector<common_file_info> fs_list_files(const std::string & path) {
+    std::vector<common_file_info> files;
+    if (path.empty()) return files;
+
+    std::filesystem::path dir(path);
+    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
+        return files;
+    }
+
+    for (const auto & entry : std::filesystem::directory_iterator(dir)) {
+        try {
+            // Only include regular files (skip directories)
+            const auto & p = entry.path();
+            if (std::filesystem::is_regular_file(p)) {
+                common_file_info info;
+                info.path = p.string();
+                info.name = p.filename().string();
+                try {
+                    info.size = static_cast<size_t>(std::filesystem::file_size(p));
+                } catch (const std::filesystem::filesystem_error &) {
+                    info.size = 0;
+                }
+                files.push_back(std::move(info));
+            }
+        } catch (const std::filesystem::filesystem_error &) {
+            // skip entries we cannot inspect
+            continue;
+        }
+    }
+
+    return files;
+}
+
 
 //
 // Model utils
 //
 
+static inline void common_init_sampler_from_model(
+    const llama_model * model,
+    common_params_sampling & sparams) {
+
+    const uint64_t config = sparams.user_sampling_config;
+
+    auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[64] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            int32_t v = strtol(buf, &end, 10);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
+        if (config & user_config) return;
+
+        char buf[128] = {0};
+        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
+            char * end = nullptr;
+            float v = strtof(buf, &end);
+            if (end && end != buf) dst = v;
+        }
+    };
+
+    // Sampling sequence
+    if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
+        char buf[512] = {0};
+        if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
+            const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
+            if (!sampler_names.empty()) {
+                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+            }
+        }
+    }
+
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K),           sparams.top_k,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P),           sparams.top_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P),           sparams.min_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD),   sparams.xtc_threshold,   common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP),            sparams.temp,            common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N),  sparams.penalty_last_n,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT),  sparams.penalty_repeat,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT),        sparams.mirostat,        common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU),    sparams.mirostat_tau,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
+}
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
@@ -924,6 +1013,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
+    common_init_sampler_from_model(model, params.sampling);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     auto cparams = common_context_params_to_llama(params);
diff --git a/llama/llama.cpp/common/common.h b/llama/llama.cpp/common/common.h
index a8cb630e..2f23d0ba 100644
--- a/llama/llama.cpp/common/common.h
+++ b/llama/llama.cpp/common/common.h
@@ -2,17 +2,15 @@
 
 #pragma once
 
+#include "ggml-opt.h"
+#include "llama-cpp.h"
+
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
-#include <sstream>
-#include <cmath>
-
-#include "ggml-opt.h"
-#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -30,6 +28,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct common_time_meas {
+    common_time_meas(int64_t & t_acc, bool disable = false);
+    ~common_time_meas();
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
+
 struct common_adapter_lora_info {
     std::string path;
     float scale;
@@ -133,6 +140,22 @@ struct common_grammar_trigger {
     llama_token token = LLAMA_TOKEN_NULL;
 };
 
+enum common_params_sampling_config : uint64_t {
+    COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS        = 1 << 0,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_K           = 1 << 1,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_P           = 1 << 2,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIN_P           = 1 << 3,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD   = 1 << 5,
+    COMMON_PARAMS_SAMPLING_CONFIG_TEMP            = 1 << 6,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N  = 1 << 7,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT  = 1 << 8,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT        = 1 << 9,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU    = 1 << 10,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
+};
+
+
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -165,6 +188,8 @@ struct common_params_sampling {
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
+    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
+
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
 
@@ -406,6 +431,8 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
+    int image_min_tokens = -1;
+    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;
@@ -458,7 +485,8 @@ struct common_params {
     float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
-    bool is_pp_shared = false;
+    bool is_pp_shared   = false;
+    bool is_tg_separate = false;
 
     std::vector<int32_t> n_pp;
     std::vector<int32_t> n_tg;
@@ -505,6 +533,10 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
+
+    bool has_speculative() const {
+        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
+    }
 };
 
 // call once at the start of a program if it uses libcommon
@@ -605,6 +637,13 @@ bool fs_create_directory_with_parents(const std::string & path);
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
 
+struct common_file_info {
+    std::string path;
+    std::string name;
+    size_t      size = 0; // in bytes
+};
+std::vector<common_file_info> fs_list_files(const std::string & path);
+
 //
 // Model utils
 //
diff --git a/llama/llama.cpp/common/json-schema-to-grammar.cpp b/llama/llama.cpp/common/json-schema-to-grammar.cpp
index d88f4320..cb659915 100644
--- a/llama/llama.cpp/common/json-schema-to-grammar.cpp
+++ b/llama/llama.cpp/common/json-schema-to-grammar.cpp
@@ -268,10 +268,10 @@ static bool is_reserved_name(const std::string & name) {
 }
 
 std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
 std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
 std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
-    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
@@ -303,6 +303,8 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
+std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
+
 class SchemaConverter {
 private:
     friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
@@ -601,7 +603,10 @@ private:
     }
 
     std::string _resolve_ref(const std::string & ref) {
-        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        auto it = ref.find('#');
+        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
+        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
+        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
             _refs_being_resolved.insert(ref);
             json resolved = _refs[ref];
@@ -774,11 +779,24 @@ public:
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
-                            if (target.is_null() || !target.contains(sel)) {
+                            if (target.is_object() && target.contains(sel)) {
+                                target = target[sel];
+                            } else if (target.is_array()) {
+                                size_t sel_index;
+                                try {
+                                    sel_index = std::stoul(sel);
+                                } catch (const std::invalid_argument & e) {
+                                    sel_index = target.size();
+                                }
+                                if (sel_index >= target.size()) {
+                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                    return;
+                                }
+                                target = target[sel_index];
+                            } else {
                                 _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
                                 return;
                             }
-                            target = target[sel];
                         }
                         _refs[ref] = target;
                     }
diff --git a/llama/llama.cpp/common/json-schema-to-grammar.h b/llama/llama.cpp/common/json-schema-to-grammar.h
index 362991b5..c89ab7f9 100644
--- a/llama/llama.cpp/common/json-schema-to-grammar.h
+++ b/llama/llama.cpp/common/json-schema-to-grammar.h
@@ -18,4 +18,6 @@ struct common_grammar_options {
     bool dotall = false;
 };
 
+std::string gbnf_format_literal(const std::string & literal);
+
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});
diff --git a/llama/llama.cpp/common/log.cpp b/llama/llama.cpp/common/log.cpp
index 4ccdbd17..a24782b7 100644
--- a/llama/llama.cpp/common/log.cpp
+++ b/llama/llama.cpp/common/log.cpp
@@ -442,3 +442,9 @@ void common_log_set_prefix(struct common_log * log, bool prefix) {
 void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
+
+void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
+    if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+        common_log_add(common_log_main(), level, "%s", text);
+    }
+}
diff --git a/llama/llama.cpp/common/log.h b/llama/llama.cpp/common/log.h
index f329b434..7edb239a 100644
--- a/llama/llama.cpp/common/log.h
+++ b/llama/llama.cpp/common/log.h
@@ -36,6 +36,8 @@ extern int common_log_verbosity_thold;
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
+void common_log_default_callback(enum ggml_log_level level, const char * text, void * user_data);
+
 // the common_log uses an internal worker thread to print/write log messages
 // when the worker thread is paused, incoming log messages are discarded
 struct common_log;
diff --git a/llama/llama.cpp/common/sampling.cpp b/llama/llama.cpp/common/sampling.cpp
index c69d525b..7a6b7be1 100644
--- a/llama/llama.cpp/common/sampling.cpp
+++ b/llama/llama.cpp/common/sampling.cpp
@@ -3,9 +3,10 @@
 #include "common.h"
 #include "log.h"
 
-#include <cmath>
-#include <unordered_map>
 #include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <unordered_map>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -112,6 +113,13 @@ struct common_sampler {
 
     llama_token_data_array cur_p;
 
+    void reset() {
+        prev.clear();
+
+        llama_sampler_reset(grmr);
+        llama_sampler_reset(chain);
+    }
+
     void set_logits(struct llama_context * ctx, int idx) {
         const auto * logits = llama_get_logits_ith(ctx, idx);
 
@@ -128,6 +136,12 @@ struct common_sampler {
 
         cur_p = { cur.data(), cur.size(), -1, false };
     }
+
+    common_time_meas tm() {
+        return common_time_meas(t_total_us, params.no_perf);
+    }
+
+    mutable int64_t t_total_us = 0;
 };
 
 std::string common_params_sampling::print() const {
@@ -298,6 +312,8 @@ void common_sampler_free(struct common_sampler * gsmpl) {
 }
 
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
+    const auto tm = gsmpl->tm();
+
     if (accept_grammar) {
         llama_sampler_accept(gsmpl->grmr, token);
     }
@@ -308,9 +324,7 @@ void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, boo
 }
 
 void common_sampler_reset(struct common_sampler * gsmpl) {
-    llama_sampler_reset(gsmpl->grmr);
-
-    llama_sampler_reset(gsmpl->chain);
+    gsmpl->reset();
 }
 
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
@@ -327,16 +341,54 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
     // TODO: measure grammar performance
 
+    const double t_sampling_ms = gsmpl ? 1e-3*gsmpl->t_total_us : 0;
+
+    llama_perf_sampler_data data_smpl;
+    llama_perf_context_data data_ctx;
+
+    memset(&data_smpl, 0, sizeof(data_smpl));
+    memset(&data_ctx,  0, sizeof(data_ctx));
+
     if (gsmpl) {
-        llama_perf_sampler_print(gsmpl->chain);
+        auto & data = data_smpl;
+
+        data = llama_perf_sampler(gsmpl->chain);
+
+        // note: the sampling time includes the samplers time + extra time spent in common/sampling
+        LOG_INF("%s:    sampling time = %10.2f ms\n", __func__, t_sampling_ms);
+        LOG_INF("%s:    samplers time = %10.2f ms / %5d tokens\n", __func__, data.t_sample_ms, data.n_sample);
     }
+
     if (ctx) {
-        llama_perf_context_print(ctx);
+        auto & data = data_ctx;
+
+        data = llama_perf_context(ctx);
+
+        const double t_end_ms = 1e-3 * ggml_time_us();
+
+        const double t_total_ms = t_end_ms - data.t_start_ms;
+        const double t_unacc_ms = t_total_ms - (t_sampling_ms + data.t_p_eval_ms + data.t_eval_ms);
+        const double t_unacc_pc = 100.0 * t_unacc_ms /  t_total_ms;
+
+        LOG_INF("%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
+        LOG_INF("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
+        LOG_INF("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
+        LOG_INF("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
+        LOG_INF("%s: unaccounted time = %10.2f ms / %5.1f %%      (total - sampling - prompt eval - eval) / (total)\n", __func__, t_unacc_ms, t_unacc_pc);
+        LOG_INF("%s:    graphs reused = %10d\n", __func__, data.n_reused);
+
         llama_memory_breakdown_print(ctx);
     }
 }
 
 llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+    llama_synchronize(ctx);
+
+    // start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
+    const auto tm = gsmpl->tm();
+
     gsmpl->set_logits(ctx, idx);
 
     auto & grmr  = gsmpl->grmr;
@@ -428,6 +480,8 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
 // helpers
 
 llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    const auto tm = gsmpl->tm();
+
     auto * res = &gsmpl->cur_p;
 
     if (do_sort && !res->sorted) {
diff --git a/llama/llama.cpp/include/llama.h b/llama/llama.cpp/include/llama.h
index a0a660bf..b52eaacf 100644
--- a/llama/llama.cpp/include/llama.h
+++ b/llama/llama.cpp/include/llama.h
@@ -83,6 +83,7 @@ extern "C" {
         LLAMA_ROPE_TYPE_NORM   = 0,
         LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
         LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+        LLAMA_ROPE_TYPE_IMROPE = GGML_ROPE_TYPE_IMROPE,
         LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
     };
 
@@ -245,6 +246,21 @@ extern "C" {
         LLAMA_KV_OVERRIDE_TYPE_STR,
     };
 
+    enum llama_model_meta_key {
+        LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_K,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIN_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD,
+        LLAMA_MODEL_META_KEY_SAMPLING_TEMP,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA,
+    };
+
     struct llama_model_kv_override {
         enum llama_model_kv_override_type tag;
 
@@ -460,7 +476,11 @@ extern "C" {
     LLAMA_API bool llama_supports_gpu_offload(void);
     LLAMA_API bool llama_supports_rpc        (void);
 
+    // NOTE: After creating a llama_context, it is recommended to query the actual values using these functions
+    //       In some cases the requested values via llama_context_params may differ from the actual values used by the context
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ctx_seq  (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
@@ -481,6 +501,7 @@ extern "C" {
 
     LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
@@ -512,6 +533,9 @@ extern "C" {
     // Get the number of metadata key/value pairs
     LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
 
+    // Get sampling metadata key name. Returns nullptr if the key is invalid
+    LLAMA_API const char * llama_model_meta_key_str(enum llama_model_meta_key key);
+
     // Get metadata key name by index
     LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
 
@@ -584,7 +608,7 @@ extern "C" {
     LLAMA_API int32_t llama_adapter_meta_val_str_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
 
     // Manually free a LoRA adapter
-    // Note: loaded adapters will be free when the associated model is deleted
+    // NOTE: loaded adapters will be free when the associated model is deleted
     LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
 
     // Get the invocation tokens if the current lora is an alora
@@ -1110,8 +1134,6 @@ extern "C" {
     //        // sample from the logits of the last token in the batch
     //        const llama_token id = llama_sampler_sample(smpl, ctx, -1);
     //
-    //        // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.)
-    //        llama_sampler_accept(smpl, id);
     //        ...
     //    }
     //
diff --git a/llama/llama.cpp/src/llama-arch.cpp b/llama/llama.cpp/src/llama-arch.cpp
index ab262ec0..b6bde25d 100644
--- a/llama/llama.cpp/src/llama-arch.cpp
+++ b/llama/llama.cpp/src/llama-arch.cpp
@@ -32,6 +32,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
     { LLM_ARCH_QWEN3,            "qwen3"            },
     { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
+    { LLM_ARCH_QWEN3NEXT,        "qwen3next"        },
+    { LLM_ARCH_QWEN3VL,          "qwen3vl"          },
+    { LLM_ARCH_QWEN3VLMOE,       "qwen3vlmoe"       },
     { LLM_ARCH_PHI2,             "phi2"             },
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
@@ -89,6 +92,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BAILINGMOE2,      "bailingmoe2"      },
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
+    { LLM_ARCH_AFMOE,            "afmoe"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
     { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
@@ -104,23 +108,39 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_SEED_OSS,         "seed_oss"         },
     { LLM_ARCH_GROVEMOE,         "grovemoe"         },
     { LLM_ARCH_APERTUS,          "apertus"          },
+    { LLM_ARCH_MINIMAX_M2,       "minimax-m2"       },
+    { LLM_ARCH_COGVLM,           "cogvlm"           },
+    { LLM_ARCH_RND1,             "rnd1"             },
+    { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
-    { LLM_KV_GENERAL_TYPE,                 "general.type"                          },
-    { LLM_KV_GENERAL_ARCHITECTURE,         "general.architecture"                  },
-    { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version"          },
-    { LLM_KV_GENERAL_ALIGNMENT,            "general.alignment"                     },
-    { LLM_KV_GENERAL_FILE_TYPE,            "general.file_type"                     },
-    { LLM_KV_GENERAL_NAME,                 "general.name"                          },
-    { LLM_KV_GENERAL_AUTHOR,               "general.author"                        },
-    { LLM_KV_GENERAL_VERSION,              "general.version"                       },
-    { LLM_KV_GENERAL_URL,                  "general.url"                           },
-    { LLM_KV_GENERAL_DESCRIPTION,          "general.description"                   },
-    { LLM_KV_GENERAL_LICENSE,              "general.license"                       },
-    { LLM_KV_GENERAL_SOURCE_URL,           "general.source.url"                    },
-    { LLM_KV_GENERAL_SOURCE_HF_REPO,       "general.source.huggingface.repository" },
+    { LLM_KV_GENERAL_TYPE,                     "general.type"                          },
+    { LLM_KV_GENERAL_ARCHITECTURE,             "general.architecture"                  },
+    { LLM_KV_GENERAL_QUANTIZATION_VERSION,     "general.quantization_version"          },
+    { LLM_KV_GENERAL_ALIGNMENT,                "general.alignment"                     },
+    { LLM_KV_GENERAL_FILE_TYPE,                "general.file_type"                     },
+    { LLM_KV_GENERAL_SAMPLING_SEQUENCE,        "general.sampling.sequence"             },
+    { LLM_KV_GENERAL_SAMPLING_TOP_K,           "general.sampling.top_k"                },
+    { LLM_KV_GENERAL_SAMPLING_TOP_P,           "general.sampling.top_p"                },
+    { LLM_KV_GENERAL_SAMPLING_MIN_P,           "general.sampling.min_p"                },
+    { LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, "general.sampling.xtc_probability"      },
+    { LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,   "general.sampling.xtc_threshold"        },
+    { LLM_KV_GENERAL_SAMPLING_TEMP,            "general.sampling.temp"                 },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,  "general.sampling.penalty_last_n"       },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,  "general.sampling.penalty_repeat"       },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT,        "general.sampling.mirostat"             },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,    "general.sampling.mirostat_tau"         },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,    "general.sampling.mirostat_eta"         },
+    { LLM_KV_GENERAL_NAME,                     "general.name"                          },
+    { LLM_KV_GENERAL_AUTHOR,                   "general.author"                        },
+    { LLM_KV_GENERAL_VERSION,                  "general.version"                       },
+    { LLM_KV_GENERAL_URL,                      "general.url"                           },
+    { LLM_KV_GENERAL_DESCRIPTION,              "general.description"                   },
+    { LLM_KV_GENERAL_LICENSE,                  "general.license"                       },
+    { LLM_KV_GENERAL_SOURCE_URL,               "general.source.url"                    },
+    { LLM_KV_GENERAL_SOURCE_HF_REPO,           "general.source.huggingface.repository" },
 
     { LLM_KV_VOCAB_SIZE,                        "%s.vocab_size"                        },
     { LLM_KV_CONTEXT_LENGTH,                    "%s.context_length"                    },
@@ -146,6 +166,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
+    { LLM_KV_NUM_DEEPSTACK_LAYERS,              "%s.n_deepstack_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -329,6 +350,36 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_AFMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_GATE,       "blk.%d.attn_gate" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
+        },
+    },
     {
         LLM_ARCH_LLAMA4,
         {
@@ -781,6 +832,77 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_QWEN3NEXT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,     "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_SSM_A,              "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_CONV1D,         "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,             "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_BETA_ALPHA,     "blk.%d.ssm_ba" },
+            { LLM_TENSOR_SSM_IN,             "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_NORM,           "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,            "blk.%d.ssm_out" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3VL,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3VLMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_PHI2,
         {
@@ -2168,7 +2290,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
-            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,       "token_embd_norm" }, // note: wrong tensor name
             { LLM_TENSOR_OUTPUT,            "output" },
         }
     },
@@ -2190,7 +2312,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
-            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,       "token_embd_norm" }, // note: wrong tensor name
             { LLM_TENSOR_FFN_GATE_INP,      "blk.%d.ffn_gate_inp" },
             { LLM_TENSOR_FFN_GATE_EXPS,     "blk.%d.ffn_gate_exps" },
             { LLM_TENSOR_FFN_DOWN_EXPS,     "blk.%d.ffn_down_exps" },
@@ -2332,6 +2454,84 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_CHEXPS,      "blk.%d.ffn_up_chexps" },
         },
     },
+    {
+        LLM_ARCH_MINIMAX_M2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        },
+    },
+    {
+        LLM_ARCH_PANGU_EMBED,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_COGVLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_VISEXP_ATTN_QKV, "blk.%d.vis_attn_qkv" },
+            { LLM_TENSOR_VISEXP_ATTN_OUT, "blk.%d.vis_attn_output" },
+            { LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" },
+            { LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" },
+            { LLM_TENSOR_VISEXP_FFN_UP,   "blk.%d.vis_up" },
+        },
+    },
+    {
+        LLM_ARCH_RND1,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2340,11 +2540,21 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
     },
 };
 
+// declare information about the model weight tensors:
+// - the layer in which the tensor is going to be used. this is needed in order to assign the correct buffer type for the weight
+// - the operator which is going to use the weight. this is needed to determine if the respective backend supports the operator
+//
+// for example, input layers are usually assigned to CPU/host buffer types
+//
+// a mismatch between the declared information and the actual layer/op in which the tensor is used can lead to sub-optimal
+//   assignment of the buffer types and extra overhead during computation
+// example: https://github.com/ggml-org/llama.cpp/pull/17548
+//
 static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TOKEN_EMBD,                 {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_POS_EMBD,                   {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
-    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_TOKEN_TYPES,                {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_INPUT, GGML_OP_MUL}},
     {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2361,6 +2571,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_GATE,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -2398,6 +2609,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_X,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_DT,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_OUT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_SSM_BETA_ALPHA,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_A1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -2509,6 +2721,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_QKV,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_OUT,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_GATE,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are currently ignored (reserved for future MTP support)
     // These tensors only exist in the last layer(s) and are treated as output tensors
     {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2592,6 +2809,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_QWEN3NEXT:
             return true;
         default:
             return false;
@@ -2603,6 +2821,7 @@ bool llm_arch_is_diffusion(const llm_arch & arch) {
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
             return true;
         default:
             return false;
diff --git a/llama/llama.cpp/src/llama-arch.h b/llama/llama.cpp/src/llama-arch.h
index ea2b4ffb..3936a468 100644
--- a/llama/llama.cpp/src/llama-arch.h
+++ b/llama/llama.cpp/src/llama-arch.h
@@ -36,6 +36,9 @@ enum llm_arch {
     LLM_ARCH_QWEN2VL,
     LLM_ARCH_QWEN3,
     LLM_ARCH_QWEN3MOE,
+    LLM_ARCH_QWEN3NEXT,
+    LLM_ARCH_QWEN3VL,
+    LLM_ARCH_QWEN3VLMOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
@@ -93,6 +96,7 @@ enum llm_arch {
     LLM_ARCH_BAILINGMOE2,
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
+    LLM_ARCH_AFMOE,
     LLM_ARCH_ERNIE4_5,
     LLM_ARCH_ERNIE4_5_MOE,
     LLM_ARCH_HUNYUAN_MOE,
@@ -108,6 +112,10 @@ enum llm_arch {
     LLM_ARCH_SEED_OSS,
     LLM_ARCH_GROVEMOE,
     LLM_ARCH_APERTUS,
+    LLM_ARCH_MINIMAX_M2,
+    LLM_ARCH_COGVLM,
+    LLM_ARCH_RND1,
+    LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -117,6 +125,18 @@ enum llm_kv {
     LLM_KV_GENERAL_QUANTIZATION_VERSION,
     LLM_KV_GENERAL_ALIGNMENT,
     LLM_KV_GENERAL_FILE_TYPE,
+    LLM_KV_GENERAL_SAMPLING_SEQUENCE,
+    LLM_KV_GENERAL_SAMPLING_TOP_K,
+    LLM_KV_GENERAL_SAMPLING_TOP_P,
+    LLM_KV_GENERAL_SAMPLING_MIN_P,
+    LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY,
+    LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,
+    LLM_KV_GENERAL_SAMPLING_TEMP,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,
     LLM_KV_GENERAL_NAME,
     LLM_KV_GENERAL_AUTHOR,
     LLM_KV_GENERAL_VERSION,
@@ -150,6 +170,7 @@ enum llm_kv {
     LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_NEXTN_PREDICT_LAYERS,
+    LLM_KV_NUM_DEEPSTACK_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -308,6 +329,7 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_POST_NORM,
     LLM_TENSOR_ATTN_ROT_EMBD,
     LLM_TENSOR_ATTN_SINKS,
+    LLM_TENSOR_ATTN_GATE,
     LLM_TENSOR_FFN_GATE_INP,
     LLM_TENSOR_FFN_GATE_INP_SHEXP,
     LLM_TENSOR_FFN_NORM,
@@ -362,6 +384,7 @@ enum llm_tensor {
     LLM_TENSOR_SSM_D,
     LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_SSM_BETA_ALPHA,      // qwen3next
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
@@ -458,6 +481,11 @@ enum llm_tensor {
     LLM_TENSOR_SHORTCONV_CONV,
     LLM_TENSOR_SHORTCONV_INPROJ,
     LLM_TENSOR_SHORTCONV_OUTPROJ,
+    LLM_TENSOR_VISEXP_ATTN_QKV,
+    LLM_TENSOR_VISEXP_ATTN_OUT,
+    LLM_TENSOR_VISEXP_FFN_GATE,
+    LLM_TENSOR_VISEXP_FFN_DOWN,
+    LLM_TENSOR_VISEXP_FFN_UP,
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,
diff --git a/llama/llama.cpp/src/llama-batch.cpp b/llama/llama.cpp/src/llama-batch.cpp
index 55d89eca..86a1a4ba 100644
--- a/llama/llama.cpp/src/llama-batch.cpp
+++ b/llama/llama.cpp/src/llama-batch.cpp
@@ -215,6 +215,7 @@ bool llama_batch_allocr::init(
             /*.n_seq_tokens =*/ (uint32_t) 1,
             /*.n_seqs       =*/ (uint32_t) batch.n_tokens,
             /*.n_seqs_unq   =*/ (uint32_t) this->seq_id_unq.size(),
+            /*.n_pos        =*/ n_pos_per_embd,
             /*.token        =*/ batch.token,
             /*.embd         =*/ batch.embd,
             /*.pos          =*/ batch.pos,
@@ -251,46 +252,72 @@ bool llama_batch_allocr::init(
     // consistency checks
     //
 
-    for (uint32_t s = 0; s < n_seq_max; ++s) {
-        if (seq_pos[s].empty()) {
-            continue;
-        }
+    if (n_pos_per_embd > 1) {
+        // M-RoPE case: allow position to "jump" forward only (non-continuous positions are allowed)
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
+            if (seq_pos[s].empty()) {
+                continue;
+            }
 
-        const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
-
-        if (p0 >= 0) {
-            bool ok = true;
+            const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
 
             if (batch.token) {
+                if (p0 >= 0 && p0 >= seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X < Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
+                }
+            } else {
+                // embedding inputs can have overlapping positions
+                if (p0 >= 0 && p0 > seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X <= Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
+                }
+            }
+        }
+    } else {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
+            if (seq_pos[s].empty()) {
+                continue;
+            }
+
+            const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
+
+            if (p0 >= 0) {
+                bool ok = true;
+
                 if (seq_pos_min(s) != p0 + 1) {
                     ok = false;
                 }
-            } else {
-                assert(batch.embd);
 
-                // for embeddings (typically used as vision input), we allow them to have repeating positions
-                // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
-                if (seq_pos_min(s) != p0 && seq_pos_min(s) != p0 + 1) {
-                    ok = false;
+                if (!ok) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " it is required that the sequence positions remain consecutive: Y = X + 1\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
                 }
             }
 
-            if (!ok) {
-                LLAMA_LOG_ERROR(
-                        "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
-                        " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
-                        " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
-                        " it is required that the sequence positions remain consecutive: Y = X + 1\n",
-                        __func__, s, s, p0, s, seq_pos_min(s));
-
+            if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
+                LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
                 return false;
             }
         }
-
-        if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
-            LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
-            return false;
-        }
     }
 
     if (memory) {
@@ -389,6 +416,7 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ n_seqs,
+        /*.n_pos        =*/ n_pos_per_embd,
 
         /*.token        =*/ udata->token.data(),
         /*.embd         =*/ nullptr,
@@ -655,10 +683,8 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
     auto udata = std::make_shared<llama_ubatch::data_t>();
 
-    const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;
-
     const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
-    const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_cur;
+    const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_per_embd;
 
     udata->token     .resize(n_tokens);
     udata->embd      .resize(n_embd_all);
@@ -680,8 +706,13 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
             memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
         }
 
-        for (int j = 0; j < n_pos_cur; ++j) {
-            udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
+        for (size_t j = 0; j < (size_t)n_pos_per_embd; ++j) {
+            // if we are using M-RoPE
+            //     if the current batch is text, we need to broadcast the same position across all RoPE sections
+            //     otherwise, the input batch is image embeddings, we copy the positions as-is
+            // if we are not using M-RoPE, there is only one position per token (this loop runs only once)
+            size_t src_off = batch.token ? 0 : j*batch.n_tokens;
+            udata->pos[j*n_tokens + i] = batch.pos[src_off + idxs[i]];
         }
 
         udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
@@ -710,6 +741,7 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
         /*.n_seq_tokens =*/ n_tokens/n_seqs,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ (uint32_t) udata->seq_id_unq.size(),
+        /*.n_pos        =*/ n_pos_per_embd,
 
         /*.token        =*/ batch.token ? udata->token.data() : nullptr,
         /*.embd         =*/ batch.embd ? udata->embd.data() : nullptr,
diff --git a/llama/llama.cpp/src/llama-batch.h b/llama/llama.cpp/src/llama-batch.h
index 0dc8cebd..209cf369 100644
--- a/llama/llama.cpp/src/llama-batch.h
+++ b/llama/llama.cpp/src/llama-batch.h
@@ -17,6 +17,16 @@ struct llama_ubatch {
         return b_equal_seqs != 0;
     }
 
+    // typical for M-RoPE cases:
+    //   0 - sequantial position of the tokens/embeddings in the sequence
+    //   1 - y position in the image
+    //   2 - x position in the image
+    //   3 - other
+    bool is_pos_2d() const {
+        // TODO @ngxson : we may need to check for model arch when more models use >1 positions
+        return n_pos >= 3;
+    }
+
     uint32_t b_equal_seqs; // note: this is a boolean, but we use an int32_t for alignment
                            //       otherwise address sanitizer complains
     // TODO: whole_seqs for embeddings?
@@ -25,6 +35,7 @@ struct llama_ubatch {
     uint32_t n_seq_tokens; // tokens per sequence set
     uint32_t n_seqs;       // sequence sets in the ubatch
     uint32_t n_seqs_unq;   // unique sequence ids in the ubatch
+    uint32_t n_pos;        // number of position inputs for each token/embedding
 
     // seq_id_unq: unique sequence ids in the ubatch
     // seq_idx:    indices of the unique sequence ids in the ubatch in [0, n_seqs_unq)
@@ -33,7 +44,7 @@ struct llama_ubatch {
     //                          // size               | idx | val
     llama_token  *  token;      // [n_tokens]         | i   | id, token
     float        *  embd;       // [n_embd, n_tokens] | i   | embd
-    llama_pos    *  pos;        // [n_tokens]         | i   | pos
+    llama_pos    *  pos;        // [n_tokens*n_pos]   | i   | pos
     int32_t      *  n_seq_id;   // [n_tokens]         | i   | -
     llama_seq_id ** seq_id;     // [n_tokens]         | s   | s0, s1, seq_id
     llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id
diff --git a/llama/llama.cpp/src/llama-chat.cpp b/llama/llama.cpp/src/llama-chat.cpp
index 0285006d..fc6a6223 100644
--- a/llama/llama.cpp/src/llama-chat.cpp
+++ b/llama/llama.cpp/src/llama-chat.cpp
@@ -73,6 +73,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
     { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
     { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
+    { "pangu-embedded",    LLM_CHAT_TEMPLATE_PANGU_EMBED       },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -213,6 +214,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_SEED_OSS;
     } else if (tmpl_contains("'Assistant: '  + message['content'] + '<|separator|>")) {
         return LLM_CHAT_TEMPLATE_GROK_2;
+    } else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) {
+        return LLM_CHAT_TEMPLATE_PANGU_EMBED;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -813,6 +816,35 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
+    }else if (tmpl == LLM_CHAT_TEMPLATE_PANGU_EMBED) {
+        // [unused9]系统：xxx[unused10]
+        // [unused9]用户：xxx[unused10]
+        // [unused9]助手：xxx[unused10]
+        // ...
+        for (size_t i = 0; i < chat.size(); ++i) {
+            const auto & msg = chat[i];
+            const std::string & role = msg->role;
+            const std::string & content = msg->content;
+
+            if (i == 0 && role != "system") {
+                ss << "[unused9]系统：[unused10]";
+            }
+
+            if (role == "system") {
+                ss << "[unused9]系统：" << content << "[unused10]";
+            } else if (role == "user") {
+                ss << "[unused9]用户：" << content << "[unused10]";
+            } else if (role == "assistant") {
+                ss << "[unused9]助手：" << content << "[unused10]";
+            } else if (role == "tool") {
+                ss << "[unused9]工具：" << content << "[unused10]";
+            } else if (role == "function") {
+                ss << "[unused9]方法：" << content << "[unused10]";
+            }
+        }
+        if (add_ass) {
+            ss << "[unused9]助手：";
+        }
     } else {
         // template not supported
         return -1;
diff --git a/llama/llama.cpp/src/llama-chat.h b/llama/llama.cpp/src/llama-chat.h
index da1b7c47..684efb4d 100644
--- a/llama/llama.cpp/src/llama-chat.h
+++ b/llama/llama.cpp/src/llama-chat.h
@@ -53,6 +53,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_SEED_OSS,
     LLM_CHAT_TEMPLATE_GROK_2,
+    LLM_CHAT_TEMPLATE_PANGU_EMBED,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 
diff --git a/llama/llama.cpp/src/llama-context.cpp b/llama/llama.cpp/src/llama-context.cpp
index 8b4a89d3..1359c614 100644
--- a/llama/llama.cpp/src/llama-context.cpp
+++ b/llama/llama.cpp/src/llama-context.cpp
@@ -1,5 +1,6 @@
 #include "llama-context.h"
 
+#include "llama-arch.h"
 #include "llama-impl.h"
 #include "llama-batch.h"
 #include "llama-io.h"
@@ -21,6 +22,8 @@ llama_context::llama_context(
               llama_context_params params) :
     model(model),
     balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
+    // TODO warning when creating llama_context with awkward ctx size that is not a power of 2,
+    //     may need to be backend-dependent
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -112,11 +115,28 @@ llama_context::llama_context(
         }
     }
 
-    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+    // ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
+    cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256);
+
+    if (cparams.kv_unified) {
+        cparams.n_ctx_seq = cparams.n_ctx;
+    } else {
+        cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max;
+        cparams.n_ctx_seq = GGML_PAD(cparams.n_ctx_seq, 256);
+
+        if (cparams.n_ctx_seq == 0) {
+            throw std::runtime_error("n_ctx_seq == 0");
+        }
+
+        if (cparams.n_ctx != cparams.n_ctx_seq * cparams.n_seq_max) {
+            cparams.n_ctx =  cparams.n_ctx_seq * cparams.n_seq_max;
+            LLAMA_LOG_WARN("%s: n_ctx is not divisible by n_seq_max - rounding down to %u\n", __func__, cparams.n_ctx);
+        }
+    }
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
     LLAMA_LOG_INFO("%s: n_ctx         = %u\n",   __func__, cparams.n_ctx);
-    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",   __func__, n_ctx_per_seq);
+    LLAMA_LOG_INFO("%s: n_ctx_seq     = %u\n",   __func__, cparams.n_ctx_seq);
     LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
@@ -125,14 +145,14 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
-    if (n_ctx_per_seq < hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq < hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
-    if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq > hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
     if (!hparams.vocab_only) {
@@ -268,9 +288,7 @@ llama_context::llama_context(
         if (pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
         }
-    }
 
-    if (!hparams.vocab_only) {
         llama_memory_context_ptr mctx;
         if (memory) {
             LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
@@ -282,7 +300,7 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         // avoid reserving graphs with zero outputs - assume one output per sequence
@@ -343,7 +361,14 @@ llama_context::llama_context(
         {
             auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
-                throw std::runtime_error("failed to allocate compute pp buffers");
+                if (pipeline_parallel) {
+                    LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
+                    sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, false, cparams.op_offload));
+                    gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+                }
+                if (!gf) {
+                    throw std::runtime_error("failed to allocate compute pp buffers");
+                }
             }
 
             n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
@@ -448,8 +473,8 @@ uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
 
-uint32_t llama_context::n_ctx_per_seq() const {
-    return cparams.n_ctx / cparams.n_seq_max;
+uint32_t llama_context::n_ctx_seq() const {
+    return cparams.n_ctx_seq;
 }
 
 uint32_t llama_context::n_batch() const {
@@ -518,7 +543,7 @@ bool llama_context::memory_update(bool optimize) {
             throw std::runtime_error("failed to initialize memory context");
         }
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -803,7 +828,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
@@ -972,7 +997,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_vocab = vocab.n_tokens();
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
 
     const bool output_all = false;
 
@@ -1223,7 +1248,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
         // make the outputs have the same order they had in the user-provided batch
         // note: this is mostly relevant for recurrent models atm
-        if (!sorted_output) {
+        if (!sorted_output && n_outputs > 1) {
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
@@ -1361,6 +1386,9 @@ void llama_context::output_reorder() {
 //
 
 uint32_t llama_context::graph_max_nodes() const {
+    if (model.arch == LLM_ARCH_QWEN3NEXT) {
+        return std::max<uint32_t>(8192u, 32u*model.n_tensors());
+    }
     return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }
 
@@ -2129,7 +2157,7 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_inp(), cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
             LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
             return;
         }
@@ -2377,6 +2405,10 @@ uint32_t llama_n_ctx(const llama_context * ctx) {
     return ctx->n_ctx();
 }
 
+uint32_t llama_n_ctx_seq(const llama_context * ctx) {
+    return ctx->n_ctx_seq();
+}
+
 uint32_t llama_n_batch(const llama_context * ctx) {
     return ctx->n_batch();
 }
diff --git a/llama/llama.cpp/src/llama-context.h b/llama/llama.cpp/src/llama-context.h
index ed6d82cb..20cbd789 100644
--- a/llama/llama.cpp/src/llama-context.h
+++ b/llama/llama.cpp/src/llama-context.h
@@ -43,11 +43,11 @@ struct llama_context {
 
     ggml_backend_sched_t get_sched() const;
 
-    uint32_t n_ctx()         const;
-    uint32_t n_ctx_per_seq() const;
-    uint32_t n_batch()       const;
-    uint32_t n_ubatch()      const;
-    uint32_t n_seq_max()     const;
+    uint32_t n_ctx()     const;
+    uint32_t n_ctx_seq() const;
+    uint32_t n_batch()   const;
+    uint32_t n_ubatch()  const;
+    uint32_t n_seq_max() const;
 
     uint32_t n_threads()       const;
     uint32_t n_threads_batch() const;
diff --git a/llama/llama.cpp/src/llama-cparams.h b/llama/llama.cpp/src/llama-cparams.h
index eae7b839..fcef8fa9 100644
--- a/llama/llama.cpp/src/llama-cparams.h
+++ b/llama/llama.cpp/src/llama-cparams.h
@@ -8,6 +8,7 @@
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
+    uint32_t n_ctx_seq;       // context for a single sequence
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;
diff --git a/llama/llama.cpp/src/llama-grammar.cpp b/llama/llama.cpp/src/llama-grammar.cpp
index b51cee09..a7307c47 100644
--- a/llama/llama.cpp/src/llama-grammar.cpp
+++ b/llama/llama.cpp/src/llama-grammar.cpp
@@ -6,8 +6,10 @@
 
 #include <cmath>
 #include <algorithm>
+#include <cstdint>
 #include <stdexcept>
 
+#define MAX_REPETITION_THRESHOLD 2000
 //
 // helpers
 //
@@ -345,8 +347,10 @@ const char * llama_grammar_parser::parse_sequence(
     size_t last_sym_start = rule.size();
     const char * pos = src;
 
-    auto handle_repetitions = [&](int min_times, int max_times) {
-
+    // use UINT64_MAX as the empty value because we aligned to the proper uint64_t type so -1 can't be used
+    // (though it's technically the same as -1 now)
+    auto handle_repetitions = [&](uint64_t min_times, uint64_t max_times) {
+        bool no_max = max_times == UINT64_MAX;
         if (last_sym_start == rule.size()) {
             throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
         }
@@ -373,20 +377,20 @@ const char * llama_grammar_parser::parse_sequence(
             rule.resize(last_sym_start);
         } else {
             // Repeat the previous elements (min_times - 1) times
-            for (int i = 1; i < min_times; i++) {
+            for (uint64_t i = 1; i < min_times; i++) {
                 rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
             }
         }
 
         uint32_t last_rec_rule_id = 0;
-        auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+        auto n_opt = no_max ? 1 : max_times - min_times;
 
         llama_grammar_rule rec_rule(prev_rule);
-        for (int i = 0; i < n_opt; i++) {
+        for (uint64_t i = 0; i < n_opt; i++) {
             rec_rule.resize(prev_rule.size());
             uint32_t rec_rule_id = generate_symbol_id( rule_name);
-            if (i > 0 || max_times < 0) {
-                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+            if (i > 0 || no_max) {
+                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, no_max ? rec_rule_id : last_rec_rule_id});
             }
             rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
             rec_rule.push_back({LLAMA_GRETYPE_END, 0});
@@ -478,10 +482,10 @@ const char * llama_grammar_parser::parse_sequence(
                 throw std::runtime_error(std::string("expecting an int at ") + pos);
             }
             const char * int_end = parse_int(pos);
-            int min_times = std::stoul(std::string(pos, int_end - pos));
+            uint64_t min_times = std::stoul(std::string(pos, int_end - pos));
             pos = parse_space(int_end, is_nested);
 
-            int max_times = -1;
+            uint64_t max_times = UINT64_MAX; // default: no max limit
 
             if (*pos == '}') {
                 max_times = min_times;
@@ -502,6 +506,10 @@ const char * llama_grammar_parser::parse_sequence(
             } else {
                 throw std::runtime_error(std::string("expecting ',' at ") + pos);
             }
+            bool has_max = max_times != UINT64_MAX;
+            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+                throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
+            }
             handle_repetitions(min_times, max_times);
         } else {
             break;
diff --git a/llama/llama.cpp/src/llama-graph.cpp b/llama/llama.cpp/src/llama-graph.cpp
index 41fa6894..1d012e09 100644
--- a/llama/llama.cpp/src/llama-graph.cpp
+++ b/llama/llama.cpp/src/llama-graph.cpp
@@ -810,6 +810,9 @@ ggml_tensor * llm_graph_context::build_ffn(
             GGML_ABORT("fatal error");
     }
 
+    //expand here so that we can fuse ffn gate
+    ggml_build_forward_expand(gf, cur);
+
     if (gate && type_gate == LLM_FFN_PAR) {
         cur = ggml_mul(ctx0, cur, tmp);
         cb(cur, "ffn_gate_par", il);
@@ -958,14 +961,14 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         // organize experts into n_expert_groups
         ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
 
-        ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
+        ggml_tensor * group_scores = ggml_argsort_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
         group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
 
         // get top n_group_used expert groups
         group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
         group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
 
-        ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
+        ggml_tensor * expert_groups = ggml_argsort_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
         cb(expert_groups, "ffn_moe_group_topk", il);
 
         // mask out the other groups
@@ -976,7 +979,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     }
 
     // select experts
-    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
+    ggml_tensor * selected_experts = ggml_argsort_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
     cb(selected_experts->src[0], "ffn_moe_argsort", il);
     cb(selected_experts, "ffn_moe_topk", il);
 
@@ -1006,10 +1009,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
         cb(weights_sum, "ffn_moe_weights_sum", il);
 
-        if (arch == LLM_ARCH_BAILINGMOE2) {
-            weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20);
-            cb(weights_sum, "ffn_moe_weights_sum_biased", il);
-        }
+        // Avoid division by zero, clamp to smallest number representable by F16
+        weights_sum = ggml_clamp(ctx0, weights_sum, 6.103515625e-5, INFINITY);
+        cb(weights_sum, "ffn_moe_weights_sum_clamped", il);
 
         weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
         cb(weights, "ffn_moe_weights_norm", il);
@@ -1091,6 +1093,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             GGML_ABORT("fatal error");
     }
 
+    //expand here so that we can fuse ffn gate
+    ggml_build_forward_expand(gf, cur);
+
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
@@ -1137,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd = hparams.n_embd_inp();
 
     auto inp = std::make_unique<llm_graph_input_embd>();
 
@@ -1274,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     //    return cur;
     //}
 
-    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
+    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
     const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
     cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
@@ -1587,9 +1592,10 @@ ggml_tensor * llm_graph_context::build_attn(
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
+    // expand k later to enable rope fusion which directly writes into k-v cache
     ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
+    ggml_build_forward_expand(gf, k_cur);
 
     const auto * mctx_cur = inp->mctx;
 
@@ -2030,7 +2036,7 @@ int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buck
 
     if (bidirectional) {
         relative_bucket += (relative_position > 0) * n_buckets;
-        relative_position = abs(relative_position);
+        relative_position = std::abs(relative_position);
     } else {
         relative_position = -std::min<int32_t>(relative_position, 0);
     }
diff --git a/llama/llama.cpp/src/llama-hparams.cpp b/llama/llama.cpp/src/llama-hparams.cpp
index b6bf6bbf..41127bf9 100644
--- a/llama/llama.cpp/src/llama-hparams.cpp
+++ b/llama/llama.cpp/src/llama-hparams.cpp
@@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_embd_inp() const {
+    uint32_t n_embd_inp = n_embd;
+
+    if (n_deepstack_layers > 0) {
+        n_embd_inp += n_embd * n_deepstack_layers;
+    }
+
+    return n_embd_inp;
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
@@ -148,7 +158,7 @@ bool llama_hparams::is_recurrent(uint32_t il) const {
 }
 
 uint32_t llama_hparams::n_pos_per_embd() const {
-    return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+    return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
 }
 
 bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
diff --git a/llama/llama.cpp/src/llama-hparams.h b/llama/llama.cpp/src/llama-hparams.h
index 24569a25..2ffe7dd3 100644
--- a/llama/llama.cpp/src/llama-hparams.h
+++ b/llama/llama.cpp/src/llama-hparams.h
@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
+#define LLAMA_MAX_EXPERTS 512 // Qwen3 Next
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE           = 0,
@@ -185,6 +185,9 @@ struct llama_hparams {
     std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
     std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
 
+    // qwen3vl deepstack
+    uint32_t n_deepstack_layers = 0;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -226,6 +229,9 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    // dimension of main + auxiliary input embeddings
+    uint32_t n_embd_inp() const;
+
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
 
diff --git a/llama/llama.cpp/src/llama-impl.cpp b/llama/llama.cpp/src/llama-impl.cpp
index 6ec709dd..c7a1880a 100644
--- a/llama/llama.cpp/src/llama-impl.cpp
+++ b/llama/llama.cpp/src/llama-impl.cpp
@@ -20,10 +20,10 @@ static llama_logger_state g_logger_state;
 time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
 
 time_meas::~time_meas() {
-        if (t_start_us >= 0) {
-            t_acc += ggml_time_us() - t_start_us;
-        }
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
     }
+}
 
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);
diff --git a/llama/llama.cpp/src/llama-kv-cache-iswa.cpp b/llama/llama.cpp/src/llama-kv-cache-iswa.cpp
index facba1d0..3a34102a 100644
--- a/llama/llama.cpp/src/llama-kv-cache-iswa.cpp
+++ b/llama/llama.cpp/src/llama-kv-cache-iswa.cpp
@@ -45,7 +45,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
 
     const uint32_t size_base = kv_size;
 
-    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
+    // note: the SWA cache is always padded to 256 for performance
+    //       https://github.com/ggml-org/llama.cpp/issues/17037
+    uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
 
     // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
     if (swa_full) {
diff --git a/llama/llama.cpp/src/llama-kv-cache.cpp b/llama/llama.cpp/src/llama-kv-cache.cpp
index 736693e1..e26385a1 100644
--- a/llama/llama.cpp/src/llama-kv-cache.cpp
+++ b/llama/llama.cpp/src/llama-kv-cache.cpp
@@ -8,6 +8,7 @@
 #include <algorithm>
 #include <cassert>
 #include <cmath>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -37,8 +38,15 @@ llama_kv_cache::llama_kv_cache(
 
     const uint32_t n_layer_kv = hparams.n_layer_kv();
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -53,13 +61,12 @@ llama_kv_cache::llama_kv_cache(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     GGML_ASSERT(n_stream == 1 || n_stream == n_seq_max);
@@ -167,11 +174,8 @@ llama_kv_cache::llama_kv_cache(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -179,7 +183,7 @@ llama_kv_cache::llama_kv_cache(
         LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
 
         ggml_backend_buffer_clear(buf, 0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -203,7 +207,7 @@ void llama_kv_cache::clear(bool data) {
     }
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -334,6 +338,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             llama_pos pos   = v_cells[s0].pos_get(i);
             llama_pos shift = v_cells[s0].get_shift(i);
 
+            llama_kv_cell_ext ext = v_cells[s0].ext_get(i);
+
             if (shift != 0) {
                 pos -= shift;
                 assert(pos >= 0);
@@ -345,6 +351,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             if (shift != 0) {
                 v_cells[s1].pos_add(i, shift);
             }
+
+            v_cells[s1].ext_set(i, ext);
         }
     }
 
@@ -379,6 +387,7 @@ void llama_kv_cache::seq_keep(llama_seq_id seq_id) {
 
 void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_add() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
     auto & head  = v_heads[seq_to_stream[seq_id]];
@@ -423,6 +432,7 @@ void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, ll
 
 void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_div() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
 
@@ -472,8 +482,8 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -896,6 +906,14 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &
 
             cells.pos_set(idx, ubatch.pos[i]);
 
+            if (ubatch.is_pos_2d()) {
+                llama_kv_cell_ext ext {
+                    /*.x =*/ ubatch.pos[i + ubatch.n_tokens*2],
+                    /*.y =*/ ubatch.pos[i + ubatch.n_tokens],
+                };
+                cells.ext_set(idx, ext);
+            }
+
             for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) {
                 cells.seq_add(idx, ubatch.seq_id[i][s]);
             }
@@ -957,10 +975,14 @@ bool llama_kv_cache::get_has_shift() const {
 uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
     uint32_t result = 0;
 
+    // pad the n_kv value so that the graph remains constant across batches and can be reused
+    // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220)
+    const uint32_t n_pad_cur = std::max(n_pad, 256u);
+
     for (uint32_t s = 0; s < sinfo.n_stream(); ++s) {
         const auto & cells = v_cells[sinfo.strm[s]];
 
-        result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result);
+        result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result);
     }
 
     return result;
@@ -1239,6 +1261,11 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
 
                 const llama_pos p1 = ubatch->pos[i];
 
+                // for M-RoPE
+                const bool is_2d = ubatch->is_pos_2d();
+                const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
+                const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens]   : 0;
+
                 const uint64_t idst = n_kv*(h*n_stream*n_tps_pad + s*n_tps_pad + ii);
 
                 for (uint32_t j = 0; j < n_kv; ++j) {
@@ -1258,6 +1285,14 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
                         continue;
                     }
 
+                    // M-RoPE causal mask
+                    if (causal_attn && is_2d && p0 == p1) {
+                        const auto & p0_ext = cells.ext_get(j);
+                        if (p0_ext.is_2d_gt(p1_x, p1_y)) {
+                            continue;
+                        }
+                    }
+
                     // apply SWA if any
                     if (is_masked_swa(p0, p1)) {
                         continue;
@@ -1298,7 +1333,7 @@ void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch
 size_t llama_kv_cache::total_size() const {
     size_t size = 0;
 
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 
@@ -1340,7 +1375,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_beta_slow  = cparams.yarn_beta_slow;
 
     const auto & n_rot     = hparams.n_rot;
-    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE
+    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
                                 // @ngxson : this is a workaround
                                 // for M-RoPE, we want to rotate the whole vector when doing KV shift
                                 // a normal RoPE should work, we just need to use the correct ordering
@@ -1551,6 +1586,9 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t
             io.write(&pos,      sizeof(pos));
             io.write(&n_seq_id, sizeof(n_seq_id));
 
+            // TODO: we also need to save llama_kv_cell_ext when apply_ubatch() support loading it
+            //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
+
             for (const auto & seq_id : seq_ids) {
                 io.write(&seq_id, sizeof(seq_id));
             }
@@ -1696,6 +1734,8 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             return false;
         }
 
+        // TODO: we cannot yet restore llama_kv_cell_ext as the apply_ubatch() does not support it yet
+        //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
         apply_ubatch(sinfo, ubatch);
 
         const auto head_cur = sinfo.head();
@@ -2010,8 +2050,3 @@ void llama_kv_cache_context::set_input_kq_mask(ggml_tensor * dst, const llama_ub
 void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_pos_bucket(dst, ubatch);
 }
-
-uint32_t llama_kv_cache::get_padding(const llama_cparams & cparams) {
-    // the FA kernels require padding to avoid extra runtime boundary checks
-    return cparams.flash_attn ? 256u : 32u;
-}
diff --git a/llama/llama.cpp/src/llama-kv-cache.h b/llama/llama.cpp/src/llama-kv-cache.h
index 85f0663d..bf7821c0 100644
--- a/llama/llama.cpp/src/llama-kv-cache.h
+++ b/llama/llama.cpp/src/llama-kv-cache.h
@@ -19,8 +19,6 @@ struct llama_context;
 
 class llama_kv_cache : public llama_memory_i {
 public:
-    static uint32_t get_padding(const llama_cparams & cparams);
-
     struct stream_copy_info {
         bool empty() const {
             assert(ssrc.size() == sdst.size());
@@ -217,8 +215,8 @@ private:
     // this is the SWA type of the cache - not to be confused with the model SWA type
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
diff --git a/llama/llama.cpp/src/llama-kv-cells.h b/llama/llama.cpp/src/llama-kv-cells.h
index 8f6bf014..10063bf4 100644
--- a/llama/llama.cpp/src/llama-kv-cells.h
+++ b/llama/llama.cpp/src/llama-kv-cells.h
@@ -5,9 +5,27 @@
 
 #include <bitset>
 #include <cassert>
-#include <vector>
-#include <set>
+#include <cstring>
 #include <map>
+#include <set>
+#include <vector>
+
+struct llama_kv_cell_ext {
+    // 2D spatial positions, typically used for M-RoPE
+    llama_pos x = 0;
+    llama_pos y = 0;
+
+    // return true if the current 2D spatial position is greater than other
+    bool is_2d_gt(llama_pos ox, llama_pos oy) const {
+        return (y > oy) || (y == oy && x > ox);
+    }
+
+    void reset() {
+        static_assert(std::is_trivially_copyable_v<llama_kv_cell_ext>);
+
+        memset(this, 0, sizeof(*this));
+    }
+};
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
@@ -16,6 +34,7 @@ public:
     void reset() {
         for (uint32_t i = 0; i < pos.size(); ++i) {
             pos[i]   = -1;
+            ext[i].reset();
             shift[i] =  0;
             seq[i].reset();
         }
@@ -43,6 +62,7 @@ public:
 
     void resize(uint32_t n) {
         pos.resize(n);
+        ext.resize(n);
         shift.resize(n);
         seq.resize(n);
 
@@ -108,6 +128,7 @@ public:
             const auto idx = i + j;
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -126,6 +147,7 @@ public:
             const auto idx = idxs[j];
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -154,6 +176,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -184,6 +207,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -203,6 +227,7 @@ public:
         seq[i].reset();
 
         pos[i] = -1;
+        ext[i].reset();
         shift[i] = 0;
 
         used.erase(i);
@@ -221,6 +246,7 @@ public:
 
         if (seq[i].none()) {
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -250,6 +276,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -340,6 +367,13 @@ public:
         return pos[i];
     }
 
+    const llama_kv_cell_ext & ext_get(uint32_t i) const {
+        assert(i < pos.size());
+        assert(pos[i] != -1);
+
+        return ext[i];
+    }
+
     // note: call only if the cell is not empty
     llama_pos get_shift(uint32_t i) const {
         assert(i < pos.size());
@@ -368,6 +402,11 @@ public:
         used.insert(i);
     }
 
+    void ext_set(uint32_t i, llama_kv_cell_ext p) {
+        assert(i < ext.size());
+        ext[i] = p;
+    }
+
     // pos[i] = pos[i] + d
     // sets "has_shift" to true
     // note: call only if the cell is not empty
@@ -424,6 +463,9 @@ private:
 
     std::vector<llama_pos> pos;
 
+    // stores extra info per cell
+    std::vector<llama_kv_cell_ext> ext;
+
     // this array accumulates any applied shifts to the pos array since the last reset_shift() call
     // this is used to queue multiple updates to the pos array, which in the end can be applied in one go:
     //
diff --git a/llama/llama.cpp/src/llama-memory-recurrent.cpp b/llama/llama.cpp/src/llama-memory-recurrent.cpp
index d67f5a5f..812bf253 100644
--- a/llama/llama.cpp/src/llama-memory-recurrent.cpp
+++ b/llama/llama.cpp/src/llama-memory-recurrent.cpp
@@ -7,6 +7,7 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -32,8 +33,15 @@ llama_memory_recurrent::llama_memory_recurrent(
     cells.clear();
     cells.resize(mem_size);
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -48,13 +56,12 @@ llama_memory_recurrent::llama_memory_recurrent(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     r_l.resize(n_layer);
@@ -93,17 +100,14 @@ llama_memory_recurrent::llama_memory_recurrent(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for rs cache");
         }
         ggml_backend_buffer_clear(buf, 0);
         LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -129,7 +133,7 @@ void llama_memory_recurrent::clear(bool data) {
     used = 0;
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -147,7 +151,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    // models like Mamba or RWKV can't have a state partially erased
+    // models like Mamba or RWKV can't have a state partially erased at the end
+    // of the sequence because their state isn't preserved for previous tokens
     if (seq_id >= (int64_t) size) {
         // could be fatal
         return false;
@@ -156,8 +161,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
             const auto & cell = cells[tail_id];
-            // partial intersection is invalid
-            if ((0 < p0 && p0 < cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+            // partial intersection is invalid if it includes the final pos
+            if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
                 //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
                 return false;
             }
@@ -364,8 +369,8 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -662,7 +667,7 @@ bool llama_memory_recurrent::get_can_shift() const {
 
 size_t llama_memory_recurrent::total_size() const {
     size_t size = 0;
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 
diff --git a/llama/llama.cpp/src/llama-memory-recurrent.h b/llama/llama.cpp/src/llama-memory-recurrent.h
index 077c6e3c..47f01d73 100644
--- a/llama/llama.cpp/src/llama-memory-recurrent.h
+++ b/llama/llama.cpp/src/llama-memory-recurrent.h
@@ -109,8 +109,8 @@ private:
 
     const uint32_t n_seq_max = 1;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     size_t total_size() const;
 
diff --git a/llama/llama.cpp/src/llama-model.cpp b/llama/llama.cpp/src/llama-model.cpp
index 54621ea3..4468de2f 100644
--- a/llama/llama.cpp/src/llama-model.cpp
+++ b/llama/llama.cpp/src/llama-model.cpp
@@ -2,7 +2,6 @@
 
 #include "llama-impl.h"
 #include "llama-mmap.h"
-#include "llama-batch.h"
 #include "llama-cparams.h"
 #include "llama-model-loader.h"
 
@@ -13,9 +12,10 @@
 
 #include "ggml-cpp.h"
 
+#include "models/models.h"
+
 #include <algorithm>
 #include <cassert>
-#include <cmath>
 #include <cfloat>
 #include <cstring>
 #include <cmath>
@@ -83,6 +83,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_15B:           return "15B";
         case LLM_TYPE_16B:           return "16B";
         case LLM_TYPE_20B:           return "20B";
+        case LLM_TYPE_26B:           return "26B";
         case LLM_TYPE_27B:           return "27B";
         case LLM_TYPE_30B:           return "30B";
         case LLM_TYPE_32B:           return "32B";
@@ -121,6 +122,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
+        case LLM_TYPE_230B_A10B:     return "230B.A10B";
         case LLM_TYPE_235B_A22B:     return "235B.A22B";
         case LLM_TYPE_300B_A47B:     return "300B.A47B";
         case LLM_TYPE_355B_A32B:     return "355B.A32B";
@@ -274,8 +276,8 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_IM2COL:
             {
-                const int n_embd = hparams.n_embd;
-                ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd, w->ne[1], 1, 1);
+                const int n_embd_inp = hparams.n_embd_inp();
+                ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd_inp, w->ne[1], 1, 1);
                 op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16);
             } break;
         case GGML_OP_SCALE:
@@ -438,7 +440,7 @@ struct llama_model::impl {
     llama_mlocks mlock_mmaps;
 
     // contexts where the model tensors metadata is stored as well ass the corresponding buffers:
-    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
+    std::vector<std::pair<ggml_context_ptr, std::vector<ggml_backend_buffer_ptr>>> ctxs_bufs;
 
     buft_list_t cpu_buft_list;
     std::map<ggml_backend_dev_t, buft_list_t> gpu_buft_list;
@@ -693,6 +695,37 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_AFMOE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
+
+                // Set up interleaved sliding window attention (ISWA)
+                // Pattern: 3 sliding - 1 full (global_attn_every_n_layers = 4)
+                if (hparams.n_swa > 0) {
+                    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+                    hparams.set_swa_pattern(4);
+                } else {
+                    hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+                }
+
+                // Default to sigmoid if not set
+                if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+                    hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                }
+
+                switch (hparams.n_layer) {
+                    case 56: type = LLM_TYPE_6B; break;
+                    case 32: type = LLM_TYPE_26B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_DECI:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1002,6 +1035,18 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_RND1:
+            {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 48: type = LLM_TYPE_30B_A3B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+                // Set non-causal attention for diffusion models
+                hparams.causal_attn = false;
+            } break;
         case LLM_ARCH_QWEN2MOE:
             {
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
@@ -1026,10 +1071,34 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_QWEN3VL:
+            {
+                ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 28: type = LLM_TYPE_1_7B; break;
+                    case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break;
+                    case 64: type = LLM_TYPE_32B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_QWEN3MOE:
             {
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
 
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 48: type = LLM_TYPE_30B_A3B; break;
+                    case 94: type = LLM_TYPE_235B_A22B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_QWEN3VLMOE:
+            {
+                ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 48: type = LLM_TYPE_30B_A3B; break;
@@ -1535,7 +1604,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_DEEPSEEK2:
             {
-                bool is_lite = (hparams.n_layer == 27);
+                // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
+                bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
                 if (!is_lite) {
@@ -1869,7 +1939,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_embd) {
-                    case 1536: type = LLM_TYPE_7B_A1B; break;
+                    case 768: type = LLM_TYPE_350M; break;
+                    case 1536: type = (hparams.n_embd == 2048 ? LLM_TYPE_7B_A1B : LLM_TYPE_1B); break;
                     case 2048: case 2560: type = LLM_TYPE_3B; break;
                     case 4096: type = LLM_TYPE_32B; break;
                     default: type = LLM_TYPE_UNKNOWN;
@@ -2140,6 +2211,57 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MINIMAX_M2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,  hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,   hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,           hparams.expert_gating_func, false);
+
+                switch (hparams.n_layer) {
+                    case 62: type = LLM_TYPE_230B_A10B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_COGVLM:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 32: type = LLM_TYPE_13B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_PANGU_EMBED:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_1B; break; // openPangu-Embedded-1B-V1.1
+                    case 34: type = LLM_TYPE_7B; break; // openPangu-Embedded-7B-V1.1
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_QWEN3NEXT:
+            {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+
+                // Load linear attention (gated delta net) parameters
+                ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+                ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+                ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+                ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+                ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+                // Mark recurrent layers (linear attention layers)
+                for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+                    hparams.recurrent_layer_arr[i] = ((i + 1) % 4 != 0); // TODO: extract the magic 4 from "full_attention_interval"
+                }
+
+                switch (hparams.n_layer) {
+                    case 80: type = LLM_TYPE_80B_A3B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -2247,7 +2369,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
     struct ggml_backend_buft_comparator {
         bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
-            return ggml_backend_buft_name(lhs) < ggml_backend_buft_name(rhs);
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
         }
     };
     std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
@@ -3293,6 +3415,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_QWEN3:
+            case LLM_ARCH_QWEN3VL:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -3327,6 +3450,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_QWEN3MOE:
+            case LLM_ARCH_QWEN3VLMOE:
+            case LLM_ARCH_RND1:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -4507,7 +4632,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_DEEPSEEK2:
                 {
-                    const bool is_lite = (hparams.n_layer == 27);
+                    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
+                    const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
 
                     const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
 
@@ -5735,6 +5861,71 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_AFMOE:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    const int64_t n_ff_exp = hparams.n_ff_exp;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        // dual attention normalization
+                        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        // attention projections
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        // Q/K normalization
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+
+                        // attention gating
+                        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+
+                        // dual ffn normalization
+                        layer.ffn_norm      = create_tensor(tn(LLM_TENSOR_FFN_NORM,      "weight", i), {n_embd}, 0);
+                        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        if (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) {
+                            // MoE layers
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+                            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+
+                            // grouped expert weights
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+
+                            // shared expert
+                            if (n_expert_shared > 0) {
+                                const int64_t n_ff_shexp = n_ff_exp * n_expert_shared;
+                                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, 0);
+                                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+                            }
+                        } else {
+                            // Dense layers
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+                        }
+                    }
+                } break;
             case LLM_ARCH_ERNIE4_5:
             case LLM_ARCH_ERNIE4_5_MOE:
                 {
@@ -6007,9 +6198,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_LFM2:
             case LLM_ARCH_LFM2MOE:
                 {
-                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
-                    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
-                    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,          "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
 
                     if (output == NULL) {
                         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
@@ -6180,6 +6372,182 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), { n_embd_head_k }, TENSOR_NOT_REQUIRED);
                     }
                 } break;
+            case LLM_ARCH_MINIMAX_M2:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_gqa }, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_gqa }, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k * n_head}, 0);
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_k_gqa}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
+                        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+                    }
+                } break;
+            case LLM_ARCH_COGVLM:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd_head_k * n_head * 3}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.visexp_attn_wqkv = create_tensor(tn(LLM_TENSOR_VISEXP_ATTN_QKV, "weight", i), {n_embd, n_embd_head_k * n_head * 3}, 0);
+                        layer.visexp_attn_wo = create_tensor(tn(LLM_TENSOR_VISEXP_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+                        layer.visexp_ffn_gate = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.visexp_ffn_down = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.visexp_ffn_up   = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                    }
+                } break;
+            case LLM_ARCH_PANGU_EMBED:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        // weight tensors
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        // bias tensors
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd_head_k * n_head}, 0);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+                            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        } else {
+                            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        }
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                    }
+                } break;
+            case LLM_ARCH_QWEN3NEXT:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+                    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
+                    }
+
+                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+
+                    // Calculate dimensions from hyperparameters
+                    const int64_t head_k_dim = hparams.ssm_d_state;
+                    const int64_t head_v_dim = hparams.ssm_d_state;
+                    const int64_t n_k_heads  = hparams.ssm_n_group;
+                    const int64_t n_v_heads  = hparams.ssm_dt_rank;
+                    const int64_t key_dim    = head_k_dim * n_k_heads;
+                    const int64_t value_dim  = head_v_dim * n_v_heads;
+                    const int64_t conv_dim   = key_dim * 2 + value_dim;
+
+                    // Calculate projection sizes
+                    const int64_t qkvz_dim = key_dim * 2 + value_dim * 2;
+                    const int64_t ba_dim   = n_v_heads * 2;
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), { n_embd }, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0);
+
+                        if (!hparams.is_recurrent(i)) {
+                            // Attention layers
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), { n_embd, n_embd_k_gqa }, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), { n_embd, n_embd_v_gqa }, 0);
+                            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
+
+                            // Q/K normalization for attention layers
+                            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
+                            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
+                        } else {
+                            // Linear attention (gated delta net) specific tensors
+                            // Create tensors with calculated dimensions
+                            layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", i), { n_embd, qkvz_dim }, 0);
+                            layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
+                            layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   i), { hparams.ssm_dt_rank }, 0);
+                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A,                    i), { hparams.ssm_dt_rank }, 0);
+                            layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0);
+                            layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", i), { head_v_dim }, 0);
+                            layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", i), { value_dim, n_embd }, 0);
+                        }
+
+                        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert }, 0);
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
+
+                        // Shared experts
+                        layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0);
+                        layer.ffn_gate_shexp     = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP,     "weight", i), { n_embd, hparams.n_ff_shexp }, 0);
+                        layer.ffn_up_shexp       = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,       "weight", i), { n_embd, hparams.n_ff_shexp }, 0);
+                        layer.ffn_down_shexp     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP,     "weight", i), { hparams.n_ff_shexp, n_embd }, 0);
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -6229,7 +6597,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
         bool is_default_buft = buft == ggml_backend_dev_buffer_type(dev);
 
-        ggml_backend_buffer_t buf = nullptr;
+        std::vector<ggml_backend_buffer_ptr> bufs;
         if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 // only the mmap region containing the tensors in the model is mapped to the backend buffer
@@ -6242,15 +6610,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     continue;
                 }
                 const size_t max_size = ggml_get_max_tensor_size(ctx);
-                buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
+                ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
                 if (buf == nullptr) {
                     throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
                 }
+                bufs.emplace_back(buf);
                 buf_map.emplace(idx, buf);
             }
         }
         else {
-            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
@@ -6260,11 +6629,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 mlock_buf->init   (ggml_backend_buffer_get_base(buf));
                 mlock_buf->grow_to(ggml_backend_buffer_get_size(buf));
             }
+            bufs.emplace_back(buf);
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 buf_map.emplace(idx, buf);
             }
         }
-        pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), buf);
+        pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), std::move(bufs));
 
         for (auto & buf : buf_map) {
             // indicate that this buffer contains weights
@@ -6290,8 +6660,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     }
 
     // print memory requirements per buffer type
-    for (auto & [_, buf] : pimpl->ctxs_bufs) {
-        LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0);
+    for (auto & [_, bufs] : pimpl->ctxs_bufs) {
+        for (auto & buf: bufs) {
+            LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n",
+                __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0);
+        }
     }
 
     // populate tensors_by_name
@@ -6343,8 +6716,10 @@ size_t llama_model::n_devices() const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : pimpl->ctxs_bufs) {
-        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [_, bufs] : pimpl->ctxs_bufs) {
+        for (const auto & buf : bufs) {
+            ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+        }
     }
     return ret;
 }
@@ -6392,6 +6767,7 @@ void llama_model::print_info() const {
     if (!hparams.vocab_only) {
         LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
         LLAMA_LOG_INFO("%s: n_embd           = %u\n",     __func__, hparams.n_embd);
+        LLAMA_LOG_INFO("%s: n_embd_inp       = %u\n",     __func__, hparams.n_embd_inp());
         LLAMA_LOG_INFO("%s: n_layer          = %u\n",     __func__, hparams.n_layer);
         LLAMA_LOG_INFO("%s: n_head           = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head(il);    }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_head_kv        = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
@@ -6412,6 +6788,8 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_expert         = %u\n",     __func__, hparams.n_expert);
         LLAMA_LOG_INFO("%s: n_expert_used    = %u\n",     __func__, hparams.n_expert_used);
+        LLAMA_LOG_INFO("%s: n_expert_groups  = %d\n",     __func__, hparams.n_expert_groups);
+        LLAMA_LOG_INFO("%s: n_group_used     = %d\n",     __func__, hparams.n_group_used);
         LLAMA_LOG_INFO("%s: causal attn      = %d\n",     __func__, hparams.causal_attn);
         LLAMA_LOG_INFO("%s: pooling type     = %d\n",     __func__, hparams.pooling_type);
         LLAMA_LOG_INFO("%s: rope type        = %d\n",     __func__, hparams.rope_type);
@@ -6420,6 +6798,10 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
+        // MRoPE (Multi-axis Rotary Position Embedding) sections
+        if (const auto & s = hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) {
+            LLAMA_LOG_INFO("%s: mrope sections   = [%d, %d, %d, %d]\n", __func__, s[0], s[1], s[2], s[3]);
+        }
         if (!classifier_labels.empty()) {
             LLAMA_LOG_INFO("%s: n_cls_out        = %u\n", __func__, hparams.n_cls_out);
 
@@ -6436,6 +6818,7 @@ void llama_model::print_info() const {
         arch == LLM_ARCH_FALCON_H1 ||
         arch == LLM_ARCH_PLAMO2 ||
         arch == LLM_ARCH_GRANITE_HYBRID ||
+        arch == LLM_ARCH_QWEN3NEXT ||
         arch == LLM_ARCH_NEMOTRON_H) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
@@ -6485,7 +6868,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
     }
 
-    if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE) {
+    if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE || arch == LLM_ARCH_QWEN3VLMOE || arch == LLM_ARCH_RND1) {
         LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
     }
 
@@ -6512,8 +6895,6 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: n_ff_shexp           = %d\n",     __func__, hparams.n_ff_shexp);
         LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
-        LLAMA_LOG_INFO("%s: n_expert_groups      = %d\n",     __func__, hparams.n_expert_groups);
-        LLAMA_LOG_INFO("%s: n_group_used         = %d\n",     __func__, hparams.n_group_used);
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
         LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
         LLAMA_LOG_INFO("%s: expert_gating_func   = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
@@ -6618,13225 +6999,21 @@ float llama_model::get_rope_freq_scale(const llama_cparams & cparams, int il) co
 }
 
 ggml_tensor * llama_model::get_rope_factors(const llama_cparams & cparams, int il) const {
-    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+    const uint32_t n_ctx_seq = cparams.n_ctx_seq;
 
     // choose long/short freq factors based on the context size
     if (layers[il].rope_freqs != nullptr) {
         return layers[il].rope_freqs;
     }
 
-    if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
+    if (n_ctx_seq > hparams.n_ctx_orig_yarn) {
         return layers[il].rope_long;
     }
 
     return layers[il].rope_short;
 }
 
-struct llm_build_llama : public llm_graph_context {
-    llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                if (hparams.use_kq_norm) {
-                    // Llama4TextL2Norm
-                    Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
-                    Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
-                    cb(Qcur, "Qcur_normed", il);
-                    cb(Kcur, "Kcur_normed", il);
-                }
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network (non-MoE)
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-                cb(cur, "ffn_moe_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_llama_iswa : public llm_graph_context {
-    llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        // temperature tuning
-        ggml_tensor * inp_attn_scale = nullptr;
-        inp_attn_scale = build_inp_attn_scale();
-
-        auto * inp_attn = build_attn_inp_kv_iswa();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
-                                  (il + 1) % hparams.n_no_rope_layer_step != 0;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                if (use_rope) {
-                    Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                    Kcur = ggml_rope_ext(
-                            ctx0, Kcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-                } else if (inp_attn_scale) {
-                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                if (use_rope && hparams.use_kq_norm) {
-                    // Llama4TextL2Norm
-                    Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
-                    Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
-                    cb(Qcur, "Qcur_normed", il);
-                    cb(Kcur, "Kcur_normed", il);
-                }
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network (non-MoE)
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, false,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
-                        il);
-
-                // Shared experts
-                ggml_tensor * shexp_out = build_ffn(ffn_inp_normed,
-                    model.layers[il].ffn_up_shexp,   NULL, NULL,
-                    model.layers[il].ffn_gate_shexp, NULL, NULL,
-                    model.layers[il].ffn_down_shexp, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(shexp_out, "ffn_moe_shexp", il);
-
-                cur = ggml_add(ctx0, moe_out, shexp_out);
-                cb(cur, "ffn_moe_out_merged", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_deci : public llm_graph_context {
-    llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-            const int64_t n_head_kv = hparams.n_head_kv(il);
-            const int64_t n_head    = hparams.n_head(il);
-            const int64_t n_ff      = hparams.n_ff(il);
-
-            if (n_head == 0) {
-                // attention-free layer of Llama-3_1-Nemotron-51B
-                cur = inpL;
-            } else {
-                // norm
-                cur = build_norm(inpL,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "attn_norm", il);
-            }
-
-            if (n_head > 0 && n_head_kv == 0) {
-                // "linear attention" of Llama-3_1-Nemotron-51B
-                cur = build_lora_mm(model.layers[il].wo, cur);
-                cb(cur, "wo", il);
-            } else if (n_head > 0) {
-                // self-attention
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
-            if (n_ff == 0) {
-                continue;
-            }
-
-            // modified to support attention-free layer of Llama-3_1-Nemotron-51B
-            ggml_tensor * ffn_inp = cur;
-            if (n_head > 0) {
-                ffn_inp = ggml_add(ctx0, cur, inpSA);
-                cb(ffn_inp, "ffn_inp", il);
-            }
-
-            // feed-forward network
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_baichuan : public llm_graph_context {
-    llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                switch (model.type) {
-                    case LLM_TYPE_7B:
-                        Qcur = ggml_rope_ext(
-                                ctx0, Qcur, inp_pos, nullptr,
-                                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                ext_factor, attn_factor, beta_fast, beta_slow
-                                );
-                        Kcur = ggml_rope_ext(
-                                ctx0, Kcur, inp_pos, nullptr,
-                                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                ext_factor, attn_factor, beta_fast, beta_slow
-                                );
-                        break;
-                    case LLM_TYPE_13B:
-                        break;
-                    default:
-                        GGML_ABORT("fatal error");
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_xverse : public llm_graph_context {
-    llm_build_xverse(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_falcon : public llm_graph_context {
-    llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * attn_norm;
-
-            attn_norm = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(attn_norm, "attn_norm", il);
-
-            // self-attention
-            {
-                if (model.layers[il].attn_norm_2) {
-                    // Falcon-40B
-                    cur = build_norm(inpL,
-                            model.layers[il].attn_norm_2,
-                            model.layers[il].attn_norm_2_b,
-                            LLM_NORM, il);
-                    cb(cur, "attn_norm_2", il);
-                } else {
-                    cur = attn_norm;
-                }
-
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                // using mode = 2 for neox mode
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);
-                inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);
-                attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = cur;
-
-            // feed forward
-            {
-                cur = build_ffn(attn_norm, // !! use the attn norm, not the result
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        NULL,                      NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = ggml_add(ctx0, cur, inpL);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        // norm
-        cur = build_norm(cur,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_grok : public llm_graph_context {
-    llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].attn_out_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_out_norm", il);
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // MoE branch
-            ggml_tensor * moe_out = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_GELU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(moe_out, "ffn_moe_out", il);
-
-            if (model.layers[il].ffn_up) {
-                ggml_tensor * ffn_out = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, il);
-                cb(ffn_out, "ffn_out", il);
-
-                cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);
-                cb(cur, "ffn_out", il);
-            } else {
-                cur = moe_out;
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_post_norm", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
-
-        // final logit soft-capping
-        if (hparams.f_final_logit_softcapping) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
-            cur = ggml_tanh(ctx0, cur);
-            cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
-        }
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_dbrx : public llm_graph_context {
-    llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = nullptr;
-                ggml_tensor * Kcur = nullptr;
-                ggml_tensor * Vcur = nullptr;
-
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                cb(cur, "wqkv_clamped", il);
-
-                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].attn_out_norm, NULL,
-                    LLM_NORM, il);
-            cb(cur, "attn_out_norm", il);
-
-            cur = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_starcoder : public llm_graph_context {
-    llm_build_starcoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
-        cb(pos, "pos_embd", -1);
-
-        inpL = ggml_add(ctx0, inpL, pos);
-        cb(inpL, "inpL", -1);
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_refact : public llm_graph_context {
-    llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_bert : public llm_graph_context {
-    llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-        ggml_tensor * inp_pos = nullptr;
-
-        if (model.arch != LLM_ARCH_JINA_BERT_V2) {
-            inp_pos = build_inp_pos();
-        }
-
-        // construct input embeddings (token, type, position)
-        inpL = build_inp_embd(model.tok_embd);
-
-        // token types are hardcoded to zero ("Sentence A")
-        if (model.type_embd) {
-            ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
-            inpL = ggml_add(ctx0, inpL, type_row0);
-        }
-        if (model.arch == LLM_ARCH_BERT) {
-            inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
-        }
-        cb(inpL, "inp_embd", -1);
-
-        // embed layer norm
-        inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
-        cb(inpL, "inp_norm", -1);
-
-        auto * inp_attn = build_attn_inp_no_cache();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * cur = inpL;
-
-            {
-                ggml_tensor * Qcur;
-                ggml_tensor * Kcur;
-                ggml_tensor * Vcur;
-
-                // self-attention
-                if (model.layers[il].wqkv) {
-                    cur = build_lora_mm(model.layers[il].wqkv, cur);
-                    cb(cur, "wqkv", il);
-
-                    if (model.layers[il].bqkv) {
-                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                        cb(cur, "bqkv", il);
-                    }
-
-                    Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                    Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                    Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-                } else {
-                    Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
-                    Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
-                    Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-                }
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head*n_head, n_tokens);
-
-                    Qcur = build_norm(Qcur,
-                            model.layers[il].attn_q_norm,
-                            model.layers[il].attn_q_norm_b,
-                            LLM_NORM, il);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                }
-
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head*n_head_kv, n_tokens);
-
-                    Kcur = build_norm(Kcur,
-                            model.layers[il].attn_k_norm,
-                            model.layers[il].attn_k_norm_b,
-                            LLM_NORM, il);
-
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                }
-
-                // RoPE
-                if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || model.arch == LLM_ARCH_JINA_BERT_V3) {
-                    Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, nullptr,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                    Kcur = ggml_rope_ext(
-                            ctx0, Kcur, inp_pos, nullptr,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-                cb(cur, "kqv_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // re-add the layer input
-            cur = ggml_add(ctx0, cur, inpL);
-
-            // attention layer norm
-            cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);
-
-            if (model.layers[il].attn_norm_2 != nullptr) {
-                cur = ggml_add(ctx0, cur, inpL); // re-add the layer input
-                cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);
-            }
-
-            ggml_tensor * ffn_inp = cur;
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
-                // MoE branch
-                cur = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        nullptr,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        hparams.n_expert,
-                        hparams.n_expert_used,
-                        LLM_FFN_GELU,
-                        false, false,
-                        0.0f,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
-                cb(cur, "ffn_moe_out", il);
-            } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || model.arch == LLM_ARCH_JINA_BERT_V3) {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL,                        NULL,
-                        model.layers[il].ffn_gate, NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // attentions bypass the intermediate layer
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            // output layer norm
-            cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cb(cur, "result_embd", -1);
-        res->t_embd = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_neo_bert : public llm_graph_context {
-    llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        // construct input embeddings (token, type, position)
-        inpL = build_inp_embd(model.tok_embd);
-        cb(inpL, "inp_embd", -1);
-
-        auto * inp_attn = build_attn_inp_no_cache();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * cur = inpL;
-
-            // pre-norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-
-            {
-                ggml_tensor * Qcur;
-                ggml_tensor * Kcur;
-                ggml_tensor * Vcur;
-
-                // self-attention
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                // RoPE
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-                cb(cur, "kqv_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // re-add the layer input
-            cur = ggml_add(ctx0, cur, inpL);
-
-            ggml_tensor * ffn_inp = cur;
-            cb(ffn_inp, "ffn_inp", il);
-
-            // pre-norm
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,
-                    NULL, NULL, NULL, NULL, NULL,
-                    model.layers[il].ffn_down,
-                    NULL, NULL, NULL,
-                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
-
-            // attentions bypass the intermediate layer
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm_enc, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_embd", -1);
-        res->t_embd = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_bloom : public llm_graph_context {
-    llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        inpL = build_norm(inpL,
-                model.tok_norm,
-                model.tok_norm_b,
-                LLM_NORM, -1);
-        cb(inpL, "inp_norm", -1);
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // Add the input
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_mpt : public llm_graph_context {
-    llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * pos;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        if (model.pos_embd) {
-            // inp_pos - contains the positions
-            ggml_tensor * inp_pos = build_inp_pos();
-            pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
-            cb(pos, "pos_embd", -1);
-
-            inpL = ggml_add(ctx0, inpL, pos);
-            cb(inpL, "inpL", -1);
-        }
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * attn_norm;
-
-            attn_norm = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(attn_norm, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = attn_norm;
-
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                if (model.layers[il].bqkv){
-                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                    cb(cur, "bqkv", il);
-                }
-
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(cur, "wqkv_clamped", il);
-                }
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                // Q/K Layernorm
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head*n_head,    n_tokens);
-                    Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head*n_head_kv, n_tokens);
-
-                    Qcur = build_norm(Qcur,
-                            model.layers[il].attn_q_norm,
-                            model.layers[il].attn_q_norm_b,
-                            LLM_NORM, il);
-
-                    Kcur = build_norm(Kcur,
-                            model.layers[il].attn_k_norm,
-                            model.layers[il].attn_k_norm_b,
-                            LLM_NORM, il);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // Add the input
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed forward
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        model.layers[il].ffn_act,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_stablelm : public llm_graph_context {
-    llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            ggml_tensor * inpSA = cur;
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = build_norm(Qcur,
-                            model.layers[il].attn_q_norm,
-                            NULL,
-                            LLM_NORM, il);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = build_norm(Kcur,
-                            model.layers[il].attn_k_norm,
-                            NULL,
-                            LLM_NORM, il);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpL  = ggml_get_rows(ctx0,  inpL, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                if (model.layers[il].ffn_norm) {
-                    cur = build_norm(ffn_inp,
-                            model.layers[il].ffn_norm,
-                            model.layers[il].ffn_norm_b,
-                            LLM_NORM, il);
-                    cb(cur, "ffn_norm", il);
-                } else {
-                    // parallel residual
-                    cur = inpSA;
-                }
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_qwen : public llm_graph_context {
-    llm_build_qwen(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 2*sizeof(float)*(n_embd));
-
-                // using mode = 2 for neox mode
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward forward
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_qwen2 : public llm_graph_context {
-    llm_build_qwen2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        if (model.output_b != nullptr) {
-            cur = ggml_add(ctx0, cur, model.output_b);
-        }
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_dream : public llm_graph_context {
-    llm_build_dream(const llama_model & model, const llm_graph_params & params) :
-        llm_graph_context(params) {
-        //copied from qwen2
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_no_cache();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur               = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur               = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur               = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                     ext_factor, attn_factor, beta_fast, beta_slow);
-
-                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                     ext_factor, attn_factor, beta_fast, beta_slow);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL,
-                            model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_llada : public llm_graph_context {
-    llm_build_llada(const llama_model & model, const llm_graph_params & params) :
-        llm_graph_context(params) {
-        // LLaDA is similar to LLaMA but uses non-causal attention for diffusion
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        // Non-causal attention for diffusion
-        auto * inp_attn = build_attn_inp_no_cache();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute separate Q, K, V projections without bias, matching LLaDALlamaBlock
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                     ext_factor, attn_factor, beta_fast, beta_slow);
-
-                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                     ext_factor, attn_factor, beta_fast, beta_slow);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL,
-                            model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_qwen2vl : public llm_graph_context {
-    llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        int sections[4];
-        std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_multi(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_multi(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_qwen2moe : public llm_graph_context {
-    llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            ggml_tensor * moe_out =
-                build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, false,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-            cb(moe_out, "ffn_moe_out", il);
-
-            // FFN shared expert
-            {
-                ggml_tensor * cur_gate_inp = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
-                cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
-
-                // sigmoid
-                ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
-                cb(cur_gate, "ffn_shexp_gate", il);
-
-                ggml_tensor * cur_ffn = build_ffn(cur,
-                        model.layers[il].ffn_up_shexp,   NULL, NULL,
-                        model.layers[il].ffn_gate_shexp, NULL, NULL,
-                        model.layers[il].ffn_down_shexp, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur_ffn, "ffn_shexp", il);
-
-                ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
-                cb(ffn_shexp_out, "ffn_shexp_out", il);
-
-                moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
-                cb(moe_out, "ffn_out", il);
-
-                cur = moe_out;
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_qwen3 : public llm_graph_context {
-    llm_build_qwen3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_qwen3moe : public llm_graph_context {
-    llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            ggml_tensor * moe_out =
-                build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-            cb(moe_out, "ffn_moe_out", il);
-            cur = moe_out;
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_phi2 : public llm_graph_context {
-    llm_build_phi2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * attn_norm_output;
-        ggml_tensor * ffn_output;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            attn_norm_output = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(attn_norm_output, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = nullptr;
-                ggml_tensor * Kcur = nullptr;
-                ggml_tensor * Vcur = nullptr;
-
-                if (model.layers[il].wqkv) {
-                    cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output);
-                    cb(cur, "wqkv", il);
-
-                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                    cb(cur, "bqkv", il);
-
-                    Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                    Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                    Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-                } else {
-                    Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
-                    Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
-                    Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-                }
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                // with phi2, we scale the Q to avoid precision issues
-                // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
-                inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
-                attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
-            }
-
-            // FF
-            {
-                ffn_output = build_ffn(attn_norm_output,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(ffn_output, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_output);
-            cur = ggml_add(ctx0, cur, inpL);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output_no_bias", -1);
-
-        cur = ggml_add(ctx0, cur, model.output_b);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-template<bool iswa>
-struct llm_build_phi3 : public llm_graph_context {
-    llm_build_phi3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
-        inp_attn_type * inp_attn = nullptr;
-
-        if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_iswa();
-        } else {
-            inp_attn = build_attn_inp_kv();
-        }
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            auto * residual = inpL;
-
-            // self-attention
-            {
-                // rope freq factors for 128k context
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                ggml_tensor* attn_norm_output = build_norm(inpL,
-                        model.layers[il].attn_norm,
-                        model.layers[il].attn_norm_b,
-                        LLM_NORM_RMS, il);
-                cb(attn_norm_output, "attn_norm", il);
-
-                ggml_tensor * Qcur = nullptr;
-                ggml_tensor * Kcur = nullptr;
-                ggml_tensor * Vcur = nullptr;
-
-                if (model.layers[il].wqkv) {
-                    cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output);
-                    cb(cur, "wqkv", il);
-
-                    Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head * sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd));
-                    Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd));
-                    Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
-                } else {
-                    Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
-                    Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
-                    Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-                }
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
-                cb(Qcur, "Qcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
-                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
-            }
-
-            cur = ggml_add(ctx0, cur, residual);
-            residual = cur;
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        NULL,                      NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                cur = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-                cb(cur, "ffn_moe_out", il);
-            }
-
-            cur = ggml_add(ctx0, residual, cur);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        if (model.output_b != nullptr) {
-            cb(cur, "result_output_no_bias", -1);
-            cur = ggml_add(ctx0, cur, model.output_b);
-        }
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_plamo : public llm_graph_context {
-    llm_build_plamo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            ggml_tensor * sa_inp = cur;
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur    = ggml_get_rows(ctx0,    cur, inp_out_ids);
-                sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids);
-                inpL   = ggml_get_rows(ctx0,   inpL, inp_out_ids);
-            }
-
-            ggml_tensor * sa_out = cur;
-
-            cur = sa_inp;
-
-            // feed-forward network
-            {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, sa_out);
-            cur = ggml_add(ctx0, cur, inpL);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_gpt2 : public llm_graph_context {
-    llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * pos;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
-        cb(pos, "pos_embd", -1);
-
-        inpL = ggml_add(ctx0, inpL, pos);
-        cb(inpL, "inpL", -1);
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_codeshell : public llm_graph_context {
-    llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_orion : public llm_graph_context {
-    llm_build_orion(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                // if (model.layers[il].bq) {
-                //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                //     cb(Qcur, "Qcur", il);
-                // }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                // if (model.layers[il].bk) {
-                //     Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                //     cb(Kcur, "Kcur", il);
-                // }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                // if (model.layers[il].bv) {
-                //     Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                //     cb(Vcur, "Vcur", il);
-                // }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_internlm2 : public llm_graph_context {
-    llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_minicpm3 : public llm_graph_context {
-    llm_build_minicpm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        //TODO: if the model varies, these parameters need to be read from the model
-        const int64_t n_embd_base = 256;
-        const float scale_embd  = 12.0f;
-        const float scale_depth = 1.4f;
-        const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k));
-
-        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
-        const uint32_t kv_lora_rank = hparams.n_lora_kv;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // scale the input embeddings
-        inpL = ggml_scale(ctx0, inpL, scale_embd);
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                ggml_tensor * q = NULL;
-                // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
-                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
-                cb(q, "q", il);
-
-                q = build_norm(q,
-                        model.layers[il].attn_q_a_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(q, "q", il);
-
-                // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
-                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
-                cb(q, "q", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        0);
-                cb(q_nope, "q_nope", il);
-
-                // and {n_head * n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        ggml_row_size(q->type, n_embd_head_qk_nope));
-                cb(q_pe, "q_pe", il);
-
-                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
-
-                // split into {kv_lora_rank, n_tokens}
-                ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        0);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // and {n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        kv_pe_compresseed->nb[1],
-                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
-                cb(k_pe, "k_pe", il);
-
-                kv_compressed = build_norm(kv_compressed,
-                        model.layers[il].attn_kv_a_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
-                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
-                cb(kv, "kv", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        0);
-                cb(k_nope, "k_nope", il);
-
-                // and {n_head * n_embd_head_v, n_tokens}
-                ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_cont(ctx0, v_states);
-                cb(v_states, "v_states", il);
-
-                q_pe = ggml_rope_ext(
-                        ctx0, q_pe, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-                cb(q_pe, "q_pe", il);
-
-                // shared RoPE key
-                k_pe = ggml_rope_ext(
-                        ctx0, k_pe, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-                cb(k_pe, "k_pe", il);
-
-                ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
-                cb(q_states, "q_states", il);
-
-                ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
-                cb(k_states, "k_states", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // scale_res - scale the hidden states for residual connection
-            const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct?
-            cur = ggml_scale(ctx0, cur, scale_res);
-            cb(cur, "hidden_scaled", il);
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // scale the hidden states for residual connection
-            cur = ggml_scale(ctx0, cur, scale_res);
-            cb(cur, "hidden_scaled_ffn", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head scaling
-        const float scale_lmhead = float(n_embd_base)/float(n_embd);
-        cur = ggml_scale(ctx0, cur, scale_lmhead);
-        cb(cur, "lmhead_scaling", -1);
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_gemma : public llm_graph_context {
-    llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
-                cb(Qcur, "Qcur_scaled", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-            cb(sa_out, "sa_out", il);
-
-            cur = build_norm(sa_out,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, sa_out);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_gemma2_iswa : public llm_graph_context {
-    llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_k;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv_iswa();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-            cb(sa_out, "sa_out", il);
-
-            cur = build_norm(sa_out,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_post_norm, NULL,
-                    LLM_NORM_RMS, -1);
-            cb(cur, "ffn_post_norm", -1);
-
-            cur = ggml_add(ctx0, cur, sa_out);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        // final logit soft-capping
-        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
-        cur = ggml_tanh(ctx0, cur);
-        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_gemma3_iswa : public llm_graph_context {
-    llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_k;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-        if (ubatch.token) {
-            inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-            cb(inpL, "inp_scaled", -1);
-        }
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        // TODO: is causal == true correct? might need some changes
-        auto * inp_attn = build_attn_inp_kv_iswa();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const float freq_base_l  = model.get_rope_freq_base (cparams, il);
-            const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
-
-            // norm
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
-                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-            cb(sa_out, "sa_out", il);
-
-            cur = build_norm(sa_out,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_post_norm, NULL,
-                    LLM_NORM_RMS, -1);
-            cb(cur, "ffn_post_norm", -1);
-
-            cur = ggml_add(ctx0, cur, sa_out);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_gemma3n_iswa : public llm_graph_context {
-    const llama_model & model;
-
-    const int64_t n_embd_head;
-    const int64_t n_embd_altup;
-    const int64_t n_altup;
-    const int     i_altup_act;
-    const int     n_layer_sparsity = 10; // number of layers using activation sparsity
-    const float   f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95)
-
-    llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params)
-            : llm_graph_context(params),
-              model(model),
-              n_embd_head(model.hparams.n_embd_head_k),
-              n_embd_altup(model.hparams.n_embd_altup),
-              n_altup(model.hparams.n_altup),
-              i_altup_act(model.hparams.i_altup_act) {
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-        if (ubatch.token) {
-            inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-            cb(inpL, "inp_scaled", -1);
-        }
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        // TODO: is causal == true correct? might need some changes
-        auto * inp_attn = build_attn_inp_kv_iswa();
-
-        // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
-        ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
-
-        // inpL now has only 1 altup, project it to the rest of the altups
-        // these "added" altups will be concat to the last dim of inpL
-        {
-            ggml_tensor * target_magnitude = calc_magnitude(inpL);
-            ggml_tensor * inp_repeated = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1);
-            ggml_tensor * altup_added = ggml_mul_mat(ctx0, model.altup_proj, inp_repeated); // shape: [n_embd, n_tokens, n_altup - 1]
-            ggml_tensor * new_magnitude = calc_magnitude(altup_added);
-            altup_added = ggml_div(ctx0,
-                                ggml_mul(ctx0, altup_added, target_magnitude),
-                                new_magnitude);
-            inpL = ggml_concat(ctx0, inpL, altup_added, 2); // shape: [n_embd, n_tokens, n_altup]
-            cb(inpL, "inp_stacked", -1);
-        }
-
-        // inpL now has shape:          [n_embd,       n_tokens, n_altup]
-        // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer]
-
-        for (int il = 0; il < n_layer; ++il) {
-            // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
-            const float freq_base_l  = model.get_rope_freq_base (cparams, il);
-            const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
-
-            ggml_tensor * cur = inpL; // [n_embd, n_tokens, n_altup]
-            ggml_tensor * predictions = altup_predict(cur, il); // [n_embd, n_tokens, n_altup]
-
-            // predicted value will go through self-attention and laurel
-            ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); // [n_embd, n_tokens]
-            cur = active_prediction;
-            cb(cur, "active_prediction", il);
-
-            // norm
-            cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // laurel
-            ggml_tensor * laurel_out = laurel(cur, il); // [n_embd, n_tokens]
-
-            // self-attention
-            if (hparams.has_kv(il)) {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
-
-                cb(Qcur, "Qcur_normed", il);
-                cb(Kcur, "Kcur_normed", il);
-                cb(Vcur, "Vcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                cb(Qcur, "Qcur_pos", il);
-                cb(Kcur, "Kcur_pos", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
-            } else {
-                // reuse KV cache of earlier layers
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Qcur, "Qcur_pos", il);
-
-                cur = build_attn(inp_attn,
-                    model.layers[il].wo, NULL,
-                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            cur = ggml_add(ctx0, cur, active_prediction); // [n_embd, n_tokens]
-            cb(cur, "attn_gated", il);
-
-            ggml_tensor * attn_laurel = ggml_scale(ctx0,
-                                            ggml_add(ctx0, cur, laurel_out),
-                                            1.0f / sqrtf(2.0f)); // [n_embd, n_tokens]
-            cb(attn_laurel, "attn_laurel", il);
-
-            cur = build_norm(attn_laurel,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                ggml_tensor * up_proj   = build_lora_mm(model.layers[il].ffn_up,   cur);
-                ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur);
-
-                if (il < n_layer_sparsity) {
-                    // apply activation sparsity
-                    gate_proj = gaussian_topk(gate_proj);
-                }
-                gate_proj = ggml_gelu(ctx0, gate_proj);
-
-                cur = ggml_mul(ctx0, up_proj, gate_proj);
-                cur = build_lora_mm(model.layers[il].ffn_down, cur);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_post_norm, NULL,
-                    LLM_NORM_RMS, -1);
-            cb(cur, "ffn_post_norm", il);
-
-            ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel); // [n_embd, n_tokens]
-            cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il);
-
-            ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il); // [n_embd, n_tokens, n_altup]
-
-            ggml_tensor * first_prediction; // [n_embd, n_tokens]
-            {
-                first_prediction = view_2d_slice(corrected, i_altup_act); // [n_embd, n_tokens]
-                first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale);
-                first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction);
-                first_prediction = ggml_gelu(ctx0, first_prediction); // [n_embd_altup, n_tokens]
-                cb(first_prediction, "first_prediction_gated", il);
-                ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il); // [n_embd_altup, n_tokens]
-                first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer); // [n_embd_altup, n_tokens]
-                cb(first_prediction, "first_prediction_scaled", il);
-
-                first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction); // [n_embd, n_tokens]
-                first_prediction = build_norm(first_prediction,
-                        model.layers[il].per_layer_post_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(first_prediction, "first_prediction_out", il);
-            }
-
-            // equivalent to python code: corrected_predictions[1:] += first_prediction
-            {
-                ggml_tensor * slice_first = view_2d_slice(corrected, 0);
-                ggml_tensor * slice_rest  = ggml_view_3d(ctx0, corrected, n_embd, n_tokens, n_altup - 1,
-                                                    ggml_row_size(corrected->type, n_embd),
-                                                    ggml_row_size(corrected->type, n_embd*n_tokens),
-                                                    n_embd*n_tokens*ggml_element_size(corrected));
-                ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction); // [n_embd, n_tokens, n_altup - 1]
-                corrected = ggml_concat(ctx0, slice_first, tmp, 2); // [n_embd, n_tokens, n_altup]
-            }
-
-            cur = corrected; // [n_embd, n_tokens, n_altup]
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL; // [n_embd, n_tokens, n_altup]
-
-        // cur now has multiple altup(s), we want to merge them back to 1 altup
-        {
-            ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act)); // [n_embd, n_tokens]
-            // do a view to skip the first slice (active altup)
-            ggml_tensor * alt_slice = ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1,
-                                                    ggml_row_size(cur->type, n_embd),
-                                                    ggml_row_size(cur->type, n_embd*n_tokens),
-                                                    n_embd*n_tokens*ggml_element_size(cur));
-            ggml_tensor * altup_unembd = ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice); // shape: [n_embd, n_tokens, n_altup - 1]
-            ggml_tensor * new_magnitude = calc_magnitude(altup_unembd);
-            altup_unembd = ggml_div(ctx0,
-                                ggml_mul(ctx0, altup_unembd, target_magnitude),
-                                new_magnitude);
-            cb(altup_unembd, "altup_unembd", -1);
-
-            // equivalent to torch.mean(hidden_states, dim=0)
-            cur = view_2d_slice(cur, 0); // [n_embd, n_tokens]
-            for (int i = 0; i < n_altup - 1; ++i) {
-                cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i));
-            }
-            cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup)); // [n_embd, n_tokens]
-            cb(cur, "unembd_merged", -1);
-        }
-
-        // cur now has shape: [n_embd, n_tokens]
-
-        // TODO: move this to right after the last KV layer
-        {
-            // skip computing output for unused tokens
-            ggml_tensor * inp_out_ids = build_inp_out_ids();
-            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-        }
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        {
-            // final logit soft-capping
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
-            cur = ggml_tanh(ctx0, cur);
-            cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
-        }
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-    ggml_tensor * calc_magnitude(ggml_tensor * x) {
-        return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x)));
-    }
-
-    // get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
-    ggml_tensor * view_2d_slice(ggml_tensor * x, int idx) {
-        GGML_ASSERT(idx < (int)x->ne[2]);
-        return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1],
-                            ggml_row_size(x->type, x->ne[0]),
-                            idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
-    }
-
-    // equivalent to get_per_layer_inputs() in python code
-    // output shape: [n_embd_altup, n_layer, n_tokens]
-    ggml_tensor * get_per_layer_inputs() {
-        auto inp = std::make_unique<llm_graph_input_embd>();
-        ggml_tensor * inp_per_layer;
-        if (ubatch.token) {
-            inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
-            ggml_set_input(inp->tokens);
-            res->t_tokens = inp->tokens;
-            inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
-            inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
-            inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float)n_embd_altup));
-            cb(inp_per_layer, "inp_per_layer_selected", -1);
-        } else {
-            GGML_ABORT("TODO: support embd input");
-        }
-        res->add_input(std::move(inp));
-        return inp_per_layer;
-    }
-
-    // equivalent to project_per_layer_inputs() in python code
-    // this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
-    // output shape: [n_embd_altup, n_tokens, n_layer]
-    ggml_tensor * project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) {
-        const float per_layer_projection_scale = 1.0f / sqrtf((float)n_embd);
-        const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
-
-        ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
-        per_layer_proj = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale);
-        per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens);
-        per_layer_proj = build_norm(per_layer_proj,
-                                    model.per_layer_proj_norm, NULL,
-                                    LLM_NORM_RMS, -1); // [n_embd_altup, n_layer, n_tokens]
-        cb(per_layer_proj, "per_layer_proj", -1);
-
-        inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
-        inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
-        cb(inp_per_layer, "inp_per_layer", -1);
-
-        // permute to shape: [n_embd_altup, n_tokens, n_layer]
-        inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
-        return inp_per_layer;
-    }
-
-    // input cur shape: [n_altup, n_tokens]
-    // output    shape: [n_altup, n_tokens]
-    ggml_tensor * laurel(ggml_tensor * cur, int il) {
-        ggml_tensor * tmp = cur;
-        tmp = build_lora_mm(model.layers[il].laurel_l, tmp);
-        tmp = build_lora_mm(model.layers[il].laurel_r, tmp);
-        tmp = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il);
-        tmp = ggml_add(ctx0, tmp, cur);
-        cb(tmp, "laurel_out", il);
-        return tmp;
-    }
-
-    // input x shape: [n_embd, n_tokens]
-    // output  shape: [n_embd, n_tokens]
-    ggml_tensor * gaussian_topk(ggml_tensor * x) {
-        ggml_tensor * mean = ggml_mean(ctx0, x);
-        ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0,
-            ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
-            1.0f / (float)(x->ne[0] - 1)
-        ));
-        ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul));
-        return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x));
-    }
-
-    //
-    // altup functions
-    //
-
-    // equivalent to compute_router_modalities() in python code
-    // input x shape: [n_embd,  n_tokens]
-    // output  shape: [n_altup, n_tokens]
-    ggml_tensor * altup_compute_router_modalities(ggml_tensor * x, int il) {
-        ggml_tensor * router_inputs = build_norm(x,
-            model.layers[il].altup_router_norm, NULL,
-            LLM_NORM_RMS, il);
-
-        // router_input_scale
-        router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float)n_embd);
-
-        ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs);
-        return ggml_tanh(ctx0, output); // [n_altup, n_tokens]
-    }
-
-    // input cur shape: [n_embd, n_tokens, n_altup]
-    // output    shape: [n_embd, n_tokens, n_altup]
-    ggml_tensor * altup_predict(ggml_tensor * cur, int il) {
-        ggml_tensor * activated = view_2d_slice(cur, i_altup_act); // [n_embd, n_tokens]
-        ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens]
-        cb(modalities, "modalities", il);
-
-        ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities);
-        cb(all_coefs, "all_coefs", il);
-        // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor)
-        all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens);
-
-        // permute to [n_altup, n_embd, n_tokens]
-        ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
-        ggml_tensor * predictions = ggml_mul_mat(ctx0, cur_permuted, all_coefs); // [n_altup, n_embd, n_tokens]
-
-        // final shape must be the same as cur: [n_embd, n_tokens, n_altup]
-        predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3));
-        predictions = ggml_add(ctx0, predictions, cur);
-        cb(predictions, "predictions", il);
-
-        return predictions;
-    }
-
-    // input predictions       shape: [n_embd, n_tokens, n_altup]
-    // input activated         shape: [n_embd, n_tokens]
-    // output                  shape: [n_embd, n_tokens, n_altup]
-    ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
-        ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens]
-        cb(modalities, "modalities", il);
-
-        ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);
-        ggml_tensor * innovation = ggml_sub(ctx0, activated, active_prediction); // [n_embd, n_tokens]
-        cb(innovation, "innovation", il);
-
-        ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities); // [n_altup, n_tokens]
-        all_coefs = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f); // + 1.0
-        cb(all_coefs, "all_coefs", il);
-        all_coefs = ggml_transpose(ctx0, all_coefs); // [n_tokens, n_altup]
-        all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup]
-
-        innovation = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
-        ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs); // [n_embd, n_tokens, n_altup]
-        corrected = ggml_add(ctx0, corrected, predictions); // [n_embd, n_tokens, n_altup]
-        cb(corrected, "corrected", il);
-
-        return corrected;
-    }
-};
-
-struct llm_build_gemma_embedding : public llm_graph_context {
-    llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_k;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-        if (ubatch.token) {
-            inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-            cb(inpL, "inp_scaled", -1);
-        }
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_no_cache();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const float freq_base_l  = model.get_rope_freq_base (cparams, il);
-            const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
-
-            // norm
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
-                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-            cb(sa_out, "sa_out", il);
-
-            cur = build_norm(sa_out,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_post_norm, NULL,
-                    LLM_NORM_RMS, -1);
-            cb(cur, "ffn_post_norm", -1);
-
-            cur = ggml_add(ctx0, cur, sa_out);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-// TODO: move up next to build_starcoder
-struct llm_build_starcoder2 : public llm_graph_context {
-    llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    NULL,                      NULL,                        NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_graph_context_mamba : public llm_graph_context {
-    llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {}
-
-    ggml_tensor * build_mamba_layer(
-        llm_graph_input_rs * inp,
-               ggml_tensor * cur,
-         const llama_model & model,
-        const llama_ubatch & ubatch,
-                       int   il) {
-
-        const auto * mctx_cur = inp->mctx;
-
-        const auto kv_head = mctx_cur->get_head();
-
-        const auto & layer = model.layers[il];
-
-        const int64_t d_conv  = hparams.ssm_d_conv;
-        const int64_t d_inner = hparams.ssm_d_inner;
-        const int64_t d_state = hparams.ssm_d_state;
-        const int64_t dt_rank = hparams.ssm_dt_rank;
-        const int64_t n_head  = d_inner;
-        const int64_t head_dim = 1;
-        const int64_t n_seqs  = ubatch.n_seqs;
-        // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers)
-        const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms;
-
-        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
-
-        GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(ubatch.equal_seqs());
-        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
-
-        ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
-        ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
-
-        ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
-        conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs);
-
-        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
-        // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
-        ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur);
-        // split the above in two
-        // => {d_inner, n_seq_tokens, n_seqs}
-        ggml_tensor * x = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0);
-        ggml_tensor * z = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner*ggml_element_size(xz));
-
-        // conv
-        {
-            // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
-            ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
-
-            // copy last (d_conv - 1) columns back into the state cache
-            ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0]));
-
-            ggml_build_forward_expand(gf,
-                ggml_cpy(ctx0, last_conv,
-                    ggml_view_1d(ctx0, conv_states_all,
-                        (d_conv - 1)*(d_inner)*(n_seqs),
-                        kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all))));
-
-            // 1D convolution
-            // The equivalent is to make a self-overlapping view of conv_x
-            // over d_conv columns at each stride in the 3rd dimension,
-            // then element-wise multiply that with the conv1d weight,
-            // then sum the elements of each row,
-            // (the last two steps are a dot product over rows (also doable with mul_mat))
-            // then permute away the ne[0] dimension,
-            // and then you're left with the resulting x tensor.
-            // For simultaneous sequences, all sequences need to have the same length.
-            x = ggml_ssm_conv(ctx0, conv_x, layer.ssm_conv1d);
-
-            // bias
-            x = ggml_add(ctx0, x, layer.ssm_conv1d_b);
-
-            x = ggml_silu(ctx0, x);
-        }
-
-        // ssm
-        {
-            // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}
-            ggml_tensor * x_db = build_lora_mm(layer.ssm_x, x);
-            // split
-            ggml_tensor * dt = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0);
-            ggml_tensor * B  = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank);
-            ggml_tensor * C  = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state));
-
-            // Some Mamba variants (e.g. FalconMamba, Jamba) apply RMS norm in B, C & Dt layers
-            if (ssm_dt_b_c_rms || (layer.ssm_dt_norm && layer.ssm_b_norm && layer.ssm_c_norm)) {
-                dt = build_norm(dt, layer.ssm_dt_norm, NULL, LLM_NORM_RMS, il);
-                B  = build_norm(B,  layer.ssm_b_norm,  NULL, LLM_NORM_RMS, il);
-                C  = build_norm(C,  layer.ssm_c_norm,  NULL, LLM_NORM_RMS, il);
-            }
-
-            // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
-            dt = build_lora_mm(layer.ssm_dt, dt);
-            dt = ggml_add(ctx0, dt, layer.ssm_dt_b);
-
-            cur = x;
-            x = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs);
-
-            ggml_tensor * A = layer.ssm_a;
-
-            // use the states and the indices provided by build_recurrent_state
-            // (this is necessary in order to properly use the states before they are overwritten,
-            //  while avoiding to make unnecessary copies of the states)
-            auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
-                ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
-
-                // Custom operator to optimize the parallel associative scan
-                // as described in the Annex D of the Mamba paper.
-                // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
-                return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
-            };
-
-            ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
-
-            // store last states
-            ggml_build_forward_expand(gf,
-                ggml_cpy(ctx0,
-                    ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]*x->ne[3]),
-                    ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all))));
-
-            ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0);
-
-            // TODO: skip computing output earlier for unused tokens
-
-            y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, layer.ssm_d));
-            y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
-
-            // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
-            cur = build_lora_mm(layer.ssm_out, y);
-        }
-
-        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
-
-        return cur;
-    }
-
-    ggml_tensor * build_mamba2_layer(
-        llm_graph_input_rs * inp,
-               ggml_tensor * cur,
-         const llama_model & model,
-        const llama_ubatch & ubatch,
-                       int   il) const {
-
-        const auto * mctx_cur = inp->mctx;
-
-        const auto kv_head = mctx_cur->get_head();
-
-        const int64_t d_conv  = hparams.ssm_d_conv;
-        const int64_t d_inner = hparams.ssm_d_inner;
-        const int64_t d_state = hparams.ssm_d_state;
-        const int64_t n_head  = hparams.ssm_dt_rank;
-        const int64_t head_dim = d_inner / n_head;
-        const int64_t n_group = hparams.ssm_n_group;
-        const int64_t n_seqs  = ubatch.n_seqs;
-
-        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
-
-        GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(ubatch.equal_seqs());
-        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
-
-        ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
-        ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
-
-        ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
-        conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs);
-
-        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
-        // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
-
-        // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
-        ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur);
-
-        // split the above in three
-        ggml_tensor * z = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*zxBCdt->nb[0], zxBCdt->nb[1], zxBCdt->nb[2], 0);
-        ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2*n_group*d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], d_inner*ggml_element_size(zxBCdt));
-        ggml_tensor * dt = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], (2*d_inner + 2*n_group*d_state)*ggml_element_size(zxBCdt));
-
-        // conv
-        {
-            // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs}
-            ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0);
-
-            // copy last (d_conv - 1) columns back into the state cache
-            ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0]));
-
-            ggml_build_forward_expand(gf,
-                ggml_cpy(ctx0, last_conv,
-                    ggml_view_1d(ctx0, conv_states_all,
-                        (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs),
-                        kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all))));
-
-            // 1D convolution
-            // The equivalent is to make a self-overlapping view of conv_x
-            // over d_conv columns at each stride in the 3rd dimension,
-            // then element-wise multiply that with the conv1d weight,
-            // then sum the elements of each row,
-            // (the last two steps are a dot product over rows (also doable with mul_mat))
-            // then permute away the ne[0] dimension,
-            // and then you're left with the resulting x tensor.
-            // For simultaneous sequences, all sequences need to have the same length.
-            xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
-
-            // bias
-            xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b);
-
-            xBC = ggml_silu(ctx0, xBC);
-        }
-
-        // ssm
-        {
-            // These correspond to V K Q in SSM/attention duality
-            ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*xBC->nb[0], xBC->nb[1], xBC->nb[2], 0);
-            ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], d_inner*ggml_element_size(xBC));
-            ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], (d_inner + n_group*d_state)*ggml_element_size(xBC));
-
-            // {n_head, n_seq_tokens, n_seqs}
-            dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b);
-
-            ggml_tensor * A = model.layers[il].ssm_a;
-
-            // use the states and the indices provided by build_recurrent_state
-            // (this is necessary in order to properly use the states before they are overwritten,
-            //  while avoiding to make unnecessary copies of the states)
-            auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
-                ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
-
-                // TODO: use semistructured matrices to implement state-space duality
-                // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
-                return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
-            };
-
-            ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
-
-            // store last states
-            ggml_build_forward_expand(gf,
-                ggml_cpy(ctx0,
-                    ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, ggml_nelements(x)*x->nb[0]),
-                    ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all))));
-
-            ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head*x->nb[1], n_seq_tokens*n_head*x->nb[1], 0);
-
-            // TODO: skip computing output earlier for unused tokens
-
-            y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
-            cb(y, "mamba2_y_add_d", il);
-            y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
-
-            // grouped RMS norm
-            if (model.layers[il].ssm_norm) {
-                y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs);
-                y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
-            }
-
-            y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
-
-            // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
-            cur = build_lora_mm(model.layers[il].ssm_out, y);
-        }
-
-        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
-        cb(cur, "mamba_out", il);
-
-        return cur;
-    }
-};
-
-struct llm_build_mamba : public llm_graph_context_mamba {
-    llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        // {n_embd, n_tokens}
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * rs_inp = build_rs_inp();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            if (model.arch == LLM_ARCH_MAMBA2) {
-                cur = build_mamba2_layer(rs_inp, cur, model, ubatch, il);
-            } else {
-                cur = build_mamba_layer(rs_inp, cur, model, ubatch, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // residual
-            cur = ggml_add(ctx0, cur, inpL);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        // final rmsnorm
-        cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-};
-
-struct llm_build_jamba : public llm_graph_context_mamba {
-    llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        // {n_embd, n_tokens}
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * inp_hybrid = build_inp_mem_hybrid();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const int64_t n_head_kv = hparams.n_head_kv(il);
-
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            if (n_head_kv == 0) {
-                cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);
-            } else {
-                // Attention
-
-                struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                // No RoPE :)
-                cur = build_attn(inp_hybrid->get_attn(),
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // residual
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur);
-            cb(cur, "ffn_inp", il);
-
-            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-                // FFN
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                cur = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, false,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-                cb(cur, "ffn_moe_out", il);
-            }
-
-            // residual
-            cur = ggml_add(ctx0, ffn_inp, cur);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        // final rmsnorm
-        cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_command_r : public llm_graph_context {
-    llm_build_command_r(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        const float f_logit_scale = hparams.f_logit_scale;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            ggml_tensor * ffn_inp = cur;
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = build_norm(Qcur,
-                            model.layers[il].attn_q_norm,
-                            NULL,
-                            LLM_NORM, il);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = build_norm(Kcur,
-                            model.layers[il].attn_k_norm,
-                            NULL,
-                            LLM_NORM, il);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur     = ggml_get_rows(ctx0,     cur, inp_out_ids);
-                inpL    = ggml_get_rows(ctx0,    inpL, inp_out_ids);
-                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
-            }
-
-            ggml_tensor * attn_out = cur;
-
-            // feed-forward network
-            {
-                cur = build_ffn(ffn_inp,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // add together residual + FFN + self-attention
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = ggml_add(ctx0, cur, attn_out);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        if (f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, f_logit_scale);
-        }
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_cohere2_iswa : public llm_graph_context {
-    llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        const float f_logit_scale = hparams.f_logit_scale;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv_iswa();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const bool is_swa = hparams.is_swa(il);
-
-            // norm
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-            ggml_tensor * ffn_inp = cur;
-
-            // self-attention
-            {
-                // rope freq factors for 128k context
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                if (is_swa) {
-                    Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                    Kcur = ggml_rope_ext(
-                            ctx0, Kcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
-                inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
-                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
-            }
-
-            ggml_tensor * attn_out = cur;
-
-            // feed-forward network
-            {
-                cur = build_ffn(ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate,
-                        NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR,
-                        il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // add together residual + FFN + self-attention
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = ggml_add(ctx0, cur, attn_out);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        if (f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, f_logit_scale);
-        }
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-// ref: https://allenai.org/olmo
-// based on the original build_llama() function, changes:
-//   * non-parametric layer norm
-//   * clamp qkv
-//   * removed bias
-//   * removed MoE
-struct llm_build_olmo : public llm_graph_context {
-    llm_build_olmo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    NULL, NULL,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    NULL, NULL,
-                    LLM_NORM, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                NULL, NULL,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-template <bool iswa>
-struct llm_build_olmo2 : public llm_graph_context {
-    llm_build_olmo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
-        inp_attn_type * inp_attn = nullptr;
-
-        if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_iswa();
-        } else {
-            inp_attn = build_attn_inp_kv();
-        }
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = inpL;
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                const bool is_swa = hparams.is_swa(il);
-
-                if (is_swa) {
-                    // For sliding window layers, Olmo3 use regular rope with no yarn rope scaling.
-                    // This is achieved here by setting freq_scale and attn_factor to 1.
-                    // We also set ext_factor to 0 to avoid a few unnecessary computations.
-                    Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
-                        0.0, 1.0, beta_fast, beta_slow
-                        );
-
-                    Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
-                        0.0, 1.0, beta_fast, beta_slow
-                        );
-                } else {
-                    Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                    Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_ffn(ffn_inp,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_post_norm, NULL,
-                    LLM_NORM_RMS, -1);
-            cb(cur, "ffn_post_norm", -1);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-// based on the build_qwen2moe() function, changes:
-//   * removed shared experts
-//   * removed bias
-//   * added q, k norm
-struct llm_build_olmoe : public llm_graph_context {
-    llm_build_olmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, false,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_llada_moe : public llm_graph_context {
-    llm_build_llada_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_no_cache();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, false,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_openelm : public llm_graph_context {
-    llm_build_openelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const int64_t n_head    = hparams.n_head(il);
-            const int64_t n_head_kv = hparams.n_head_kv(il);
-            const int64_t n_head_qkv = 2*n_head_kv + n_head;
-
-            cur = inpL;
-            ggml_tensor * residual = cur;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, cur->nb[1], cur->nb[2], 0);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv)));
-                cb(Vcur, "Vcur", il);
-
-                Qcur = build_norm(Qcur,
-                        model.layers[il].attn_q_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur", il);
-
-                Kcur = build_norm(Kcur,
-                        model.layers[il].attn_k_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, NULL,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, NULL,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Qcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
-                cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        // norm
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_gptneox : public llm_graph_context {
-    llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // ffn
-            if (hparams.use_par_res) {
-                // attention and ffn are computed in parallel
-                // x = x + attn(ln1(x)) + ffn(ln2(x))
-
-                ggml_tensor * attn_out = cur;
-
-                cur = build_norm(inpL,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-
-                cur = ggml_add(ctx0, cur, inpL);
-                cb(cur, "ffn_out", il);
-
-                cur = ggml_add(ctx0, cur, attn_out);
-
-                cur = build_cvec(cur, il);
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
-            } else {
-                // attention and ffn are computed sequentially
-                // x = x + attn(ln1(x))
-                // x = x + ffn(ln2(x))
-
-                ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-                cb(ffn_inp, "ffn_inp", il);
-
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-
-                cur = ggml_add(ctx0, cur, ffn_inp);
-
-                cur = build_cvec(cur, il);
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
-            }
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_arctic : public llm_graph_context {
-    llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
-            cb(ffn_out, "ffn_out", il);
-
-            // MoE
-            cur = build_norm(inpSA,
-                    model.layers[il].ffn_norm_exps, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm_exps", il);
-
-            cur = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_out);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_deepseek : public llm_graph_context {
-    llm_build_deepseek(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            if ((uint32_t) il < hparams.n_layer_dense_lead) {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                ggml_tensor * moe_out =
-                    build_moe_ffn(cur,
-                            model.layers[il].ffn_gate_inp,
-                            model.layers[il].ffn_up_exps,
-                            model.layers[il].ffn_gate_exps,
-                            model.layers[il].ffn_down_exps,
-                            nullptr,
-                            n_expert, n_expert_used,
-                            LLM_FFN_SILU, false,
-                            false, hparams.expert_weights_scale,
-                            LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                            il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                // FFN shared expert
-                {
-                    ggml_tensor * ffn_shexp = build_ffn(cur,
-                            model.layers[il].ffn_up_shexp,   NULL, NULL,
-                            model.layers[il].ffn_gate_shexp, NULL, NULL,
-                            model.layers[il].ffn_down_shexp, NULL, NULL,
-                            NULL,
-                            LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                    cb(cur, "ffn_out", il);
-                }
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_deepseek2 : public llm_graph_context {
-    llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        bool is_lite = (hparams.n_layer == 27);
-
-        const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
-
-        // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
-        const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
-        const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
-
-        const int64_t n_embd_head_qk_rope = hparams.n_rot;
-        const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
-
-        const uint32_t kv_lora_rank = hparams.n_lora_kv;
-
-        // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
-        // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
-        const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
-        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k));
-        const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        // {n_embd, n_tokens}
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                ggml_tensor * q = NULL;
-                if (!is_lite) {
-                    q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
-                    cb(q, "q", il);
-
-                    q = build_norm(q,
-                            model.layers[il].attn_q_a_norm, nullptr,
-                            LLM_NORM_RMS, il);
-                    cb(q, "q", il);
-
-                    q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
-                    cb(q, "q", il);
-                } else {
-                    q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                    cb(q, "q", il);
-                }
-
-                // split into {n_embd_head_qk_nope, n_head, n_tokens}
-                ggml_tensor * q_nope = ggml_view_3d(ctx0, q,
-                        n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(q->type, n_embd_head_k),
-                        ggml_row_size(q->type, n_embd_head_k) * n_head,
-                        0);
-                cb(q_nope, "q_nope", il);
-
-                // and {n_embd_head_qk_rope, n_head, n_tokens}
-                ggml_tensor * q_pe = ggml_view_3d(ctx0, q,
-                        n_embd_head_qk_rope, n_head, n_tokens,
-                        ggml_row_size(q->type, n_embd_head_k),
-                        ggml_row_size(q->type, n_embd_head_k) * n_head,
-                        ggml_row_size(q->type, n_embd_head_qk_nope));
-                cb(q_pe, "q_pe", il);
-
-                ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_cmpr_pe, "kv_cmpr_pe", il);
-
-                // split into {kv_lora_rank, n_tokens}
-                ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe,
-                        kv_lora_rank, n_tokens,
-                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
-                        0);
-                cb(kv_cmpr, "kv_cmpr", il);
-
-                // and {n_embd_head_qk_rope, 1, n_tokens}
-                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe,
-                        n_embd_head_qk_rope, 1, n_tokens,
-                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
-                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
-                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
-                cb(k_pe, "k_pe", il);
-
-                q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(q_pe, "q_pe", il);
-
-                k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(k_pe, "k_pe", il);
-
-                kv_cmpr = build_norm(kv_cmpr,
-                        model.layers[il].attn_kv_a_norm, nullptr,
-                        LLM_NORM_RMS, il);
-                cb(kv_cmpr, "kv_cmpr", il);
-
-                if (is_mla) {
-                    // {n_embd_head_qk_nope, n_tokens, n_head}
-                    q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
-                    cb(q_nope, "q_nope_perm", il);
-
-                    // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
-                    ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
-                    cb(q_nope_absorbed, "q_nope_absorbed", il);
-
-                    // {kv_lora_rank, n_head, n_tokens}
-                    q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
-                    cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
-
-                    // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
-                    // note: rope must go first for in-place context shifting in build_rope_shift()
-                    ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
-                    cb(Qcur, "Qcur", il);
-
-                    kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
-                    cb(kv_cmpr, "kv_cmpr_reshape", il);
-
-                    // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
-                    ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
-                    cb(Kcur, "Kcur", il);
-
-                    // {kv_lora_rank, 1, n_tokens}
-                    ggml_tensor * Vcur = kv_cmpr;
-                    cb(Vcur, "Vcur", il);
-
-                    // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
-                    cur = build_attn(inp_attn,
-                            model.layers[il].wo, NULL,
-                            Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
-                } else {
-                    ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
-                    cb(kv, "kv", il);
-
-                    // split into {n_embd_head_qk_nope, n_head, n_tokens}
-                    ggml_tensor * k_nope = ggml_view_3d(ctx0, kv,
-                            n_embd_head_qk_nope, n_head, n_tokens,
-                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
-                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
-                            0);
-                    cb(k_nope, "k_nope_view", il);
-
-                    // and {n_embd_head_v, n_head, n_tokens}
-                    ggml_tensor * Vcur = ggml_view_3d(ctx0, kv,
-                            n_embd_head_v, n_head, n_tokens,
-                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
-                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
-                            ggml_row_size(kv->type, n_embd_head_qk_nope));
-                    cb(Vcur, "Vcur_view", il);
-
-                    Vcur = ggml_cont(ctx0, Vcur);
-                    cb(Vcur, "Vcur_cont", il);
-
-                    // note: rope must go first for in-place context shifting in build_rope_shift()
-                    ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
-                    cb(Qcur, "Qcur", il);
-
-                    ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
-                    cb(Kcur, "Kcur", il);
-
-                    // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
-                    cur = build_attn(inp_attn,
-                            model.layers[il].wo, NULL,
-                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                }
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            if ((uint32_t) il < hparams.n_layer_dense_lead) {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                ggml_tensor * moe_out =
-                    build_moe_ffn(cur,
-                            model.layers[il].ffn_gate_inp,
-                            model.layers[il].ffn_up_exps,
-                            model.layers[il].ffn_gate_exps,
-                            model.layers[il].ffn_down_exps,
-                            model.layers[il].ffn_exp_probs_b,
-                            n_expert, n_expert_used,
-                            LLM_FFN_SILU, hparams.expert_weights_norm,
-                            true, hparams.expert_weights_scale,
-                            (llama_expert_gating_func_type) hparams.expert_gating_func,
-                            il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                // FFN shared expert
-                {
-                    ggml_tensor * ffn_shexp = build_ffn(cur,
-                            model.layers[il].ffn_up_shexp,   NULL, NULL,
-                            model.layers[il].ffn_gate_shexp, NULL, NULL,
-                            model.layers[il].ffn_down_shexp, NULL, NULL,
-                            NULL,
-                            LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                    cb(cur, "ffn_out", il);
-                }
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = ggml_mul_mat(ctx0, model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_bitnet : public llm_graph_context {
-    llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                if (model.layers[il].wq_scale) {
-                    Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
-                }
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                // B1.K
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                if (model.layers[il].wk_scale) {
-                    Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
-                }
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                // B1.V
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                if (model.layers[il].wv_scale) {
-                    Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
-                }
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        NULL, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-
-                cur = build_norm(cur,
-                        model.layers[il].attn_sub_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "attn_sub_norm", il);
-
-                cur = build_lora_mm(model.layers[il].wo, cur);
-                if (model.layers[il].wo_scale) {
-                    cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
-                }
-                if (model.layers[il].bo) {
-                    cur = ggml_add(ctx0, cur, model.layers[il].bo);
-                }
-                cb(cur, "attn_o_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward forward
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_scale,
-                    model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
-                    NULL,                      NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_sub_out", il);
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_sub_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_sub_norm", il);
-
-            cur = build_lora_mm(model.layers[il].ffn_down, cur);
-            if (model.layers[il].ffn_down_scale) {
-                cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
-            }
-            cb(cur, "ffn_down", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        // FIXME: do not use model.tok_embd directly, duplicate as model.output
-        cur = build_lora_mm(model.tok_embd, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_t5_enc : public llm_graph_context {
-    llm_build_t5_enc(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc();
-
-        auto * inp_attn = build_attn_inp_no_cache();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm_enc, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
-                ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo_enc, nullptr,
-                        Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
-                cb(cur, "kqv_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm_enc, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                // T5 uses relu, flan-T5 uses gelu-gated
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up_enc,   NULL, NULL,
-                        model.layers[il].ffn_gate_enc, NULL, NULL,
-                        model.layers[il].ffn_down_enc, NULL, NULL,
-                        NULL,
-                        model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
-                        model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
-                        il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cb(cur, "result_embd", -1);
-
-        cur = build_norm(cur,
-                model.output_norm_enc, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_t5_dec : public llm_graph_context {
-    llm_build_t5_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        //const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        ggml_tensor * embd_enc       = build_inp_cross_embd();
-        ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec();
-
-        const int64_t n_outputs_enc = embd_enc->ne[1];
-
-        auto * inp_attn_self  = build_attn_inp_kv();
-        auto * inp_attn_cross = build_attn_inp_cross();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        const int64_t dec_n_layer = hparams.dec_n_layer;
-
-        for (int il = 0; il < dec_n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
-                ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b);
-
-                cur = build_attn(inp_attn_self,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
-                cb(cur, "kqv_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, inpSA);
-            cb(cur, "cross_inp", il);
-
-            ggml_tensor * inpCA = cur;
-
-            // norm
-            cur = build_norm(cur,
-                    model.layers[il].attn_norm_cross, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm_cross", il);
-
-            // cross-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc);
-
-                cur = build_attn(inp_attn_cross,
-                        model.layers[il].wo_cross, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
-                cb(cur, "kqv_out", il);
-
-                //ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-                //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
-
-                //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                //cb(kq, "kq", il);
-
-                //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
-                //cb(kq, "kq_soft_max_ext", il);
-
-                //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
-                //cb(v, "v", il);
-
-                //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
-                //cb(kqv, "kqv", il);
-
-                //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                //cb(kqv_merged, "kqv_merged", il);
-
-                //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                //cb(cur, "kqv_merged_cont", il);
-
-                //ggml_build_forward_expand(gf, cur);
-
-                //cur = build_lora_mm(model.layers[il].wo_cross, cur);
-                //cb(cur, "kqv_out", il);
-            }
-
-            if (il == dec_n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                // T5 uses relu, flan-T5 uses gelu-gated
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU,
-                        model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ,
-                        il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cb(cur, "result_embd", -1);
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_jais : public llm_graph_context {
-    llm_build_jais(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*cur->nb[0]*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            inpL = ggml_add(ctx0, cur, ffn_inp);
-            cb(inpL, "l_out", il);
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_chatglm : public llm_graph_context {
-    llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = nullptr;
-                ggml_tensor * Kcur = nullptr;
-                ggml_tensor * Vcur = nullptr;
-
-                if (model.layers[il].wqkv == nullptr) {
-                    Qcur = build_lora_mm(model.layers[il].wq, cur);
-                    if (model.layers[il].bq) {
-                        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    }
-                    Kcur = build_lora_mm(model.layers[il].wk, cur);
-                    if (model.layers[il].bk) {
-                        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    }
-                    Vcur = build_lora_mm(model.layers[il].wv, cur);
-                    if (model.layers[il].bv) {
-                        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    }
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-                } else {
-                    cur = build_lora_mm(model.layers[il].wqkv, cur);
-                    cb(cur, "wqkv", il);
-                    if (model.layers[il].bqkv) {
-                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                        cb(cur, "bqkv", il);
-                    }
-                    Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                    Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                    Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-                }
-
-                //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // Add the input
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        NULL,                      NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-
-            }
-
-            inpL = ggml_add(ctx0, cur, ffn_inp);
-            cb(inpL, "l_out", il);
-        }
-
-        cur = build_norm(inpL,
-                model.output_norm,
-                NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_glm4 : public llm_graph_context {
-    llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // Pre-attention norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = nullptr;
-                ggml_tensor * Kcur = nullptr;
-                ggml_tensor * Vcur = nullptr;
-
-                if (model.layers[il].wqkv == nullptr) {
-                    Qcur = build_lora_mm(model.layers[il].wq, cur);
-                    if (model.layers[il].bq) {
-                        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    }
-                    Kcur = build_lora_mm(model.layers[il].wk, cur);
-                    if (model.layers[il].bk) {
-                        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    }
-                    Vcur = build_lora_mm(model.layers[il].wv, cur);
-                    if (model.layers[il].bv) {
-                        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    }
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-                } else {
-                    cur = build_lora_mm(model.layers[il].wqkv, cur);
-                    cb(cur, "wqkv", il);
-                    if (model.layers[il].bqkv) {
-                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                        cb(cur, "bqkv", il);
-                    }
-                    Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                    Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                    Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-                }
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // Post-attention norm (new!)
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm,
-                    NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "post_attn_norm", il);
-
-            // Add the input (residual connection after post-attention norm)
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                // Pre-MLP norm
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm,
-                        NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                // MLP
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        NULL,                      NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
-                cb(cur, "ffn_out", il);
-
-                // Post-MLP norm
-                cur = build_norm(cur,
-                        model.layers[il].ffn_post_norm,
-                        NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "post_mlp_norm", il);
-            }
-
-            // Add residual connection after post-MLP norm
-            inpL = ggml_add(ctx0, cur, ffn_inp);
-            cb(inpL, "l_out", il);
-        }
-
-        // Final norm
-        cur = build_norm(inpL,
-                model.output_norm,
-                NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // Output projection
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_glm4_moe : public llm_graph_context {
-    llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        // Only process up to last layer (skip final NextN layer)
-        // Final layer tensors are loaded but not processed in forward pass
-        const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
-        for (int il = 0; il < n_transformer_layers; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // Pre-attention norm
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                }
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                }
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                }
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                // Apply Q/K norm if available (GLM-4.5 355B variant)
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                    cb(Qcur, "Qcur_normed", il);
-                }
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                    cb(Kcur, "Kcur_normed", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_transformer_layers - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // Post-attention norm
-            cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "post_attn_norm", il);
-
-            // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
-            if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
-                // Dense FFN layer
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // Process routed experts using existing MoE infrastructure
-                ggml_tensor * routed_out = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        model.layers[il].ffn_exp_probs_b,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, hparams.expert_weights_norm,
-                        true, hparams.expert_weights_scale,
-                        (llama_expert_gating_func_type) hparams.expert_gating_func,
-                        il);
-                cb(routed_out, "ffn_moe_out", il);
-
-                // Process shared expert on original input
-                ggml_tensor * shared_out = build_ffn(cur,
-                        model.layers[il].ffn_up_shexp,   NULL, NULL,
-                        model.layers[il].ffn_gate_shexp, NULL, NULL,
-                        model.layers[il].ffn_down_shexp, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(shared_out, "ffn_shexp_out", il);
-
-                // Final output: routed_output + shared_output
-                cur = ggml_add(ctx0, routed_out, shared_out);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_nemotron : public llm_graph_context {
-    llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        //GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm,
-                    model.layers[il].ffn_norm_b,
-                    LLM_NORM, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    NULL,                      NULL,                        NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, model.output_norm_b,
-                LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_nemotron_h : public llm_graph_context_mamba {
-    llm_build_nemotron_h(
-            const llama_model      & model,
-            const llm_graph_params & params) :
-        llm_graph_context_mamba(params) {
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-        ggml_build_forward_expand(gf, inpL);
-
-        auto * inp = build_inp_mem_hybrid();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            if (hparams.is_recurrent(il)) {
-                // ssm layer //
-                cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
-            } else if (hparams.n_ff(il) == 0) {
-                // attention layer //
-                cur = build_attention_layer(cur, inp->get_attn(), model, n_embd_head, il);
-            } else {
-                cur = build_ffn_layer(cur, model, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // add residual
-            cur = ggml_add(ctx0, cur, inpSA);
-            cb(cur, "nemotron_h_block_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-    ggml_tensor * build_attention_layer(
-              ggml_tensor             * cur,
-              llm_graph_input_attn_kv * inp_attn,
-        const llama_model             & model,
-        const int64_t                   n_embd_head,
-        const int                       il) {
-
-        // compute Q and K and (optionally) RoPE them
-        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-        cb(Qcur, "Qcur", il);
-        if (model.layers[il].bq) {
-            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-            cb(Qcur, "Qcur", il);
-        }
-
-        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-        cb(Kcur, "Kcur", il);
-        if (model.layers[il].bk) {
-            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-            cb(Kcur, "Kcur", il);
-        }
-
-        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-        cb(Vcur, "Vcur", il);
-        if (model.layers[il].bv) {
-            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-            cb(Vcur, "Vcur", il);
-        }
-
-        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
-        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
-        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
-
-        cb(Qcur, "Qcur", il);
-        cb(Kcur, "Kcur", il);
-        cb(Vcur, "Vcur", il);
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-        cur = build_attn(inp_attn,
-                model.layers[il].wo, model.layers[il].bo,
-                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-        return cur;
-    }
-
-    ggml_tensor * build_ffn_layer(
-              ggml_tensor * cur,
-        const llama_model & model,
-        const int           il) {
-
-        cur = build_ffn(cur,
-                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                NULL,                      NULL,                        NULL,
-                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                NULL,
-                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
-        cb(cur, "ffn_out", il);
-
-        cur = build_cvec(cur, il);
-        cb(cur, "l_out", il);
-
-        return cur;
-    }
-};
-
-struct llm_build_exaone : public llm_graph_context {
-    llm_build_exaone(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-template <bool iswa>
-struct llm_build_exaone4 : public llm_graph_context {
-    llm_build_exaone4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_k;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
-        inp_attn_type * inp_attn = nullptr;
-
-        if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_iswa();
-        } else {
-            inp_attn = build_attn_inp_kv();
-        }
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // use RoPE for SWA layers or non-SWA models
-            const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;
-
-            cur = inpL;
-
-            // self-attention
-            {
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-                cb(Kcur, "Kcur_normed", il);
-
-                if (use_rope) {
-                    Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                    Kcur = ggml_rope_ext(
-                            ctx0, Kcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_ffn(ffn_inp,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = build_norm(cur,
-                    model.layers[il].ffn_post_norm, NULL,
-                    LLM_NORM_RMS, -1);
-            cb(cur, "ffn_post_norm", -1);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_rwkv6_base : public llm_graph_context {
-    const llama_model & model;
-
-    llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) {
-    }
-
-    ggml_tensor * build_rwkv6_channel_mix(
-            const llama_layer * layer,
-            ggml_tensor * cur,
-            ggml_tensor * x_prev,
-            llm_arch arch) const {
-        ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);
-        switch (arch) {
-            case LLM_ARCH_RWKV6:
-                {
-                    ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur);
-                    ggml_tensor * xr = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_r), cur);
-
-                    ggml_tensor * r = ggml_sigmoid(ctx0, build_lora_mm(layer->channel_mix_receptance, xr));
-                    ggml_tensor * k = ggml_sqr(
-                            ctx0,
-                            ggml_relu(
-                                ctx0,
-                                build_lora_mm(layer->channel_mix_key, xk)
-                                )
-                            );
-                    cur = ggml_mul(ctx0, r, build_lora_mm(layer->channel_mix_value, k));
-                } break;
-            default:
-                GGML_ABORT("fatal error");
-        }
-
-        return cur;
-    }
-
-    ggml_tensor * build_rwkv6_time_mix(
-            llm_graph_input_rs * inp,
-            ggml_tensor * cur,
-            ggml_tensor * x_prev,
-            const llama_ubatch & ubatch,
-            int   il) const {
-        const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
-
-        const auto n_tokens = ubatch.n_tokens;
-        const auto n_seqs = ubatch.n_seqs;
-        const auto n_seq_tokens = ubatch.n_seq_tokens;
-        const auto n_embd = hparams.n_embd;
-        const auto head_size = hparams.wkv_head_size;
-        const auto n_head = n_embd / head_size;
-        const auto n_head_kv = hparams.n_head_kv(il);
-
-        const auto kv_head = mctx_cur->get_head();
-
-        const auto & layer = model.layers[il];
-
-        bool is_qrwkv = layer.time_mix_first == nullptr;
-
-        ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);
-
-        sx  = ggml_reshape_2d(ctx0, sx,  n_embd, n_tokens);
-        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
-
-        ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_x), cur);
-
-        xxx = ggml_reshape_4d(
-                ctx0,
-                ggml_tanh(
-                    ctx0,
-                    ggml_mul_mat(ctx0, layer.time_mix_w1, xxx)
-                    ),
-                layer.time_mix_w1->ne[1] / 5, 1, 5, n_tokens
-                );
-
-        xxx = ggml_cont(ctx0, ggml_permute(ctx0, xxx, 0, 1, 3, 2));
-
-        xxx = ggml_mul_mat(
-                ctx0,
-                ggml_reshape_4d(
-                    ctx0,
-                    layer.time_mix_w2,
-                    layer.time_mix_w2->ne[0], layer.time_mix_w2->ne[1], 1, 5
-                    ),
-                xxx
-                );
-
-        ggml_tensor *xw, *xk, *xv, *xr, *xg;
-        if (layer.time_mix_lerp_fused) {
-            // fusing these weights makes some performance improvement
-            sx  = ggml_reshape_3d(ctx0, sx,  n_embd, 1, n_tokens);
-            cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
-            xxx = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xxx, layer.time_mix_lerp_fused), sx), cur);
-            xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);
-            xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
-            xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
-            xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
-            xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
-        } else {
-            // for backward compatibility
-            xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);
-            xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
-            xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
-            xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
-            xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
-
-            xw = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xw, layer.time_mix_lerp_w), sx), cur);
-            xk = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xk, layer.time_mix_lerp_k), sx), cur);
-            xv = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xv, layer.time_mix_lerp_v), sx), cur);
-            xr = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xr, layer.time_mix_lerp_r), sx), cur);
-            xg = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xg, layer.time_mix_lerp_g), sx), cur);
-        }
-
-        ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr);
-        ggml_tensor * k = build_lora_mm(layer.time_mix_key,        xk);
-        ggml_tensor * v = build_lora_mm(layer.time_mix_value,      xv);
-        if (layer.time_mix_receptance_b) {
-            r = ggml_add(ctx0, r, layer.time_mix_receptance_b);
-        }
-        if (layer.time_mix_key_b) {
-            k = ggml_add(ctx0, k, layer.time_mix_key_b);
-        }
-        if (layer.time_mix_value_b) {
-            v = ggml_add(ctx0, v, layer.time_mix_value_b);
-        }
-
-        ggml_tensor * g = build_lora_mm(layer.time_mix_gate, xg);
-        if (is_qrwkv) {
-            g = ggml_sigmoid(ctx0, g);
-        } else {
-            g = ggml_silu(ctx0, g);
-        }
-
-        if (n_head_kv != 0 && n_head_kv != n_head) {
-            GGML_ASSERT(n_head % n_head_kv == 0);
-            k = ggml_reshape_4d(ctx0, k, head_size, 1, n_head_kv, n_tokens);
-            v = ggml_reshape_4d(ctx0, v, head_size, 1, n_head_kv, n_tokens);
-            ggml_tensor * tmp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_size, n_head / n_head_kv, n_head_kv, n_tokens);
-            k = ggml_repeat(ctx0, k, tmp);
-            v = ggml_repeat(ctx0, v, tmp);
-        }
-
-        k = ggml_reshape_3d(ctx0, k, head_size, n_head, n_tokens);
-        v = ggml_reshape_3d(ctx0, v, head_size, n_head, n_tokens);
-        r = ggml_reshape_3d(ctx0, r, head_size, n_head, n_tokens);
-
-        ggml_tensor * w = ggml_mul_mat(
-                ctx0,
-                layer.time_mix_decay_w2,
-                ggml_tanh(
-                    ctx0,
-                    ggml_mul_mat(ctx0, layer.time_mix_decay_w1, xw)
-                    )
-                );
-
-        w = ggml_add(ctx0, w, layer.time_mix_decay);
-        w = ggml_exp(ctx0, ggml_neg(ctx0, ggml_exp(ctx0, w)));
-        w = ggml_reshape_3d(ctx0, w, head_size, n_head, n_tokens);
-
-        if (is_qrwkv) {
-            // k = k * (1 - w)
-            k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w));
-        }
-
-        ggml_tensor * wkv_state = build_rs(
-                inp, mctx_cur->get_s_l(il),
-                hparams.n_embd_s(), n_seqs);
-
-        ggml_tensor * wkv_output;
-        if (is_qrwkv) {
-            wkv_output = ggml_gated_linear_attn(ctx0, k, v, r, w, wkv_state, pow(head_size, -0.5f));
-        } else {
-            wkv_output = ggml_rwkv_wkv6(ctx0, k, v, r, layer.time_mix_first, w, wkv_state);
-        }
-        cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0);
-        wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));
-
-        ggml_build_forward_expand(
-                gf,
-                ggml_cpy(
-                    ctx0,
-                    wkv_state,
-                    ggml_view_1d(
-                        ctx0,
-                        mctx_cur->get_s_l(il),
-                        hparams.n_embd_s() * n_seqs,
-                        hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il))
-                        )
-                    )
-                );
-
-        if (!is_qrwkv) {
-            // group norm with head_count groups
-            cur = ggml_reshape_3d(ctx0, cur, n_embd / n_head, n_head, n_tokens);
-            cur = ggml_norm(ctx0, cur, 64e-5f);
-
-            // Convert back to regular vectors.
-            cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
-            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b);
-        } else {
-            cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
-        }
-
-        cur = ggml_mul(ctx0, cur, g);
-        cur = build_lora_mm(layer.time_mix_output, cur);
-
-        return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs);
-    }
-};
-
-struct llm_build_rwkv6 : public llm_build_rwkv6_base {
-    llm_build_rwkv6(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv6_base(model, params) {
-        GGML_ASSERT(hparams.token_shift_count == 2);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-        inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
-
-        auto * rs_inp = build_rs_inp();
-
-        const auto n_embd = hparams.n_embd;
-        const auto n_seq_tokens = ubatch.n_seq_tokens;
-        const auto n_seqs = ubatch.n_seqs;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const llama_layer * layer = &model.layers[il];
-            inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
-
-            ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
-
-            ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
-            ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift));
-
-            ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il);
-            cb(att_norm, "attn_norm", il);
-
-            ggml_tensor * x_prev = ggml_concat(
-                    ctx0,
-                    att_shift,
-                    ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
-                    1
-                    );
-
-            cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il);
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il);
-            cb(ffn_norm, "ffn_norm", il);
-
-            x_prev = ggml_concat(
-                    ctx0,
-                    ffn_shift,
-                    ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0),
-                    1
-                    );
-
-            token_shift = ggml_concat(ctx0,
-                    ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)),
-                    ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(ffn_norm)),
-                    1
-                    );
-            ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
-
-            ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp,  n_embd, n_tokens);
-            ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
-            x_prev   = ggml_reshape_2d(ctx0, x_prev,   n_embd, n_tokens);
-            cur      = ggml_reshape_2d(ctx0, cur,      n_embd, n_tokens);
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                ffn_inp  = ggml_get_rows(ctx0, ffn_inp,  inp_out_ids);
-                ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids);
-                x_prev   = ggml_get_rows(ctx0, x_prev,   inp_out_ids);
-                cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
-            }
-
-            cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6);
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) {
-                cur = ggml_scale(ctx0, cur, 0.5F);
-            }
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-// ref: https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1/blob/main/modeling_rwkv6qwen2.py
-struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base {
-    llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv6_base(model, params) {
-        GGML_ASSERT(n_embd == hparams.n_embd_r());
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * rs_inp = build_rs_inp();
-
-        const auto n_embd = hparams.n_embd;
-        const auto n_seq_tokens = ubatch.n_seq_tokens;
-        const auto n_seqs = ubatch.n_seqs;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const llama_layer * layer = &model.layers[il];
-            inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
-
-            ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
-
-            ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
-            cb(att_norm, "attn_norm", il);
-
-            ggml_tensor * x_prev = ggml_concat(
-                    ctx0,
-                    token_shift,
-                    ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
-                    1
-                    );
-
-            cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il);
-
-            token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
-            ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
-            ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
-                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
-            }
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_rwkv7_base : public llm_graph_context {
-    const llama_model & model;
-
-    llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) {
-    }
-
-    ggml_tensor * build_rwkv7_channel_mix(
-            const llama_layer * layer,
-            ggml_tensor * cur,
-            ggml_tensor * x_prev,
-            llm_arch arch) const {
-        ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);
-        switch (arch) {
-            case LLM_ARCH_RWKV7:
-                {
-                    ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur);
-
-                    ggml_tensor * k = ggml_sqr(
-                        ctx0,
-                        ggml_relu(
-                            ctx0,
-                            build_lora_mm(layer->channel_mix_key, xk)
-                        )
-                    );
-
-                    cur = build_lora_mm(layer->channel_mix_value, k);
-                } break;
-            default:
-                GGML_ABORT("fatal error");
-        }
-
-        return cur;
-    }
-
-    ggml_tensor * build_rwkv7_time_mix(
-            llm_graph_input_rs * inp,
-            ggml_tensor * cur,
-            ggml_tensor * x_prev,
-            ggml_tensor *& first_layer_value,
-            const llama_ubatch & ubatch,
-            int   il) const {
-        const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
-
-        const auto n_tokens = ubatch.n_tokens;
-        const auto n_seqs = ubatch.n_seqs;
-        const auto n_embd = hparams.n_embd;
-        const auto head_size = hparams.wkv_head_size;
-        const auto head_count = n_embd / head_size;
-        const auto n_seq_tokens = ubatch.n_seq_tokens;
-
-        const auto kv_head = mctx_cur->get_head();
-
-        const auto & layer = model.layers[il];
-
-        bool has_gating = layer.time_mix_g1 && layer.time_mix_g2;
-
-        ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);
-        ggml_tensor * dummy = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_embd, n_seq_tokens, n_seqs, has_gating ? 6 : 5);
-        sx = ggml_repeat(ctx0, sx, dummy);
-
-        ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_fused), cur);
-
-        ggml_tensor * xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);
-        ggml_tensor * xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
-        ggml_tensor * xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
-        ggml_tensor * xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
-        ggml_tensor * xa = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
-        ggml_tensor * xg = has_gating ? ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 5 * sizeof(float)) : nullptr;
-
-        ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr);
-        ggml_tensor * w = ggml_add(
-            ctx0,
-            ggml_mul_mat(ctx0, layer.time_mix_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xw))),
-            layer.time_mix_w0
-        );
-        w = ggml_exp(ctx0, ggml_scale(ctx0, ggml_sigmoid(ctx0, w), -0.606531));
-
-        ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk);
-        ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv);
-        if (first_layer_value == nullptr) {
-            first_layer_value = v;
-        } else {
-            // Add the first layer value as a residual connection.
-            v = ggml_add(ctx0, v,
-                ggml_mul(ctx0,
-                    ggml_sub(ctx0, first_layer_value, v),
-                    ggml_sigmoid(ctx0, ggml_add(ctx0,
-                            ggml_mul_mat(ctx0, layer.time_mix_v2, ggml_mul_mat(ctx0, layer.time_mix_v1, xv)),
-                            layer.time_mix_v0
-                        )
-                    )
-                )
-            );
-        }
-
-        ggml_tensor * g = nullptr;
-        if (layer.time_mix_g1 && layer.time_mix_g2) {
-            g = ggml_mul_mat(ctx0, layer.time_mix_g2, ggml_sigmoid(ctx0, ggml_mul_mat(ctx0, layer.time_mix_g1, xg)));
-        }
-
-        ggml_tensor * a = ggml_sigmoid(ctx0,
-            ggml_add(
-                ctx0,
-                ggml_mul_mat(ctx0, layer.time_mix_a2, ggml_mul_mat(ctx0, layer.time_mix_a1, xa)),
-                layer.time_mix_a0
-            )
-        );
-
-        ggml_tensor * kk = ggml_reshape_3d(ctx0, ggml_mul(ctx0, k, layer.time_mix_k_k), head_size, head_count, n_tokens);
-        kk = ggml_l2_norm(ctx0, kk, 1e-12);
-
-        ggml_tensor * ka = ggml_mul(ctx0, k, layer.time_mix_k_a);
-        k = ggml_add(ctx0, k, ggml_sub(ctx0, ggml_mul(ctx0, a, ka), ka));
-
-        r = ggml_reshape_3d(ctx0, r, head_size, head_count, n_tokens);
-        w = ggml_reshape_3d(ctx0, w, head_size, head_count, n_tokens);
-        k = ggml_reshape_3d(ctx0, k, head_size, head_count, n_tokens);
-        v = ggml_reshape_3d(ctx0, v, head_size, head_count, n_tokens);
-        a = ggml_reshape_3d(ctx0, a, head_size, head_count, n_tokens);
-
-        ggml_tensor * wkv_state = build_rs(
-                inp, mctx_cur->get_s_l(il),
-                hparams.n_embd_s(), n_seqs);
-
-        ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx0, r, w, k, v, ggml_neg(ctx0, kk), ggml_mul(ctx0, kk, a), wkv_state);
-        cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0);
-        wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));
-
-        ggml_build_forward_expand(
-                gf,
-                ggml_cpy(
-                    ctx0,
-                    wkv_state,
-                    ggml_view_1d(
-                        ctx0,
-                        mctx_cur->get_s_l(il),
-                        hparams.n_embd_s() * n_seqs,
-                        hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il))
-                        )
-                    )
-                );
-
-        if (layer.time_mix_ln && layer.time_mix_ln_b) {
-            // group norm with head_count groups
-            cur = ggml_reshape_3d(ctx0, cur, n_embd / head_count, head_count, n_tokens);
-            cur = ggml_norm(ctx0, cur, 64e-5f);
-
-            // Convert back to regular vectors.
-            cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
-            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b);
-        } else {
-            cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
-        }
-
-        ggml_tensor * rk = ggml_sum_rows(ctx0,
-                ggml_mul(ctx0, ggml_mul(ctx0, k, r), ggml_reshape_2d(ctx0, layer.time_mix_r_k, head_size, head_count)));
-        cur = ggml_add(ctx0, cur, ggml_reshape_2d(ctx0, ggml_mul(ctx0, v, rk), n_embd, n_tokens));
-
-        if (has_gating) {
-            cur = ggml_mul(ctx0, cur, g);
-        }
-        cur = build_lora_mm(layer.time_mix_output, cur);
-
-        return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs);
-    }
-};
-
-struct llm_build_rwkv7 : public llm_build_rwkv7_base {
-    llm_build_rwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
-        GGML_ASSERT(hparams.token_shift_count == 2);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-        ggml_tensor * v_first = nullptr;
-
-        inpL = build_inp_embd(model.tok_embd);
-        inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
-
-        auto * rs_inp = build_rs_inp();
-
-        const auto n_embd = hparams.n_embd;
-        const auto n_seq_tokens = ubatch.n_seq_tokens;
-        const auto n_seqs = ubatch.n_seqs;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const llama_layer * layer = &model.layers[il];
-            inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
-
-            ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
-
-            ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
-            ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift));
-
-            ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il);
-            cb(att_norm, "attn_norm", il);
-
-            ggml_tensor * x_prev = ggml_concat(
-                    ctx0,
-                    att_shift,
-                    ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
-                    1
-                    );
-
-            cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il);
-            cb(ffn_norm, "ffn_norm", il);
-
-            x_prev = ggml_concat(
-                    ctx0,
-                    ffn_shift,
-                    ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0),
-                    1
-                    );
-
-            token_shift = ggml_concat(ctx0,
-                    ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)),
-                    ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(ffn_norm)),
-                    1
-                    );
-            ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
-
-            ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp,  n_embd, n_tokens);
-            ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
-            x_prev   = ggml_reshape_2d(ctx0, x_prev,   n_embd, n_tokens);
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                ffn_inp  = ggml_get_rows(ctx0, ffn_inp,  inp_out_ids);
-                ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids);
-                x_prev   = ggml_get_rows(ctx0, x_prev,   inp_out_ids);
-            }
-
-            cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7);
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-
-struct llm_build_arwkv7 : public llm_build_rwkv7_base {
-    llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
-        GGML_ASSERT(n_embd == hparams.n_embd_r());
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-        ggml_tensor * v_first = nullptr;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * rs_inp = build_rs_inp();
-
-        const auto n_embd = hparams.n_embd;
-        const auto n_seq_tokens = ubatch.n_seq_tokens;
-        const auto n_seqs = ubatch.n_seqs;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const llama_layer * layer = &model.layers[il];
-            inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
-
-            ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
-
-            ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
-            cb(att_norm, "attn_norm", il);
-
-            ggml_tensor * x_prev = ggml_concat(
-                    ctx0,
-                    token_shift,
-                    ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
-                    1
-                    );
-
-            cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
-
-            token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
-            ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
-            ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
-                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
-            }
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_granite : public llm_graph_context {
-    llm_build_granite(
-        const llama_model & model,
-        const llm_graph_params & params)
-        : llm_graph_context(params) {
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - built only if rope enabled
-        ggml_tensor * inp_pos = nullptr;
-        if (hparams.rope_finetuned) {
-            inp_pos = build_inp_pos();
-        }
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            cur = build_attention_layer(
-                cur, inp_pos, inp_attn,
-                model, n_embd_head, il);
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // ffn
-            cur = build_layer_ffn(cur, inpSA, model, il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        // For Granite architectures - scale logits
-        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-    ggml_tensor * build_attention_layer(
-              ggml_tensor             * cur,
-              ggml_tensor             * inp_pos,
-              llm_graph_input_attn_kv * inp_attn,
-        const llama_model             & model,
-        const int64_t                 n_embd_head,
-        const int                     il) {
-
-        // compute Q and K and (optionally) RoPE them
-        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-        cb(Qcur, "Qcur", il);
-        if (model.layers[il].bq) {
-            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-            cb(Qcur, "Qcur", il);
-        }
-
-        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-        cb(Kcur, "Kcur", il);
-        if (model.layers[il].bk) {
-            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-            cb(Kcur, "Kcur", il);
-        }
-
-        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-        cb(Vcur, "Vcur", il);
-        if (model.layers[il].bv) {
-            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-            cb(Vcur, "Vcur", il);
-        }
-
-        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
-        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
-        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
-
-        const bool use_rope = hparams.rope_finetuned;
-        if (use_rope) {
-            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-            Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-
-            Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-        }
-
-        cb(Qcur, "Qcur", il);
-        cb(Kcur, "Kcur", il);
-        cb(Vcur, "Vcur", il);
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-        cur = build_attn(inp_attn,
-                model.layers[il].wo, model.layers[il].bo,
-                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-        return cur;
-    }
-
-    ggml_tensor * build_layer_ffn(
-              ggml_tensor       * cur,
-              ggml_tensor       * inpSA,
-        const llama_model       & model,
-        const int                 il) {
-
-        // For Granite architectures - scale residual
-        if (hparams.f_residual_scale) {
-            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-        }
-        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-        cb(ffn_inp, "ffn_inp", il);
-
-        // feed-forward network (non-MoE)
-        if (model.layers[il].ffn_gate_inp == nullptr) {
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-                    cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(cur, "ffn_out", il);
-
-        } else {
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-                    cb(cur, "ffn_norm", il);
-
-            ggml_tensor * moe_out = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(moe_out, "ffn_moe_out", il);
-
-            // For Granite MoE Shared
-            if (hparams.n_ff_shexp > 0) {
-                ggml_tensor * ffn_shexp = build_ffn(cur,
-                    model.layers[il].ffn_up_shexp,   NULL, NULL,
-                    model.layers[il].ffn_gate_shexp, NULL, NULL,
-                    model.layers[il].ffn_down_shexp, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(ffn_shexp, "ffn_shexp", il);
-
-                cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                cb(cur, "ffn_out", il);
-            } else {
-                cur = moe_out;
-            }
-        }
-
-        // For Granite architectures - scale residual
-        if (hparams.f_residual_scale) {
-            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-        }
-        cur = ggml_add(ctx0, cur, ffn_inp);
-        cb(cur, "ffn_out", il);
-
-        cur = build_cvec(cur, il);
-        cb(cur, "l_out", il);
-
-        return cur;
-    }
-};
-
-struct llm_build_granite_hybrid : public llm_graph_context_mamba {
-    llm_build_granite_hybrid(
-                 const llama_model & model,
-            const llm_graph_params & params) :
-        llm_graph_context_mamba(params) {
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        auto * inp = build_inp_mem_hybrid();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        // Positional embeddings populated if rope enabled
-        ggml_tensor * inp_pos = nullptr;
-        if (hparams.rope_finetuned) {
-            inp_pos = build_inp_pos();
-        }
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            if (hparams.is_recurrent(il)) {
-                // ssm layer //
-                cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
-            } else {
-                // attention layer //
-                cur = build_attention_layer(
-                    cur, inp_pos, inp->get_attn(), model,
-                    n_embd_head, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // ffn
-            cur = build_layer_ffn(cur, inpSA, model, il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        // For Granite architectures - scale logits
-        if (hparams.f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        }
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-    ggml_tensor * build_attention_layer(
-              ggml_tensor             * cur,
-              ggml_tensor             * inp_pos,
-              llm_graph_input_attn_kv * inp_attn,
-        const llama_model             & model,
-        const int64_t                 n_embd_head,
-        const int                     il) {
-
-        // compute Q and K and (optionally) RoPE them
-        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-        cb(Qcur, "Qcur", il);
-        if (model.layers[il].bq) {
-            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-            cb(Qcur, "Qcur", il);
-        }
-
-        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-        cb(Kcur, "Kcur", il);
-        if (model.layers[il].bk) {
-            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-            cb(Kcur, "Kcur", il);
-        }
-
-        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-        cb(Vcur, "Vcur", il);
-        if (model.layers[il].bv) {
-            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-            cb(Vcur, "Vcur", il);
-        }
-
-        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
-        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
-        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
-
-        const bool use_rope = hparams.rope_finetuned;
-        if (use_rope) {
-            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-            Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-
-            Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-        }
-
-        cb(Qcur, "Qcur", il);
-        cb(Kcur, "Kcur", il);
-        cb(Vcur, "Vcur", il);
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-        cur = build_attn(inp_attn,
-                model.layers[il].wo, model.layers[il].bo,
-                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-        return cur;
-    }
-
-    ggml_tensor * build_layer_ffn(
-              ggml_tensor * cur,
-              ggml_tensor * inpSA,
-        const llama_model & model,
-        const int           il) {
-
-        // For Granite architectures - scale residual
-        if (hparams.f_residual_scale) {
-            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-        }
-        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-        cb(ffn_inp, "ffn_inp", il);
-
-        // feed-forward network (non-MoE)
-        if (model.layers[il].ffn_gate_inp == nullptr) {
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-                    cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(cur, "ffn_out", il);
-
-        } else {
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-                    cb(cur, "ffn_norm", il);
-
-            ggml_tensor * moe_out = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(moe_out, "ffn_moe_out", il);
-
-            // For Granite MoE Shared
-            if (hparams.n_ff_shexp > 0) {
-                ggml_tensor * ffn_shexp = build_ffn(cur,
-                    model.layers[il].ffn_up_shexp,   NULL, NULL,
-                    model.layers[il].ffn_gate_shexp, NULL, NULL,
-                    model.layers[il].ffn_down_shexp, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(ffn_shexp, "ffn_shexp", il);
-
-                cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                cb(cur, "ffn_out", il);
-            } else {
-                cur = moe_out;
-            }
-        }
-
-        // For Granite architectures - scale residual
-        if (hparams.f_residual_scale) {
-            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-        }
-        cur = ggml_add(ctx0, cur, ffn_inp);
-        cb(cur, "ffn_out", il);
-
-        cur = build_cvec(cur, il);
-        cb(cur, "l_out", il);
-
-        return cur;
-    }
-};
-
-struct llm_build_solar : public llm_graph_context {
-    llm_build_solar(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        struct ggml_tensor * bskcn_1;
-        struct ggml_tensor * bskcn_2;
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            if (hparams.n_bskcn(0, il)) {
-                bskcn_1 = inpSA;
-            }
-
-            if (hparams.n_bskcn(1, il)) {
-                bskcn_2 = inpSA;
-            }
-
-            if (hparams.n_bskcn(2, il)) {
-                inpSA = ggml_add(
-                   ctx0,
-                   ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
-                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
-            }
-
-            if (hparams.n_bskcn(3, il)) {
-                inpSA = ggml_add(
-                   ctx0,
-                   ggml_mul(ctx0, bskcn_2, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
-                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
-            }
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-// ref: https://github.com/facebookresearch/chameleon
-// based on the original build_llama() function, changes:
-//   * qk-norm
-//   * swin-norm
-//   * removed bias
-//   * removed MoE
-struct llm_build_chameleon : public llm_graph_context {
-    llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            if (hparams.swin_norm) {
-                cur = inpL;
-            } else {
-                cur = build_norm(inpL,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "attn_norm", il);
-            }
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
-                            ggml_element_size(Qcur) * n_embd_head,
-                            ggml_element_size(Qcur) * n_embd_head * n_head,
-                            0);
-                    cb(Qcur, "Qcur", il);
-
-                    Qcur = build_norm(Qcur,
-                            model.layers[il].attn_q_norm,
-                            model.layers[il].attn_q_norm_b,
-                            LLM_NORM, il);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
-                            ggml_element_size(Kcur) * n_embd_head,
-                            ggml_element_size(Kcur) * n_embd_head * n_head_kv,
-                            0);
-                    cb(Kcur, "Kcur", il);
-
-                    Kcur = build_norm(Kcur,
-                            model.layers[il].attn_k_norm,
-                            model.layers[il].attn_k_norm_b,
-                            LLM_NORM, il);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            if (hparams.swin_norm) {
-                cur = build_norm(cur,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, il);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            if (!hparams.swin_norm) {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-            }
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            if (hparams.swin_norm) {
-                cur = build_norm(cur,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output_with_img_logits", -1);
-
-        // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
-        // Needs to be removed once image outputs are supported.
-        int img_token_end_idx = 8196;
-        int img_token_start_idx = 4;
-        int num_img_tokens = img_token_end_idx - img_token_start_idx;
-        // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
-        // which ensures that text token values are always at least larger than image token values
-        ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
-        img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
-        cb(img_logits, "img_logits", -1);
-
-        cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_wavtokenizer_dec : public llm_graph_context {
-    llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL));
-
-        cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1);
-        cur = ggml_add(ctx0, cur, model.conv1d_b);
-
-        // posnet
-        for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) {
-            const auto & layer = model.layers[il].posnet;
-
-            inpL = cur;
-
-            switch (il) {
-                case 0:
-                case 1:
-                case 3:
-                case 4:
-                    {
-                        cur = build_norm(cur,
-                                layer.norm1,
-                                layer.norm1_b,
-                                LLM_NORM_GROUP, 0);
-
-                        cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur);
-
-                        cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1);
-                        cur = ggml_add(ctx0, cur, layer.conv1_b);
-
-                        cur = build_norm(cur,
-                                layer.norm2,
-                                layer.norm2_b,
-                                LLM_NORM_GROUP, 0);
-
-                        cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur);
-
-                        cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1);
-                        cur = ggml_add(ctx0, cur, layer.conv2_b);
-
-                        cur = ggml_add(ctx0, cur, inpL);
-                    } break;
-                case 2:
-                    {
-                        cur = build_norm(cur,
-                                layer.attn_norm,
-                                layer.attn_norm_b,
-                                LLM_NORM_GROUP, 0);
-
-                        ggml_tensor * q;
-                        ggml_tensor * k;
-                        ggml_tensor * v;
-
-                        q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1);
-                        k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1);
-                        v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1);
-
-                        q = ggml_add(ctx0, q, layer.attn_q_b);
-                        k = ggml_add(ctx0, k, layer.attn_k_b);
-                        v = ggml_add(ctx0, v, layer.attn_v_b);
-
-                        q = ggml_cont(ctx0, ggml_transpose(ctx0, q));
-                        k = ggml_cont(ctx0, ggml_transpose(ctx0, k));
-
-                        ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-
-                        kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f);
-
-                        cur = ggml_mul_mat(ctx0, kq, v);
-
-                        cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1);
-                        cur = ggml_add(ctx0, cur, layer.attn_o_b);
-
-                        cur = ggml_add(ctx0, cur, inpL);
-                    } break;
-                case 5:
-                    {
-                        cur = build_norm(cur,
-                                layer.norm,
-                                layer.norm_b,
-                                LLM_NORM_GROUP, 0);
-                    } break;
-                default: GGML_ABORT("unknown posnet layer");
-            };
-        }
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-        cur = build_norm(cur,
-                model.tok_norm,
-                model.tok_norm_b,
-                LLM_NORM, -1);
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-        inpL = cur;
-
-        // convnext
-        for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) {
-            const auto & layer = model.layers[il].convnext;
-
-            cur = inpL;
-
-            cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1);
-            cur = ggml_add(ctx0, cur, layer.dw_b);
-
-            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-            cur = build_norm(cur,
-                    layer.norm,
-                    layer.norm_b,
-                    LLM_NORM, -1);
-
-            cur = build_ffn(cur,
-                    layer.pw1, layer.pw1_b, NULL,
-                    NULL,      NULL,        NULL,
-                    layer.pw2, layer.pw2_b, NULL,
-                    NULL,
-                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
-
-            cur = ggml_mul(ctx0, cur, layer.gamma);
-
-            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-            inpL = ggml_add(ctx0, cur, inpL);
-        }
-
-        cur = inpL;
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-        cur = build_norm(cur,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cur = ggml_add(ctx0, cur, model.output_b);
-
-        cb(cur, "result_embd", -1);
-        res->t_embd = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_plm : public llm_graph_context {
-    llm_build_plm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k));
-
-        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
-        const uint32_t kv_lora_rank = hparams.n_lora_kv;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        // {n_embd, n_tokens}
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                ggml_tensor * q = NULL;
-                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                cb(q, "q", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        0);
-                cb(q_nope, "q_nope", il);
-
-                // and {n_head * n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        ggml_row_size(q->type, n_embd_head_qk_nope));
-                cb(q_pe, "q_pe", il);
-
-                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
-
-                // split into {kv_lora_rank, n_tokens}
-                ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        0);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // and {n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        kv_pe_compresseed->nb[1],
-                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
-                cb(k_pe, "k_pe", il);
-
-                kv_compressed = build_norm(kv_compressed,
-                        model.layers[il].attn_kv_a_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
-                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
-                cb(kv, "kv", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        0);
-                cb(k_nope, "k_nope", il);
-
-                // and {n_head * n_embd_head_v, n_tokens}
-                ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_cont(ctx0, v_states);
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
-                        ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
-                        0);
-                cb(v_states, "v_states", il);
-
-                q_pe = ggml_rope_ext(
-                        ctx0, q_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-                cb(q_pe, "q_pe", il);
-
-                // shared RoPE key
-                k_pe = ggml_rope_ext(
-                        ctx0, k_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-                cb(k_pe, "k_pe", il);
-
-                ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
-                cb(q_states, "q_states", il);
-
-                ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
-                cb(k_states, "k_states", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    NULL, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_bailingmoe : public llm_graph_context {
-    llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            ggml_tensor * moe_out =
-                build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, hparams.expert_weights_norm,
-                        false, hparams.expert_weights_scale,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-            cb(moe_out, "ffn_moe_out", il);
-
-            // FFN shared expert
-            {
-                ggml_tensor * ffn_shexp = build_ffn(cur,
-                        model.layers[il].ffn_up_shexp,   NULL, NULL,
-                        model.layers[il].ffn_gate_shexp, NULL, NULL,
-                        model.layers[il].ffn_down_shexp, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(ffn_shexp, "ffn_shexp", il);
-
-                cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_bailingmoe2 : public llm_graph_context {
-    llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
-        for (int il = 0; il < n_transformer_layers; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
-                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_transformer_layers - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
-            cb(sa_out, "sa_out", il);
-
-            // MoE branch
-            cur = build_norm(sa_out,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                ggml_tensor * moe_out =
-                    build_moe_ffn(cur,
-                            model.layers[il].ffn_gate_inp,
-                            model.layers[il].ffn_up_exps,
-                            model.layers[il].ffn_gate_exps,
-                            model.layers[il].ffn_down_exps,
-                            model.layers[il].ffn_exp_probs_b,
-                            n_expert, n_expert_used,
-                            LLM_FFN_SILU, hparams.expert_weights_norm,
-                            true, hparams.expert_weights_scale,
-                            (llama_expert_gating_func_type) hparams.expert_gating_func,
-                            il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                {
-                    ggml_tensor * ffn_shexp = build_ffn(cur,
-                            model.layers[il].ffn_up_shexp,   NULL, NULL,
-                            model.layers[il].ffn_gate_shexp, NULL, NULL,
-                            model.layers[il].ffn_down_shexp, NULL, NULL,
-                            NULL,
-                            LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                    cb(cur, "ffn_out", il);
-                }
-            }
-
-            cur = ggml_add(ctx0, cur, sa_out);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_dots1 : public llm_graph_context {
-    llm_build_dots1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            if ((uint32_t) il < hparams.n_layer_dense_lead) {
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                ggml_tensor * moe_out =
-                    build_moe_ffn(cur,
-                            model.layers[il].ffn_gate_inp,
-                            model.layers[il].ffn_up_exps,
-                            model.layers[il].ffn_gate_exps,
-                            model.layers[il].ffn_down_exps,
-                            model.layers[il].ffn_exp_probs_b,
-                            n_expert, n_expert_used,
-                            LLM_FFN_SILU, hparams.expert_weights_norm,
-                            true, hparams.expert_weights_scale,
-                            (llama_expert_gating_func_type) hparams.expert_gating_func,
-                            il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                {
-                    ggml_tensor * ffn_shexp = build_ffn(cur,
-                            model.layers[il].ffn_up_shexp,   NULL, NULL,
-                            model.layers[il].ffn_gate_shexp, NULL, NULL,
-                            model.layers[il].ffn_down_shexp, NULL, NULL,
-                            NULL,
-                            LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                    cb(cur, "ffn_out", il);
-                }
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_ernie4_5 : public llm_graph_context {
-    llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            {
-                cur = build_norm(inpL,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "attn_norm", il);
-            }
-
-            // self-attention
-            {
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_ernie4_5_moe : public llm_graph_context {
-    llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-            // norm
-            {
-                cur = build_norm(inpL,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "attn_norm", il);
-            }
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            bool is_moe_layer = static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;
-
-            if (!is_moe_layer) {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                ggml_tensor * moe_out = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        model.layers[il].ffn_exp_probs_b,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                // Shared expert (if present)
-                if (hparams.n_ff_shexp > 0) {
-                    ggml_tensor * ffn_shexp = build_ffn(cur,
-                        model.layers[il].ffn_up_shexp,   NULL, NULL,
-                        model.layers[il].ffn_gate_shexp, NULL, NULL,
-                        model.layers[il].ffn_down_shexp, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                } else {
-                    cur = moe_out;
-                }
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_falcon_h1 : public llm_graph_context_mamba {
-    llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        // Build the inputs in the recurrent & kv cache
-        auto * inp = build_inp_mem_hybrid();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-            cb(Qcur, "Qcur", il);
-
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-            cb(Kcur, "Kcur", il);
-
-            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-            cb(Vcur, "Vcur", il);
-
-            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-            Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow);
-
-            Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-
-            cb(Qcur, "Qcur-post-rope", il);
-            cb(Kcur, "Kcur-post-rope", il);
-            cb(Vcur, "Vcur-post-rope", il);
-
-            ggml_tensor * attn_out = build_attn(inp->get_attn(),
-                    model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-            cb(attn_out, "attn_out", il);
-
-            cur = build_norm(inpL,
-                model.layers[il].attn_norm, NULL,
-                LLM_NORM_RMS, il);
-            // Mamba2 layer
-            cb(cur, "ssm_in", il);
-
-            ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
-            cb(ssm_out, "ssm_out", il);
-
-            // // Aggregation
-            cur = ggml_add(ctx0, attn_out, ssm_out);
-            inpSA = ggml_add(ctx0, cur, inpSA);
-            cb(cur, "layer_out", il);
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = inpSA;
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b, NULL,
-                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, inpSA);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_plamo2 : public llm_graph_context_mamba {
-    llm_build_plamo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        // {n_embd, n_tokens}
-        inpL = build_inp_embd(model.tok_embd);
-        cb(inpL, "embedding_output", -1);
-
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_hybrid = build_inp_mem_hybrid();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * residual = inpL;
-
-            // ggml_graph_add_node(gf, model.layers[il].attn_norm);
-            // cb(model.layers[il].attn_norm, "attn_norm", il);
-
-            // pre_mixer_norm
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-
-            // check if this layer is Mamba or Attention
-            bool is_mamba_layer = hparams.is_recurrent(il);
-
-            if (is_mamba_layer) {
-                // PLaMo-2 Mamba layer
-                cur = build_plamo2_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);
-            } else {
-                // PLaMo-2 Attention layer
-                cur = build_plamo2_attn_layer(inp_hybrid->get_attn(), inp_pos, cur, model, il);
-            }
-
-            // post_mixer_norm
-            cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            // residual connection
-            cur = ggml_add(ctx0, cur, residual);
-            cb(cur, "attn_residual", il);
-            residual = cur;
-
-            // pre-ffn norm
-            cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "ffn_pre_norm", il);
-
-            // feed-forward network
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    NULL,                      NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
-            cb(cur, "ffn_out", il);
-
-            // post ffn norm
-            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "ffn_post_norm", il);
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
-            }
-
-            // residual connection
-            cur = ggml_add(ctx0, cur, residual);
-            cb(cur, "ffn_residual", il);
-
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        // final norm
-        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
-        cb(cur, "result_norm", -1);
-
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output", -1);
-
-        // Explicitly mark as output tensor to ensure proper backend assignment
-        ggml_set_output(cur);
-
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-private:
-    ggml_tensor * build_plamo2_attn_layer(
-            llm_graph_input_attn_kv * inp,
-            ggml_tensor * inp_pos,
-            ggml_tensor * cur,
-            const llama_model & model,
-            int il) {
-
-        // self-attention
-        {
-            // PLaMo-2 uses combined QKV tensor
-            ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
-            cb(qkv, "wqkv", il);
-
-            // split QKV tensor into Q, K, V
-            const int64_t n_embd_head_q = hparams.n_embd_head_k;
-            const int64_t n_embd_head_k = hparams.n_embd_head_k;
-            const int64_t n_embd_head_v = hparams.n_embd_head_v;
-            int32_t n_head = hparams.n_head(il);
-            int32_t n_head_kv = hparams.n_head_kv(il);
-
-            const int64_t q_offset = 0;
-            const int64_t k_offset = n_embd_head_q * n_head;
-            const int64_t v_offset = k_offset + n_embd_head_k * n_head_kv;
-
-            ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head,    n_tokens, n_embd_head_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));
-            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));
-            ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv, n_tokens, n_embd_head_v * sizeof(float), qkv->nb[1], v_offset * ggml_element_size(qkv));
-
-            cb(Qcur, "Qcur", il);
-            cb(Kcur, "Kcur", il);
-            cb(Vcur, "Vcur", il);
-
-            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-            cb(Qcur, "Qcur_normed", il);
-
-            Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-
-            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-            cb(Kcur, "Kcur_normed", il);
-
-            Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-
-            cur = build_attn(inp,
-                    model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il);
-        }
-
-        cb(cur, "attn_out", il);
-
-        return cur;
-    }
-
-    ggml_tensor * build_plamo2_mamba_layer(
-         llm_graph_input_rs * inp,
-               ggml_tensor * cur,
-         const llama_model & model,
-        const llama_ubatch & ubatch,
-                       int   il) {
-
-        const auto * mctx_cur = inp->mctx;
-
-        const auto kv_head = mctx_cur->get_head();
-
-        const int64_t d_conv   = hparams.ssm_d_conv;
-        const int64_t d_inner  = hparams.ssm_d_inner;
-        const int64_t d_state  = hparams.ssm_d_state;
-        const int64_t n_heads  = hparams.ssm_dt_rank;
-        const int64_t head_dim = d_inner / n_heads;
-        const int64_t n_group  = hparams.ssm_n_group;
-        const int64_t n_seqs   = ubatch.n_seqs;
-
-        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
-
-        GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(ubatch.equal_seqs());
-        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
-
-        ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
-        ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
-
-        ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
-        conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs);
-
-        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
-        // in_proj: {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
-        ggml_tensor * zx = build_lora_mm(model.layers[il].ssm_in, cur);
-        cb(zx, "mamba_in_proj", il);
-        // {8192, 5, 1, 1} -> {8192, 1, 5, 1}
-        zx = ggml_permute(ctx0, zx, 0, 2, 1, 3);
-        zx = ggml_cont_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs);
-        cb(zx, "mamba_in_proj_out", il);
-
-        // split into z and x
-        // => {head_dim * n_heads, n_seq_tokens, n_seqs}
-        ggml_tensor * x = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], head_dim*ggml_element_size(zx));
-        x = ggml_cont_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs);
-        // x = ggml_permute(ctx0, x, 0, 2, 1, 3);
-        cb(x, "mamba_x_split", il);
-
-        ggml_tensor * z = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], 0);
-        cb(z, "mamba_z_split", il);
-
-        // conv1d
-        {
-            // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
-            ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
-            cb(conv_x, "mamba_conv1d_input", il);
-
-            // copy last (d_conv - 1) columns back into the state cache
-            ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs,
-                    conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0]));
-
-            ggml_build_forward_expand(gf,
-                ggml_cpy(ctx0, last_conv,
-                    ggml_view_1d(ctx0, conv_states_all,
-                        (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs),
-                        kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all))));
-            cb(conv_states_all, "mamba_conv1d_state", il);
-
-            // 1D convolution
-            x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
-            cb(x, "mamba_conv1d", il);
-
-            x = ggml_silu(ctx0, x);
-            cb(x, "mamba_conv1d_silu", il);
-        }
-
-        // SSM
-        {
-            // bcdt_proj: {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}
-            ggml_tensor * x_bcdt = build_lora_mm(model.layers[il].ssm_x, x);
-            cb(x_bcdt, "mamba_bcdt_proj", il);
-
-            // split into dt, B, C
-            const int64_t dt_dim = std::max(64, int(hparams.n_embd / 16));
-            ggml_tensor * B = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], 0);
-            ggml_tensor * C  = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], ggml_element_size(x_bcdt)*d_state);
-            ggml_tensor * dt  = ggml_view_3d(ctx0, x_bcdt, dt_dim, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], ggml_element_size(x_bcdt)*(2*d_state));
-            cb(B, "mamba_B_raw", il);
-            cb(C, "mamba_C_raw", il);
-            cb(dt, "mamba_dt_raw", il);
-
-            // Apply RMS norm to dt, B, C (PLaMo-2 specific)
-            B = build_norm(B, model.layers[il].ssm_b_norm, NULL, LLM_NORM_RMS, il);
-            C = build_norm(C, model.layers[il].ssm_c_norm, NULL, LLM_NORM_RMS, il);
-            dt = build_norm(dt, model.layers[il].ssm_dt_norm, NULL, LLM_NORM_RMS, il);
-            cb(B, "mamba_B_normed", il);
-            cb(C, "mamba_C_normed", il);
-            cb(dt, "mamba_dt_normed", il);
-
-            // dt_proj: {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
-            dt = build_lora_mm(model.layers[il].ssm_dt, dt);
-            dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b);
-            cb(dt, "mamba_dt_proj", il);
-
-            ggml_tensor * A = ggml_reshape_2d(ctx0, model.layers[il].ssm_a, 1, n_heads);
-            cb(A, "mamba_A", il);
-
-            x = ggml_view_4d(ctx0, x, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x), head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
-            B = ggml_view_4d(ctx0, B, d_state, 1, n_seq_tokens, n_seqs, d_state * B->nb[0], B->nb[1], B->nb[2], 0);
-            C = ggml_view_4d(ctx0, C, d_state, 1, n_seq_tokens, n_seqs, d_state * C->nb[0], C->nb[1], C->nb[2], 0);
-
-            // use the states and the indices provided by build_recurrent_state
-            // (this is necessary in order to properly use the states before they are overwritten,
-            //  while avoiding to make unnecessary copies of the states)
-            auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
-                ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_heads, mctx_cur->get_size());
-
-                // Custom operator to optimize the parallel associative scan
-                // as described in the Annex D of the Mamba paper.
-                // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
-                return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
-            };
-
-            ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
-            cb(y_ssm, "mamba_ssm_scan", il);
-
-            // store last states
-            ggml_build_forward_expand(gf,
-                ggml_cpy(ctx0,
-                    ggml_view_1d(ctx0, y_ssm, n_heads*head_dim*d_state*n_seqs, n_heads*head_dim*n_seq_tokens*n_seqs*ggml_element_size(y_ssm)),
-                    ggml_view_1d(ctx0, ssm_states_all, n_heads*head_dim*d_state*n_seqs, kv_head*n_seqs*n_heads*head_dim*d_state*ggml_element_size(ssm_states_all))));
-            cb(ssm_states_all, "mamba_ssm_states", il);
-
-            ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x), head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
-            cb(y, "mamba_y_view", il);
-
-            // Add D parameter and apply gating with z
-            // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs}
-            ggml_tensor * D = ggml_reshape_2d(ctx0, model.layers[il].ssm_d, 1, n_heads);
-            y = ggml_add(ctx0, y, ggml_mul(ctx0, x, D));
-            cb(y, "mamba_y_add_d", il);
-
-            y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
-            cb(y, "mamba_y_swiglu_z", il);
-
-            // out_proj: {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
-            y = ggml_view_3d(ctx0, y, head_dim * n_heads, n_seq_tokens, n_seqs, y->nb[2], y->nb[3], 0);
-            cur = build_lora_mm(model.layers[il].ssm_out, y);
-            cb(cur, "mamba_out_proj", il);
-        }
-
-        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
-        cb(cur, "mamba_out", il);
-
-        return cur;
-    }
-};
-
-struct llm_build_arcee : public llm_graph_context {
-    llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            // ARCEE uses relu^2 instead of silu
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    NULL,                      NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_hunyuan_moe : public llm_graph_context {
-    llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = build_norm(Kcur,
-                        model.layers[il].attn_k_norm, nullptr,
-                        LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_norm", il);
-
-                Qcur = build_norm(Qcur,
-                        model.layers[il].attn_q_norm, nullptr,
-                        LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_norm", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = build_norm(ffn_inp,
-                model.layers[il].ffn_norm, NULL,
-                LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network (non-MoE)
-            ggml_tensor * cur_mlp = build_ffn(cur,
-                    model.layers[il].ffn_up_shexp,   NULL, NULL,
-                    model.layers[il].ffn_gate_shexp, NULL, NULL,
-                    model.layers[il].ffn_down_shexp, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur_mlp, "ffn_mlp", il);
-
-            // MoE branch
-            ggml_tensor * cur_moe = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU,
-                    true, // norm_topk_prob
-                    false,
-                    0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
-            cb(cur_moe, "ffn_moe_out", il);
-
-            ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp);
-            cb(ffn_out, "ffn_out", il);
-
-            cur = ggml_add(ctx0, ffn_out, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_hunyuan_dense : public llm_graph_context {
-    llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                         ctx0, Qcur, inp_pos, rope_factors,
-                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                         ext_factor, attn_factor, beta_fast, beta_slow
-                         );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Kcur = ggml_rope_ext(
-                         ctx0, Kcur, inp_pos, rope_factors,
-                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                         ext_factor, attn_factor, beta_fast, beta_slow
-                         );
-
-                Kcur = build_norm(Kcur,
-                         model.layers[il].attn_k_norm, nullptr,
-                         LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_norm", il);
-
-                Qcur = build_norm(Qcur,
-                         model.layers[il].attn_q_norm, nullptr,
-                         LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_norm", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-            // feed-forward network (non-MoE)
-            ggml_tensor * cur_mlp = build_ffn(cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur_mlp, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur_mlp, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_smollm3 : public llm_graph_context {
-    llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                if (use_rope) {
-                    Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, nullptr,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                    Kcur = ggml_rope_ext(
-                            ctx0, Kcur, inp_pos, nullptr,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_openai_moe_iswa : public llm_graph_context {
-    llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv_iswa();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, nullptr,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il);
-
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = ffn_inp;
-            cur = build_norm(cur,
-                    model.layers[il].attn_post_norm, nullptr,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            // MoE branch
-            cur = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,  model.layers[il].ffn_gate_inp_b,
-                    model.layers[il].ffn_up_exps,   model.layers[il].ffn_up_exps_b,
-                    model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b,
-                    model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SWIGLU_OAI_MOE, false,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT,
-                    il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_lfm2 : public llm_graph_context {
-    const llama_model & model;
-
-    llm_build_lfm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) {
-
-        ggml_tensor * cur = build_inp_embd(model.tok_embd);
-        cb(cur, "model.embed_tokens", -1);
-
-        ggml_tensor * inp_pos     = build_inp_pos();
-        auto        * inp_hybrid  = build_inp_mem_hybrid();
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);
-
-            auto * prev_cur = cur;
-            cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "model.layers.{}.operator_norm", il);
-
-            cur = hparams.is_recurrent(il) ?
-                build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
-                build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il) ;
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur      = ggml_get_rows(ctx0,      cur, inp_out_ids);
-                prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
-            }
-
-            cur = ggml_add(ctx0, prev_cur, cur);
-
-            auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
-            cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
-
-            ggml_tensor * ffn_out = is_moe_layer ?
-                build_moe_feed_forward(ffn_norm_out, il) :
-                build_dense_feed_forward(ffn_norm_out, il);
-            cb(ffn_norm_out, "model.layers.{}.ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_out);
-        }
-
-        cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1);
-        cb(cur, "model.embedding_norm", -1);
-        res->t_embd = cur;
-
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "lm_head", -1);
-
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-    ggml_tensor * build_moe_feed_forward(ggml_tensor * cur,
-                                         int           il) const {
-        return build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    model.layers[il].ffn_exp_probs_b,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
-                    il);
-    }
-
-    ggml_tensor * build_dense_feed_forward(ggml_tensor * cur,
-                                           int           il) const {
-        GGML_ASSERT(!model.layers[il].ffn_up_b);
-        GGML_ASSERT(!model.layers[il].ffn_gate_b);
-        GGML_ASSERT(!model.layers[il].ffn_down_b);
-        return build_ffn(cur,
-                model.layers[il].ffn_up,   NULL, NULL,
-                model.layers[il].ffn_gate, NULL, NULL,
-                model.layers[il].ffn_down, NULL, NULL,
-                NULL,
-                LLM_FFN_SILU, LLM_FFN_PAR, il);
-    }
-
-    ggml_tensor * build_attn_block(ggml_tensor             * cur,
-                                   ggml_tensor             * inp_pos,
-                                   llm_graph_input_attn_kv * inp_attn,
-                                   int                     il) const {
-        GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
-        auto const n_embd_head = hparams.n_embd_head_v;
-        auto const n_head_kv = hparams.n_head_kv(il);
-
-        auto * q = build_lora_mm(model.layers[il].wq, cur);
-        cb(q, "model.layers.{}.self_attn.q_proj", il);
-        auto * k = build_lora_mm(model.layers[il].wk, cur);
-        cb(k, "model.layers.{}.self_attn.k_proj", il);
-        auto * v = build_lora_mm(model.layers[il].wv, cur);
-        cb(v, "model.layers.{}.self_attn.v_proj", il);
-
-        q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head,    n_tokens);
-        k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
-        v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
-
-        // qk norm
-        q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-        cb(q, "model.layers.{}.self_attn.q_layernorm", il);
-        k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-        cb(k, "model.layers.{}.self_attn.k_layernorm", il);
-
-        // RoPE
-        q = ggml_rope_ext(
-                ctx0, q, inp_pos, nullptr,
-                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                ext_factor, attn_factor, beta_fast, beta_slow
-                );
-        k = ggml_rope_ext(
-                ctx0, k, inp_pos, nullptr,
-                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                ext_factor, attn_factor, beta_fast, beta_slow
-                );
-
-        cur = build_attn(inp_attn, model.layers[il].wo, NULL,
-                q, k, v, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-
-        cb(cur, "model.layers.{}.self_attn.out_proj", il);
-
-        return cur;
-    }
-
-    ggml_tensor * build_shortconv_block(ggml_tensor        * cur,
-                                        llm_graph_input_rs * inp_recr,
-                                        int                il) {
-        const auto *   mctx_cur     = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
-        const uint32_t kv_head      = mctx_cur->get_head();
-        const int64_t  n_seq_tokens = ubatch.n_seq_tokens;
-        const int64_t  n_seqs       = ubatch.n_seqs;
-        GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(ubatch.equal_seqs());
-        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
-
-        GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
-        const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
-
-        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
-        auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
-        cb(bcx, "model.layers.{}.conv.in_proj", il);
-
-        constexpr auto n_chunks = 3;
-        GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
-        auto const chunk_size = bcx->ne[0] / n_chunks;
-        auto * b = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], 0*chunk_size*ggml_element_size(bcx));
-        auto * c = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], 1*chunk_size*ggml_element_size(bcx));
-        auto * x = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], 2*chunk_size*ggml_element_size(bcx));
-
-        auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
-
-        // read conv state
-        auto * conv_state = mctx_cur->get_r_l(il);
-        auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
-        auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
-
-        bx = ggml_concat(ctx0, conv, bx, 0);
-        GGML_ASSERT(bx->ne[0] > conv->ne[0]);
-
-        // last d_conv columns is a new conv state
-        auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2], (bx->ne[0] - conv->ne[0])*ggml_element_size(bx));
-        GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
-
-        // write new conv conv state
-        ggml_build_forward_expand(
-                gf,
-                ggml_cpy(
-                    ctx0,
-                    new_conv,
-                    ggml_view_1d(
-                        ctx0,
-                        conv_state,
-                        ggml_nelements(new_conv),
-                        kv_head*d_conv*n_embd*ggml_element_size(new_conv)
-                        )
-                    )
-                );
-
-        auto * conv_kernel = model.layers[il].shortconv.conv;
-        auto * conv_out = ggml_ssm_conv(ctx0, bx, conv_kernel);
-        cb(conv_out, "model.layers.{}.conv.conv", il);
-
-        auto * y = ggml_mul(ctx0, c, conv_out);
-        y = build_lora_mm(model.layers[il].shortconv.out_proj, y);
-        cb(y, "model.layers.{}.conv.out_proj", il);
-        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-        y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);
-
-        return y;
-    }
-};
-
-struct llm_build_seed_oss : public llm_graph_context {
-    llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = build_norm(ffn_inp,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_post_norm", il);
-
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-template <bool iswa>
-struct llm_build_smallthinker : public llm_graph_context{
-    llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
-        inp_attn_type * inp_attn = nullptr;
-
-        if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_iswa();
-        } else {
-            inp_attn = build_attn_inp_kv();
-        }
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA  = inpL;
-            ggml_tensor * probs  = nullptr;
-
-            probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
-            cb(probs, "ffn_moe_logits", il);
-
-            // norm
-            cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                if (hparams.n_no_rope_layer_step == n_layer || il % hparams.n_no_rope_layer_step != 0) {
-                    Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                     ext_factor, attn_factor, beta_fast, beta_slow);
-
-                    Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                                     ext_factor, attn_factor, beta_fast, beta_slow);
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur = ggml_get_rows(ctx0, cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-                probs = ggml_get_rows(ctx0, probs, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            ggml_tensor * ffn_out =
-                build_moe_ffn(cur,
-                        nullptr,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_RELU, true,
-                        false, 0.0,
-                        static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
-                        il, probs);
-
-            cb(ffn_out, "ffn_out", il);
-            cur = ffn_out;
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_grovemoe : public llm_graph_context {
-    llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv();
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = build_norm(ffn_inp,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "ffn_norm", il);
-
-            ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, cur); // [n_expert, n_tokens]
-            cb(probs, "ffn_moe_logits", il);
-
-            ggml_tensor * moe_out =
-                build_moe_ffn(cur,
-                        nullptr,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il, probs);
-            cb(moe_out, "ffn_moe_out", il);
-            cur = moe_out;
-
-            // TODO: Only do the expert selection and weights once
-            moe_out =
-                build_moe_ffn(cur,
-                        nullptr,
-                        model.layers[il].ffn_up_chexps,
-                        model.layers[il].ffn_gate_chexps,
-                        model.layers[il].ffn_down_chexps,
-                        nullptr,
-                        n_chunk_expert, n_expert_used > n_chunk_expert ? n_chunk_expert : n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il, probs);
-            cb(moe_out, "ffn_adj_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ggml_scale(ctx0, moe_out, hparams.expert_group_scale));
-            cb(cur, "ffn_final_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-struct llm_build_apertus : public llm_graph_context {
-    llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-
-        inpL = build_inp_embd(model.tok_embd);
-
-        ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
-        ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, nullptr,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur_pos", il);
-                cb(Kcur, "Kcur_pos", il);
-                cb(Vcur, "Vcur_pos", il);
-
-                cur = build_attn(inp_attn,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
-
-            if (il == n_layer - 1 && inp_out_ids) {
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network with xIELU activation
-            {
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, nullptr,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                // Up projection
-                ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
-                cb(up, "ffn_up", il);
-
-                float alpha_n_val = hparams.xielu_alpha_n[il];
-                float alpha_p_val = hparams.xielu_alpha_p[il];
-                float beta_val = hparams.xielu_beta[il];
-                float eps_val = hparams.xielu_eps[il];
-
-                // Apply xIELU activation
-                ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
-                cb(activated, "ffn_xielu", il);
-
-                // Down projection
-                cur = build_lora_mm(model.layers[il].ffn_down, activated);
-                cb(cur, "ffn_down", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, nullptr,
-                LLM_NORM_RMS, -1);
-
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
-llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
+llama_memory_i * llama_model::create_memory(const llama_memory_params & params, const llama_cparams & cparams) const {
     llama_memory_i * res;
 
     switch (arch) {
@@ -19853,6 +7030,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
             {
                 res = nullptr;
             } break;
@@ -19887,17 +7065,13 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         };
                     }
 
-                    const auto padding = llama_kv_cache::get_padding(cparams);
-
-                    cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
-
                     res = new llama_memory_hybrid(
                         /* model             */ *this,
                         /* attn_type_k       */ params.type_k,
                         /* attn_type_v       */ params.type_v,
                         /* attn_v_trans      */ !cparams.flash_attn,
                         /* attn_kv_size      */ cparams.n_ctx,
-                        /* attn_n_pad        */ padding,
+                        /* attn_n_pad        */ 1,
                         /* attn_n_swa        */ hparams.n_swa,
                         /* attn_swa_type     */ hparams.swa_type,
                         /* recurrent_type_k  */ GGML_TYPE_F32,
@@ -19909,23 +7083,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         /* filter_attn       */ std::move(filter_attn),
                         /* filter_recr       */ std::move(filter_recr));
                 } else {
-                    const auto padding = llama_kv_cache::get_padding(cparams);
-
-                    uint32_t n_ctx_per_stream = cparams.n_ctx;
-
-                    if (!cparams.kv_unified) {
-                        n_ctx_per_stream = (cparams.n_ctx + cparams.n_seq_max - 1)/cparams.n_seq_max;
-                        n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding);
-
-                        cparams.n_ctx = n_ctx_per_stream*cparams.n_seq_max;
-                    } else {
-                        n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding);
-
-                        cparams.n_ctx = n_ctx_per_stream;
-                    }
-
-                    LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
-
                     llama_memory_i::layer_reuse_cb reuse = nullptr;
 
                     if (arch == LLM_ARCH_GEMMA3N) {
@@ -19949,10 +7106,10 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 cparams.offload_kqv,
                                 params.swa_full,
                                 cparams.kv_unified,
-                                n_ctx_per_stream,
+                                cparams.n_ctx_seq,
                                 cparams.n_seq_max,
                                 cparams.n_ubatch,
-                                padding,
+                                1,
                                 nullptr,
                                 reuse);
                     } else {
@@ -19965,9 +7122,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 !cparams.flash_attn,
                                 cparams.offload_kqv,
                                 cparams.kv_unified,
-                                n_ctx_per_stream,
+                                cparams.n_ctx_seq,
                                 cparams.n_seq_max,
-                                padding,
+                                1,
                                 hparams.n_swa,
                                 hparams.swa_type,
                                 nullptr,
@@ -20067,6 +7224,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_llada_moe>(*this, params);
             }
             break;
+        case LLM_ARCH_RND1:
+            {
+                llm = std::make_unique<llm_build_rnd1>(*this, params);
+            }
+            break;
         case LLM_ARCH_QWEN2VL:
             {
                 llm = std::make_unique<llm_build_qwen2vl>(*this, params);
@@ -20083,6 +7245,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_qwen3moe>(*this, params);
             } break;
+        case LLM_ARCH_QWEN3VL:
+            {
+                llm = std::make_unique<llm_build_qwen3vl>(*this, params);
+            } break;
+        case LLM_ARCH_QWEN3VLMOE:
+            {
+                llm = std::make_unique<llm_build_qwen3vlmoe>(*this, params);
+            } break;
         case LLM_ARCH_PHI2:
             {
                 llm = std::make_unique<llm_build_phi2>(*this, params);
@@ -20330,6 +7500,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_arcee>(*this, params);
             } break;
+        case LLM_ARCH_AFMOE:
+            {
+                llm = std::make_unique<llm_build_afmoe>(*this, params);
+            } break;
         case LLM_ARCH_ERNIE4_5:
             {
                 llm = std::make_unique<llm_build_ernie4_5>(*this, params);
@@ -20379,6 +7553,22 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_apertus>(*this, params);
             } break;
+        case LLM_ARCH_MINIMAX_M2:
+            {
+                llm = std::make_unique<llm_build_minimax_m2>(*this, params);
+            } break;
+        case LLM_ARCH_COGVLM:
+            {
+                llm = std::make_unique<llm_build_cogvlm>(*this, params);
+            } break;
+        case LLM_ARCH_PANGU_EMBED:
+            {
+                llm = std::make_unique<llm_build_pangu_embedded>(*this, params);
+            } break;
+        case LLM_ARCH_QWEN3NEXT:
+            {
+                llm = std::make_unique<llm_build_qwen3next>(*this, params);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -20442,6 +7632,10 @@ int32_t llama_model_n_embd(const llama_model * model) {
     return model->hparams.n_embd;
 }
 
+int32_t llama_model_n_embd_inp(const llama_model * model) {
+    return model->hparams.n_embd_inp();
+}
+
 int32_t llama_model_n_layer(const llama_model * model) {
     return model->hparams.n_layer;
 }
@@ -20564,6 +7758,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_QWEN3:
         case LLM_ARCH_QWEN3MOE:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
         case LLM_ARCH_OLMO2:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
@@ -20597,10 +7792,18 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_SEED_OSS:
         case LLM_ARCH_GROVEMOE:
         case LLM_ARCH_APERTUS:
+        case LLM_ARCH_MINIMAX_M2:
+        case LLM_ARCH_COGVLM:
+        case LLM_ARCH_PANGU_EMBED:
+        case LLM_ARCH_AFMOE:
+        case LLM_ARCH_QWEN3NEXT:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:
             return LLAMA_ROPE_TYPE_MROPE;
+        case LLM_ARCH_QWEN3VL:
+        case LLM_ARCH_QWEN3VLMOE:
+            return LLAMA_ROPE_TYPE_IMROPE;
 
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
@@ -20629,6 +7832,24 @@ int32_t llama_model_meta_count(const llama_model * model) {
     return (int)model->gguf_kv.size();
 }
 
+const char * llama_model_meta_key_str(llama_model_meta_key key) {
+    switch (key) {
+        case LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE:        return "general.sampling.sequence";
+        case LLAMA_MODEL_META_KEY_SAMPLING_TOP_K:           return "general.sampling.top_k";
+        case LLAMA_MODEL_META_KEY_SAMPLING_TOP_P:           return "general.sampling.top_p";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIN_P:           return "general.sampling.min_p";
+        case LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY: return "general.sampling.xtc_probability";
+        case LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD:   return "general.sampling.xtc_threshold";
+        case LLAMA_MODEL_META_KEY_SAMPLING_TEMP:            return "general.sampling.temp";
+        case LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N:  return "general.sampling.penalty_last_n";
+        case LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT:  return "general.sampling.penalty_repeat";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT:        return "general.sampling.mirostat";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU:    return "general.sampling.mirostat_tau";
+        case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA:    return "general.sampling.mirostat_eta";
+        default:                                            return nullptr;
+    }
+}
+
 int32_t llama_model_meta_key_by_index(const llama_model * model, int i, char * buf, size_t buf_size) {
     if (i < 0 || i >= (int)model->gguf_kv.size()) {
         if (buf_size > 0) {
diff --git a/llama/llama.cpp/src/llama-model.h b/llama/llama.cpp/src/llama-model.h
index 4a7924aa..cbf4e1bf 100644
--- a/llama/llama.cpp/src/llama-model.h
+++ b/llama/llama.cpp/src/llama-model.h
@@ -77,6 +77,7 @@ enum llm_type {
     LLM_TYPE_16B,
     LLM_TYPE_20B,
     LLM_TYPE_22B,
+    LLM_TYPE_26B,
     LLM_TYPE_27B,
     LLM_TYPE_30B,
     LLM_TYPE_32B,
@@ -113,8 +114,10 @@ enum llm_type {
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
+    LLM_TYPE_230B_A10B, // Minimax M2
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_355B_A32B, // GLM-4.5
@@ -234,6 +237,7 @@ struct llama_layer {
     struct ggml_tensor * wk_enc    = nullptr;
     struct ggml_tensor * wv_enc    = nullptr;
     struct ggml_tensor * wo_enc    = nullptr;
+    struct ggml_tensor * wqkv_gate = nullptr;
 
     // attention bias
     struct ggml_tensor * bq   = nullptr;
@@ -307,6 +311,9 @@ struct llama_layer {
     struct ggml_tensor * ssm_conv1d_b = nullptr;
     struct ggml_tensor * ssm_dt_b     = nullptr;
 
+    // qwen3next
+    struct ggml_tensor * ssm_beta_alpha = nullptr;
+
     // rwkv
     struct ggml_tensor * time_mix_w1         = nullptr;
     struct ggml_tensor * time_mix_w2         = nullptr;
@@ -385,6 +392,13 @@ struct llama_layer {
     // openai-moe
     struct ggml_tensor * attn_sinks = nullptr;
 
+    // cogvlm
+    struct ggml_tensor * visexp_attn_wqkv = nullptr;
+    struct ggml_tensor * visexp_attn_wo   = nullptr;
+    struct ggml_tensor * visexp_ffn_gate  = nullptr;
+    struct ggml_tensor * visexp_ffn_down  = nullptr;
+    struct ggml_tensor * visexp_ffn_up    = nullptr;
+
     // xIELU activation parameters for Apertus
     struct ggml_tensor * ffn_act_alpha_n = nullptr;
     struct ggml_tensor * ffn_act_alpha_p = nullptr;
@@ -503,9 +517,8 @@ struct llama_model {
 
     ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;
 
-    // note: can mutate `cparams`
     // TODO: move this to new llm_arch_model_i interface
-    llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
+    llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
     ggml_cgraph * build_graph(const llm_graph_params & params) const;
diff --git a/llama/llama.cpp/src/llama-quant.cpp b/llama/llama.cpp/src/llama-quant.cpp
index 6dd40412..0b23eaef 100644
--- a/llama/llama.cpp/src/llama-quant.cpp
+++ b/llama/llama.cpp/src/llama-quant.cpp
@@ -653,7 +653,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
                 // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
-                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)std::abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
@@ -681,7 +681,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             }
             LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
             continue;
-        } else if (remapped_name != it.first) {
+        }
+
+        if (remapped_name != it.first) {
             ggml_set_name(it.second.tensor, remapped_name.c_str());
             LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
         }
@@ -726,13 +728,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads
-        int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
+        int32_t n_layer_attn = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
         if (llama_model_has_encoder(&model)) {
-            // now n_attn_layer is the number of attention layers in the encoder
+            // now n_layer_attn is the number of attention layers in the encoder
             // for each decoder block, there are 2 attention layers
-            n_attn_layer += 2 * model.hparams.dec_n_layer;
+            n_layer_attn += 2 * model.hparams.dec_n_layer;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
+
+        // note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
+        const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
+
+        LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
+
+        GGML_ASSERT((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
     }
 
     size_t total_size_org = 0;
diff --git a/llama/llama.cpp/src/llama-sampling.cpp b/llama/llama.cpp/src/llama-sampling.cpp
index da34526b..38a30ea0 100644
--- a/llama/llama.cpp/src/llama-sampling.cpp
+++ b/llama/llama.cpp/src/llama-sampling.cpp
@@ -4,6 +4,7 @@
 #include "llama-vocab.h"
 #include "llama-grammar.h"
 
+#include <array>
 #include <algorithm>
 #include <cassert>
 #include <cfloat>
@@ -471,9 +472,6 @@ static void llama_sampler_chain_reset(struct llama_sampler * smpl) {
     for (auto * smpl : chain->samplers) {
         llama_sampler_reset(smpl);
     }
-
-    chain->t_sample_us = 0;
-    chain->n_sample    = 0;
 }
 
 static struct llama_sampler * llama_sampler_chain_clone(const struct llama_sampler * smpl) {
@@ -1625,10 +1623,12 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
+        std::string trigger_pattern;
+        llama_grammar * grammar = nullptr;
         // TODO: remove trigger_words support.
         if (trigger_words != nullptr && num_trigger_words > 0) {
             GGML_ASSERT(trigger_patterns == nullptr && num_trigger_patterns == 0);
-            std::string trigger_pattern("[\\s\\S]*?(");
+            trigger_pattern = "[\\s\\S]*?(";
             for (size_t i = 0; i < num_trigger_words; ++i) {
                 static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
                 if (i > 0) {
@@ -1637,15 +1637,17 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                 trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
             }
             trigger_pattern += ")[\\s\\S]*";
-            const auto * trigger_pattern_c = trigger_pattern.c_str();
-            trigger_patterns = &trigger_pattern_c;
-            num_trigger_patterns = 1;
+
+            std::array<const char *, 1> tmp_trigger_patterns = { trigger_pattern.c_str() };
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, tmp_trigger_patterns.data(), tmp_trigger_patterns.size(), trigger_tokens, num_trigger_tokens);
+        } else {
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens);
         }
         *ctx = {
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
+            /* .grammar      = */ grammar,
         };
         if (!ctx->grammar) {
             delete ctx;
@@ -2665,8 +2667,7 @@ struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * c
 void llama_perf_sampler_print(const struct llama_sampler * chain) {
     const auto data = llama_perf_sampler(chain);
 
-    LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
+    LLAMA_LOG_INFO("%s:    samplers time = %10.2f ms / %5d runs\n", __func__, data.t_sample_ms, data.n_sample);
 }
 
 void llama_perf_sampler_reset(struct llama_sampler * chain) {
@@ -2676,5 +2677,6 @@ void llama_perf_sampler_reset(struct llama_sampler * chain) {
 
     auto * ctx = (struct llama_sampler_chain *) chain->ctx;
 
-    ctx->t_sample_us = ctx->n_sample = 0;
+    ctx->t_sample_us = 0;
+    ctx->n_sample    = 0;
 }
diff --git a/llama/llama.cpp/src/llama-vocab.cpp b/llama/llama.cpp/src/llama-vocab.cpp
index 31f49801..ea450c36 100644
--- a/llama/llama.cpp/src/llama-vocab.cpp
+++ b/llama/llama.cpp/src/llama-vocab.cpp
@@ -401,6 +401,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+            case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -442,6 +443,17 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_AFMOE:
+                regex_exprs = {
+                    // Digit handling - uses custom implementation in unicode.cpp
+                    // Groups digits with leading 1-2 based on total length modulo 3
+                    "\\p{AFMoE_digits}",
+                    // CJK and Asian scripts (using direct Unicode literals)
+                    "[一-鿿㐀-䶿豈-﫿぀-ゟ゠-ヿ･-ﾟ⼀-⿟เ-๿຀-໿ក-៿က-႟ꩠ-ꩿꧠ-꧿가-힯ᄀ-ᇿ]+",
+                    // Main BPE pattern
+                    "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\\r\\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1012,7 +1024,7 @@ private:
         }
     private:
         uint32_t get_node(size_t index) {
-            if (index > xcda_array_size) {
+            if (index >= xcda_array_size) {
                 throw std::runtime_error("Index out of array bounds in XCDA array!");
             }
             return xcda_array[index];
@@ -1269,6 +1281,7 @@ struct llm_tokenizer_plamo2 : llm_tokenizer {
 
         // Build suffix list in lexicographical order of reversed strings
         std::vector<std::string> suffixes;
+        suffixes.reserve(suffix_to_score.size() + 1);
         for (const auto & pair : suffix_to_score) {
             suffixes.push_back(pair.first);
         }
@@ -1981,6 +1994,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "grok-2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "afmoe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_AFMOE;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minimax-m2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
+                clean_spaces = false;
             } else {
                 LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
diff --git a/llama/llama.cpp/src/llama-vocab.h b/llama/llama.cpp/src/llama-vocab.h
index 5e468675..55f8f392 100644
--- a/llama/llama.cpp/src/llama-vocab.h
+++ b/llama/llama.cpp/src/llama-vocab.h
@@ -49,6 +49,8 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE   = 38,
     LLAMA_VOCAB_PRE_TYPE_GROK_2          = 39,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
+    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
+    LLAMA_VOCAB_PRE_TYPE_AFMOE           = 42,
 };
 
 struct LLM_KV;
diff --git a/llama/llama.cpp/src/llama.go b/llama/llama.cpp/src/llama.go
index ddbd5378..1face83a 100644
--- a/llama/llama.cpp/src/llama.go
+++ b/llama/llama.cpp/src/llama.go
@@ -5,4 +5,8 @@ package llama
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../ml/backend/ggml/ggml/include
 // #cgo windows CPPFLAGS: -D_WIN32_WINNT=0x0602
 import "C"
-import _ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+
+import (
+	_ "github.com/ollama/ollama/llama/llama.cpp/src/models"
+	_ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+)
diff --git a/llama/llama.cpp/src/models/afmoe.cpp b/llama/llama.cpp/src/models/afmoe.cpp
new file mode 100644
index 00000000..0192e344
--- /dev/null
+++ b/llama/llama.cpp/src/models/afmoe.cpp
@@ -0,0 +1,187 @@
+#include "models.h"
+
+llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // MuP scaling: embeddings * sqrt(hidden_size)
+    // mup_enabled = true, hidden_size = 1024, scale = 32.0
+    inpL = ggml_scale(ctx0, inpL, sqrtf(float(n_embd)));
+    cb(inpL, "inp_embd_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // dual attention normalization (pre)
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * attn_inp = cur;  // save input for gate computation
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            // compute gate from input
+            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+            cb(gate, "attn_gate_proj", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+            // Q/K normalization
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            // RoPE only for sliding_attention layers
+            const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
+                                ((il + 1) % hparams.n_no_rope_layer_step) != 0;
+            if (use_rope) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur_rope", il);
+            }
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,  // wo will be applied after gating
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+
+            // attention gating: attn_out * sigmoid(gate) BEFORE o_proj
+            gate = ggml_sigmoid(ctx0, gate);
+            cb(gate, "attn_gate_sig", il);
+            cur = ggml_mul(ctx0, cur, gate);
+            cb(cur, "attn_gated", il);
+
+            // now apply output projection
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "attn_o_proj", il);
+        }
+
+        // dual attention normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // dual ffn normalization (pre)
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // MoE or dense FFN
+        if ((uint32_t)il >= hparams.n_layer_dense_lead) {
+            // MoE layer with sigmoid routing, normalization, and scaling
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU,
+                    hparams.expert_weights_norm,           // norm_w (route_norm=True)
+                    hparams.expert_weights_scale,          // scale_w
+                    hparams.expert_weights_scale,          // w_scale (route_scale=2.826)
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // shared expert
+            if (hparams.n_expert_shared > 0) {
+                ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            } else {
+                cur = moe_out;
+            }
+        } else {
+            // dense layer
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // dual ffn normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/apertus.cpp b/llama/llama.cpp/src/models/apertus.cpp
new file mode 100644
index 00000000..9af19c1b
--- /dev/null
+++ b/llama/llama.cpp/src/models/apertus.cpp
@@ -0,0 +1,125 @@
+#include "models.h"
+
+
+
+llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos  = build_inp_pos();
+    auto *        inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+            cb(Vcur, "Vcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network with xIELU activation
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // Up projection
+            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
+            cb(up, "ffn_up", il);
+
+            float alpha_n_val = hparams.xielu_alpha_n[il];
+            float alpha_p_val = hparams.xielu_alpha_p[il];
+            float beta_val    = hparams.xielu_beta[il];
+            float eps_val     = hparams.xielu_eps[il];
+
+            // Apply xIELU activation
+            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
+            cb(activated, "ffn_xielu", il);
+
+            // Down projection
+            cur = build_lora_mm(model.layers[il].ffn_down, activated);
+            cb(cur, "ffn_down", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/arcee.cpp b/llama/llama.cpp/src/models/arcee.cpp
new file mode 100644
index 00000000..aa6167db
--- /dev/null
+++ b/llama/llama.cpp/src/models/arcee.cpp
@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // ARCEE uses relu^2 instead of silu
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                NULL,                      NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/arctic.cpp b/llama/llama.cpp/src/models/arctic.cpp
new file mode 100644
index 00000000..e8f028a7
--- /dev/null
+++ b/llama/llama.cpp/src/models/arctic.cpp
@@ -0,0 +1,138 @@
+#include "models.h"
+
+
+llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+        cb(ffn_out, "ffn_out", il);
+
+        // MoE
+        cur = build_norm(inpSA,
+                model.layers[il].ffn_norm_exps, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm_exps", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_out);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/arwkv7.cpp b/llama/llama.cpp/src/models/arwkv7.cpp
new file mode 100644
index 00000000..107a3bef
--- /dev/null
+++ b/llama/llama.cpp/src/models/arwkv7.cpp
@@ -0,0 +1,86 @@
+#include "models.h"
+
+
+llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
+    GGML_ASSERT(n_embd == hparams.n_embd_r());
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * v_first = nullptr;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+                ctx0,
+                token_shift,
+                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
+                1
+                );
+
+        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
+
+        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/baichuan.cpp b/llama/llama.cpp/src/models/baichuan.cpp
new file mode 100644
index 00000000..c04b0c98
--- /dev/null
+++ b/llama/llama.cpp/src/models/baichuan.cpp
@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            switch (model.type) {
+                case LLM_TYPE_7B:
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    break;
+                case LLM_TYPE_13B:
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/bailingmoe.cpp b/llama/llama.cpp/src/models/bailingmoe.cpp
new file mode 100644
index 00000000..ed56b9c4
--- /dev/null
+++ b/llama/llama.cpp/src/models/bailingmoe.cpp
@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    false, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // FFN shared expert
+        {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/bailingmoe2.cpp b/llama/llama.cpp/src/models/bailingmoe2.cpp
new file mode 100644
index 00000000..fbf7b210
--- /dev/null
+++ b/llama/llama.cpp/src/models/bailingmoe2.cpp
@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 0 * sizeof(float) * (n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
+        cb(sa_out, "sa_out", il);
+
+        // MoE branch
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/bert.cpp b/llama/llama.cpp/src/models/bert.cpp
new file mode 100644
index 00000000..3274fa3b
--- /dev/null
+++ b/llama/llama.cpp/src/models/bert.cpp
@@ -0,0 +1,176 @@
+#include "models.h"
+
+
+
+llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * inp_pos = nullptr;
+
+    if (model.arch != LLM_ARCH_JINA_BERT_V2) {
+        inp_pos = build_inp_pos();
+    }
+
+    // construct input embeddings (token, type, position)
+    inpL = build_inp_embd(model.tok_embd);
+
+    // token types are hardcoded to zero ("Sentence A")
+    if (model.type_embd) {
+        ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
+        inpL                    = ggml_add(ctx0, inpL, type_row0);
+    }
+    if (model.arch == LLM_ARCH_BERT) {
+        inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
+    }
+    cb(inpL, "inp_embd", -1);
+
+    // embed layer norm
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * cur = inpL;
+
+        {
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            // self-attention
+            if (model.layers[il].wqkv) {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            } else {
+                Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
+                Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
+                Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
+
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            // RoPE
+            if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                model.arch == LLM_ARCH_JINA_BERT_V3) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "kqv_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // re-add the layer input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        // attention layer norm
+        cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);
+
+        if (model.layers[il].attn_norm_2 != nullptr) {
+            cur = ggml_add(ctx0, cur, inpL);  // re-add the layer input
+            cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
+            // MoE branch
+            cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, nullptr,
+                                model.layers[il].ffn_down_exps, nullptr, hparams.n_expert, hparams.n_expert_used,
+                                LLM_FFN_GELU, false, false, 0.0f, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+            cb(cur, "ffn_moe_out", il);
+        } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                   model.arch == LLM_ARCH_JINA_BERT_V3) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // attentions bypass the intermediate layer
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        // output layer norm
+        cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cb(cur, "result_embd", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/bitnet.cpp b/llama/llama.cpp/src/models/bitnet.cpp
new file mode 100644
index 00000000..331a3f11
--- /dev/null
+++ b/llama/llama.cpp/src/models/bitnet.cpp
@@ -0,0 +1,160 @@
+#include "models.h"
+
+
+llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            if (model.layers[il].wq_scale) {
+                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
+            }
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            // B1.K
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            if (model.layers[il].wk_scale) {
+                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
+            }
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            // B1.V
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            if (model.layers[il].wv_scale) {
+                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
+            }
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+
+            cur = build_norm(cur,
+                    model.layers[il].attn_sub_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_sub_norm", il);
+
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            if (model.layers[il].wo_scale) {
+                cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
+            }
+            if (model.layers[il].bo) {
+                cur = ggml_add(ctx0, cur, model.layers[il].bo);
+            }
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward forward
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_scale,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
+                NULL,                      NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_sub_out", il);
+
+        cur = build_norm(cur,
+                model.layers[il].ffn_sub_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_sub_norm", il);
+
+        cur = build_lora_mm(model.layers[il].ffn_down, cur);
+        if (model.layers[il].ffn_down_scale) {
+            cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
+        }
+        cb(cur, "ffn_down", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    // FIXME: do not use model.tok_embd directly, duplicate as model.output
+    cur = build_lora_mm(model.tok_embd, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/bloom.cpp b/llama/llama.cpp/src/models/bloom.cpp
new file mode 100644
index 00000000..2c552d1d
--- /dev/null
+++ b/llama/llama.cpp/src/models/bloom.cpp
@@ -0,0 +1,101 @@
+#include "models.h"
+
+llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    inpL = build_norm(inpL,
+            model.tok_norm,
+            model.tok_norm_b,
+            LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/chameleon.cpp b/llama/llama.cpp/src/models/chameleon.cpp
new file mode 100644
index 00000000..184511ae
--- /dev/null
+++ b/llama/llama.cpp/src/models/chameleon.cpp
@@ -0,0 +1,178 @@
+#include "models.h"
+
+#include <float.h>
+
+llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        if (hparams.swin_norm) {
+            cur = inpL;
+        } else {
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
+                        ggml_element_size(Qcur) * n_embd_head,
+                        ggml_element_size(Qcur) * n_embd_head * n_head,
+                        0);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm,
+                        model.layers[il].attn_q_norm_b,
+                        LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
+                        ggml_element_size(Kcur) * n_embd_head,
+                        ggml_element_size(Kcur) * n_embd_head * n_head_kv,
+                        0);
+                cb(Kcur, "Kcur", il);
+
+                Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm,
+                        model.layers[il].attn_k_norm_b,
+                        LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (!hparams.swin_norm) {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output_with_img_logits", -1);
+
+    // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
+    // Needs to be removed once image outputs are supported.
+    int img_token_end_idx = 8196;
+    int img_token_start_idx = 4;
+    int num_img_tokens = img_token_end_idx - img_token_start_idx;
+    // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
+    // which ensures that text token values are always at least larger than image token values
+    ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
+    img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
+    cb(img_logits, "img_logits", -1);
+
+    cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/chatglm.cpp b/llama/llama.cpp/src/models/chatglm.cpp
new file mode 100644
index 00000000..2685d4fb
--- /dev/null
+++ b/llama/llama.cpp/src/models/chatglm.cpp
@@ -0,0 +1,132 @@
+#include "models.h"
+
+
+llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+            }
+
+            //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+        }
+
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/codeshell.cpp b/llama/llama.cpp/src/models/codeshell.cpp
new file mode 100644
index 00000000..0b3bdbff
--- /dev/null
+++ b/llama/llama.cpp/src/models/codeshell.cpp
@@ -0,0 +1,111 @@
+#include "models.h"
+
+llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/cogvlm.cpp b/llama/llama.cpp/src/models/cogvlm.cpp
new file mode 100644
index 00000000..edf0d142
--- /dev/null
+++ b/llama/llama.cpp/src/models/cogvlm.cpp
@@ -0,0 +1,100 @@
+#include "models.h"
+
+llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    float         kq_scale    = 1.0f / sqrtf(float(n_embd_head));
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor *inpL, *cur;
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    // check ubatch to see if we have input tokens (text)
+    // or an input embedding vector (image)
+    bool is_text;
+    if (ubatch.token) {
+        is_text = true;
+    } else {
+        is_text = false;
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        // get either the text or image weight tensors
+        ggml_tensor *wqkv, *wo;
+        ggml_tensor *ffn_gate, *ffn_down, *ffn_up;
+
+        if (is_text) {
+            wqkv     = model.layers[il].wqkv;
+            wo       = model.layers[il].wo;
+            ffn_gate = model.layers[il].ffn_gate;
+            ffn_down = model.layers[il].ffn_down;
+            ffn_up   = model.layers[il].ffn_up;
+        } else {
+            wqkv     = model.layers[il].visexp_attn_wqkv;
+            wo       = model.layers[il].visexp_attn_wo;
+            ffn_gate = model.layers[il].visexp_ffn_gate;
+            ffn_down = model.layers[il].visexp_ffn_down;
+            ffn_up   = model.layers[il].visexp_ffn_up;
+        }
+
+        ggml_tensor * inpSA = inpL;
+        cur                 = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+
+        // build self attention
+        {
+            ggml_tensor * qkv = build_lora_mm(wqkv, cur);
+
+            // split qkv into Q, K, V along the first dimension
+            ggml_tensor * Qcur =
+                ggml_view_3d(ctx0, qkv, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), qkv->nb[1], 0);
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], n_embd * ggml_element_size(qkv));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], 2 * n_embd * ggml_element_size(qkv));
+
+            Qcur = ggml_rope(ctx0, Qcur, inp_pos, n_embd_head, rope_type);
+            Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type);
+
+            cur = build_attn(inp_attn,
+                wo, nullptr,
+                Qcur, Kcur, Vcur,
+                nullptr, nullptr, nullptr,
+                kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                ffn_up, NULL, NULL,
+                ffn_gate, NULL, NULL,
+                ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/cohere2-iswa.cpp b/llama/llama.cpp/src/models/cohere2-iswa.cpp
new file mode 100644
index 00000000..b18aa8c4
--- /dev/null
+++ b/llama/llama.cpp/src/models/cohere2-iswa.cpp
@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const bool is_swa = hparams.is_swa(il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // rope freq factors for 128k context
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (is_swa) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/command-r.cpp b/llama/llama.cpp/src/models/command-r.cpp
new file mode 100644
index 00000000..4d3b643b
--- /dev/null
+++ b/llama/llama.cpp/src/models/command-r.cpp
@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+
+llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/dbrx.cpp b/llama/llama.cpp/src/models/dbrx.cpp
new file mode 100644
index 00000000..6d2a0ebf
--- /dev/null
+++ b/llama/llama.cpp/src/models/dbrx.cpp
@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+            cb(cur, "wqkv_clamped", il);
+
+            Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].attn_out_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_out_norm", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/deci.cpp b/llama/llama.cpp/src/models/deci.cpp
new file mode 100644
index 00000000..7410a3a4
--- /dev/null
+++ b/llama/llama.cpp/src/models/deci.cpp
@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA     = inpL;
+        const int64_t n_head_kv = hparams.n_head_kv(il);
+        const int64_t n_head    = hparams.n_head(il);
+        const int64_t n_ff      = hparams.n_ff(il);
+
+        if (n_head == 0) {
+            // attention-free layer of Llama-3_1-Nemotron-51B
+            cur = inpL;
+        } else {
+            // norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        if (n_head > 0 && n_head_kv == 0) {
+            // "linear attention" of Llama-3_1-Nemotron-51B
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "wo", il);
+        } else if (n_head > 0) {
+            // self-attention
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
+        if (n_ff == 0) {
+            continue;
+        }
+        // modified to support attention-free layer of Llama-3_1-Nemotron-51B
+        ggml_tensor * ffn_inp = cur;
+        if (n_head > 0) {
+            ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+        }
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/deepseek.cpp b/llama/llama.cpp/src/models/deepseek.cpp
new file mode 100644
index 00000000..17866c0d
--- /dev/null
+++ b/llama/llama.cpp/src/models/deepseek.cpp
@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+
+llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, false,
+                false, hparams.expert_weights_scale,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/deepseek2.cpp b/llama/llama.cpp/src/models/deepseek2.cpp
new file mode 100644
index 00000000..0b41f7ba
--- /dev/null
+++ b/llama/llama.cpp/src/models/deepseek2.cpp
@@ -0,0 +1,237 @@
+#include "models.h"
+
+
+
+llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
+    bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
+
+    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+    const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+    const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+    const float mscale      = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+    const float kq_scale    = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
+    const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * q = NULL;
+            if (!is_lite) {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+                cb(q, "q", il);
+
+                q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+                cb(q, "q", il);
+
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+                cb(q, "q", il);
+            } else {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(q, "q", il);
+            }
+            // split into {n_embd_head_qk_nope, n_head, n_tokens}
+            ggml_tensor * q_nope =
+                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                             ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_embd_head_qk_rope, n_head, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(
+                ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_cmpr =
+                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // and {n_embd_head_qk_rope, 1, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(q_pe, "q_pe", il);
+
+            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(k_pe, "k_pe", il);
+
+            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            if (is_mla) {
+                // {n_embd_head_qk_nope, n_tokens, n_head}
+                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                cb(q_nope, "q_nope_perm", il);
+
+                // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+                // {kv_lora_rank, n_head, n_tokens}
+                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                cb(Qcur, "Qcur", il);
+
+                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                cb(Kcur, "Kcur", il);
+
+                // {kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Vcur = kv_cmpr;
+                cb(Vcur, "Vcur", il);
+
+                // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
+            } else {
+                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
+                cb(kv, "kv", il);
+
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                ggml_tensor * k_nope =
+                    ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0);
+                cb(k_nope, "k_nope_view", il);
+
+                // and {n_embd_head_v, n_head, n_tokens}
+                ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope));
+                cb(Vcur, "Vcur_view", il);
+
+                Vcur = ggml_cont(ctx0, Vcur);
+                cb(Vcur, "Vcur_cont", il);
+
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                cb(Kcur, "Kcur", il);
+
+                // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
+                cur = build_attn(inp_attn,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            }
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/dots1.cpp b/llama/llama.cpp/src/models/dots1.cpp
new file mode 100644
index 00000000..09c36f82
--- /dev/null
+++ b/llama/llama.cpp/src/models/dots1.cpp
@@ -0,0 +1,134 @@
+#include "models.h"
+
+
+
+llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/dream.cpp b/llama/llama.cpp/src/models/dream.cpp
new file mode 100644
index 00000000..2aafbae1
--- /dev/null
+++ b/llama/llama.cpp/src/models/dream.cpp
@@ -0,0 +1,105 @@
+#include "models.h"
+
+
+
+llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    //copied from qwen2
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur               = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur               = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur               = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+            model.layers[il].ffn_up, NULL, NULL,
+            model.layers[il].ffn_gate, NULL, NULL,
+            model.layers[il].ffn_down, NULL, NULL,
+            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/ernie4-5-moe.cpp b/llama/llama.cpp/src/models/ernie4-5-moe.cpp
new file mode 100644
index 00000000..0d96d14e
--- /dev/null
+++ b/llama/llama.cpp/src/models/ernie4-5-moe.cpp
@@ -0,0 +1,150 @@
+#include "models.h"
+
+
+
+llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        bool is_moe_layer =
+            static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;
+
+        if (!is_moe_layer) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                                        model.layers[il].ffn_gate_inp,
+                                        model.layers[il].ffn_up_exps,
+                                        model.layers[il].ffn_gate_exps,
+                                        model.layers[il].ffn_down_exps,
+                                        model.layers[il].ffn_exp_probs_b,
+                                        n_expert, n_expert_used,
+                                        LLM_FFN_SILU, true,
+                                        false, 0.0,
+                                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                                        il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // Shared expert (if present)
+            if (hparams.n_ff_shexp > 0) {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            } else {
+                cur = moe_out;
+            }
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/ernie4-5.cpp b/llama/llama.cpp/src/models/ernie4-5.cpp
new file mode 100644
index 00000000..99aead53
--- /dev/null
+++ b/llama/llama.cpp/src/models/ernie4-5.cpp
@@ -0,0 +1,110 @@
+#include "models.h"
+
+llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1) {
+            // skip computing output for unused tokens
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/exaone.cpp b/llama/llama.cpp/src/models/exaone.cpp
new file mode 100644
index 00000000..62602b28
--- /dev/null
+++ b/llama/llama.cpp/src/models/exaone.cpp
@@ -0,0 +1,114 @@
+#include "models.h"
+
+
+
+llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/exaone4.cpp b/llama/llama.cpp/src/models/exaone4.cpp
new file mode 100644
index 00000000..8b7e3dc0
--- /dev/null
+++ b/llama/llama.cpp/src/models/exaone4.cpp
@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+template <bool iswa>
+llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // use RoPE for SWA layers or non-SWA models
+        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;
+
+        cur = inpL;
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            if (use_rope) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_ffn(ffn_inp,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL, NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_exaone4<false>;
+template struct llm_build_exaone4<true>;
diff --git a/llama/llama.cpp/src/models/falcon-h1.cpp b/llama/llama.cpp/src/models/falcon-h1.cpp
new file mode 100644
index 00000000..b641a094
--- /dev/null
+++ b/llama/llama.cpp/src/models/falcon-h1.cpp
@@ -0,0 +1,113 @@
+#include "models.h"
+
+
+
+llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // Build the inputs in the recurrent & kv cache
+    auto * inp = build_inp_mem_hybrid();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+        cb(Qcur, "Qcur", il);
+
+        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+        cb(Kcur, "Kcur", il);
+
+        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        cb(Qcur, "Qcur-post-rope", il);
+        cb(Kcur, "Kcur-post-rope", il);
+        cb(Vcur, "Vcur-post-rope", il);
+
+        ggml_tensor * attn_out = build_attn(inp->get_attn(),
+                                    model.layers[il].wo, NULL,
+                                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        cb(attn_out, "attn_out", il);
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        // Mamba2 layer
+        cb(cur, "ssm_in", il);
+
+        ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        cb(ssm_out, "ssm_out", il);
+
+        // // Aggregation
+        cur   = ggml_add(ctx0, attn_out, ssm_out);
+        inpSA = ggml_add(ctx0, cur, inpSA);
+        cb(cur, "layer_out", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = inpSA;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, inpSA);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/falcon.cpp b/llama/llama.cpp/src/models/falcon.cpp
new file mode 100644
index 00000000..db1ccdb5
--- /dev/null
+++ b/llama/llama.cpp/src/models/falcon.cpp
@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * attn_norm;
+
+        attn_norm = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(attn_norm, "attn_norm", il);
+
+        // self-attention
+        {
+            if (model.layers[il].attn_norm_2) {
+                // Falcon-40B
+                cur = build_norm(inpL,
+                        model.layers[il].attn_norm_2,
+                        model.layers[il].attn_norm_2_b,
+                        LLM_NORM, il);
+                cb(cur, "attn_norm_2", il);
+            } else {
+                cur = attn_norm;
+            }
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            // using mode = 2 for neox mode
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);
+            inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);
+            attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+
+        // feed forward
+        {
+            cur = build_ffn(attn_norm, // !! use the attn norm, not the result
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = ggml_add(ctx0, cur, inpL);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    cur = build_norm(cur,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/gemma-embedding.cpp b/llama/llama.cpp/src/models/gemma-embedding.cpp
new file mode 100644
index 00000000..90a98f7a
--- /dev/null
+++ b/llama/llama.cpp/src/models/gemma-embedding.cpp
@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+
+llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur =
+                build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/gemma.cpp b/llama/llama.cpp/src/models/gemma.cpp
new file mode 100644
index 00000000..4893d9af
--- /dev/null
+++ b/llama/llama.cpp/src/models/gemma.cpp
@@ -0,0 +1,112 @@
+#include "models.h"
+
+
+llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
+            cb(Qcur, "Qcur_scaled", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/gemma2-iswa.cpp b/llama/llama.cpp/src/models/gemma2-iswa.cpp
new file mode 100644
index 00000000..9cc59a53
--- /dev/null
+++ b/llama/llama.cpp/src/models/gemma2-iswa.cpp
@@ -0,0 +1,125 @@
+#include "models.h"
+
+llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // final logit soft-capping
+    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+    cur = ggml_tanh(ctx0, cur);
+    cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/gemma3-iswa.cpp b/llama/llama.cpp/src/models/gemma3-iswa.cpp
new file mode 100644
index 00000000..839ff6d3
--- /dev/null
+++ b/llama/llama.cpp/src/models/gemma3-iswa.cpp
@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // TODO: is causal == true correct? might need some changes
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/gemma3n-iswa.cpp b/llama/llama.cpp/src/models/gemma3n-iswa.cpp
new file mode 100644
index 00000000..a0bdd6a1
--- /dev/null
+++ b/llama/llama.cpp/src/models/gemma3n-iswa.cpp
@@ -0,0 +1,377 @@
+#include "models.h"
+
+
+
+llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params),
+    model(model),
+    n_embd_head(model.hparams.n_embd_head_k),
+    n_embd_altup(model.hparams.n_embd_altup),
+    n_altup(model.hparams.n_altup),
+    i_altup_act(model.hparams.i_altup_act) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // TODO: is causal == true correct? might need some changes
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
+    ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
+
+    // inpL now has only 1 altup, project it to the rest of the altups
+    // these "added" altups will be concat to the last dim of inpL
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(inpL);
+        ggml_tensor * inp_repeated     = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1);
+        ggml_tensor * altup_added =
+            ggml_mul_mat(ctx0, model.altup_proj, inp_repeated);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_added);
+        altup_added                 = ggml_div(ctx0, ggml_mul(ctx0, altup_added, target_magnitude), new_magnitude);
+        inpL                        = ggml_concat(ctx0, inpL, altup_added, 2);  // shape: [n_embd, n_tokens, n_altup]
+        cb(inpL, "inp_stacked", -1);
+    }
+    // inpL now has shape:          [n_embd,       n_tokens, n_altup]
+    // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer]
+
+    for (int il = 0; il < n_layer; ++il) {
+        // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        ggml_tensor * cur         = inpL;                    // [n_embd, n_tokens, n_altup]
+        ggml_tensor * predictions = altup_predict(cur, il);  // [n_embd, n_tokens, n_altup]
+
+        // predicted value will go through self-attention and laurel
+        ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);  // [n_embd, n_tokens]
+        cur                             = active_prediction;
+        cb(cur, "active_prediction", il);
+
+        // norm
+        cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // laurel
+        ggml_tensor * laurel_out = laurel(cur, il);  // [n_embd, n_tokens]
+
+        // self-attention
+        if (hparams.has_kv(il)) {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
+
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+            cb(Vcur, "Vcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+
+            cur = build_attn(inp_attn, model.layers[il].wo,
+                    NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+                    hparams.f_attention_scale, il);
+        } else {
+            // reuse KV cache of earlier layers
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(Qcur, "Qcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, active_prediction);  // [n_embd, n_tokens]
+        cb(cur, "attn_gated", il);
+
+        ggml_tensor * attn_laurel = ggml_scale(ctx0, ggml_add(ctx0, cur, laurel_out),
+                                               1.0f / sqrtf(2.0f));  // [n_embd, n_tokens]
+        cb(attn_laurel, "attn_laurel", il);
+
+        cur = build_norm(attn_laurel, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            ggml_tensor * up_proj   = build_lora_mm(model.layers[il].ffn_up, cur);
+            ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur);
+
+            if (il < n_layer_sparsity) {
+                // apply activation sparsity
+                gate_proj = gaussian_topk(gate_proj);
+            }
+            gate_proj = ggml_gelu(ctx0, gate_proj);
+
+            cur = ggml_mul(ctx0, up_proj, gate_proj);
+            cur = build_lora_mm(model.layers[il].ffn_down, cur);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", il);
+
+        ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel);  // [n_embd, n_tokens]
+        cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il);
+
+        ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il);  // [n_embd, n_tokens, n_altup]
+
+        ggml_tensor * first_prediction;                                                   // [n_embd, n_tokens]
+        {
+            first_prediction = view_2d_slice(corrected, i_altup_act);                     // [n_embd, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale);
+            first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction);
+            first_prediction = ggml_gelu(ctx0, first_prediction);                 // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_gated", il);
+            ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il);      // [n_embd_altup, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer);  // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_scaled", il);
+
+            first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction);  // [n_embd, n_tokens]
+            first_prediction =
+                build_norm(first_prediction, model.layers[il].per_layer_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(first_prediction, "first_prediction_out", il);
+        }
+        // equivalent to python code: corrected_predictions[1:] += first_prediction
+        {
+            ggml_tensor * slice_first = view_2d_slice(corrected, 0);
+            ggml_tensor * slice_rest  = ggml_view_3d(
+                ctx0, corrected, n_embd, n_tokens, n_altup - 1, ggml_row_size(corrected->type, n_embd),
+                ggml_row_size(corrected->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(corrected));
+            ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction);  // [n_embd, n_tokens, n_altup - 1]
+            corrected         = ggml_concat(ctx0, slice_first, tmp, 2);        // [n_embd, n_tokens, n_altup]
+        }
+        cur = corrected;                                                       // [n_embd, n_tokens, n_altup]
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;  // [n_embd, n_tokens, n_altup]
+
+    // cur now has multiple altup(s), we want to merge them back to 1 altup
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act));  // [n_embd, n_tokens]
+        // do a view to skip the first slice (active altup)
+        ggml_tensor * alt_slice =
+            ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1, ggml_row_size(cur->type, n_embd),
+                         ggml_row_size(cur->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(cur));
+        ggml_tensor * altup_unembd =
+            ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_unembd);
+        altup_unembd                = ggml_div(ctx0, ggml_mul(ctx0, altup_unembd, target_magnitude), new_magnitude);
+        cb(altup_unembd, "altup_unembd", -1);
+
+        // equivalent to torch.mean(hidden_states, dim=0)
+        cur = view_2d_slice(cur, 0);  // [n_embd, n_tokens]
+        for (int i = 0; i < n_altup - 1; ++i) {
+            cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i));
+        }
+        cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup));  // [n_embd, n_tokens]
+        cb(cur, "unembd_merged", -1);
+    }
+    // cur now has shape: [n_embd, n_tokens]
+
+    // TODO: move this to right after the last KV layer
+    {
+        // skip computing output for unused tokens
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+        cur                       = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    {
+        // final logit soft-capping
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_gemma3n_iswa::calc_magnitude(ggml_tensor * x) {
+    return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x)));
+}
+
+// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
+ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
+    GGML_ASSERT(idx < (int) x->ne[2]);
+    return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]),
+                        idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
+}
+
+// equivalent to get_per_layer_inputs() in python code
+// output shape: [n_embd_altup, n_layer, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
+    auto          inp = std::make_unique<llm_graph_input_embd>();
+    ggml_tensor * inp_per_layer;
+    if (ubatch.token) {
+        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+        ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
+        inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
+        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
+        inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
+        cb(inp_per_layer, "inp_per_layer_selected", -1);
+    } else {
+        GGML_ABORT("TODO: support embd input");
+    }
+    res->add_input(std::move(inp));
+    return inp_per_layer;
+}
+
+// equivalent to project_per_layer_inputs() in python code
+// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
+// output shape: [n_embd_altup, n_tokens, n_layer]
+ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) {
+    const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
+    const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
+
+    ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
+    per_layer_proj               = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale);
+    per_layer_proj               = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens);
+    per_layer_proj               = build_norm(per_layer_proj, model.per_layer_proj_norm, NULL, LLM_NORM_RMS,
+                                              -1);  // [n_embd_altup, n_layer, n_tokens]
+    cb(per_layer_proj, "per_layer_proj", -1);
+
+    inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
+    inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
+    cb(inp_per_layer, "inp_per_layer", -1);
+
+    // permute to shape: [n_embd_altup, n_tokens, n_layer]
+    inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
+    return inp_per_layer;
+}
+
+// input cur shape: [n_altup, n_tokens]
+// output    shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::laurel(ggml_tensor * cur, int il) {
+    ggml_tensor * tmp = cur;
+    tmp               = build_lora_mm(model.layers[il].laurel_l, tmp);
+    tmp               = build_lora_mm(model.layers[il].laurel_r, tmp);
+    tmp               = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il);
+    tmp               = ggml_add(ctx0, tmp, cur);
+    cb(tmp, "laurel_out", il);
+    return tmp;
+}
+
+// input x shape: [n_embd, n_tokens]
+// output  shape: [n_embd, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) {
+    ggml_tensor * mean = ggml_mean(ctx0, x);
+    ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
+                                                    1.0f / (float) (x->ne[0] - 1)));
+    ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul));
+    return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x));
+}
+
+//
+// altup functions
+//
+
+// equivalent to compute_router_modalities() in python code
+// input x shape: [n_embd,  n_tokens]
+// output  shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::altup_compute_router_modalities(ggml_tensor * x, int il) {
+    ggml_tensor * router_inputs = build_norm(x, model.layers[il].altup_router_norm, NULL, LLM_NORM_RMS, il);
+
+    // router_input_scale
+    router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float) n_embd);
+
+    ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs);
+    return ggml_tanh(ctx0, output);  // [n_altup, n_tokens]
+}
+
+// input cur shape: [n_embd, n_tokens, n_altup]
+// output    shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_predict(ggml_tensor * cur, int il) {
+    ggml_tensor * activated  = view_2d_slice(cur, i_altup_act);                 // [n_embd, n_tokens]
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities);
+    cb(all_coefs, "all_coefs", il);
+    // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor)
+    all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens);
+
+    // permute to [n_altup, n_embd, n_tokens]
+    ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
+    ggml_tensor * predictions  = ggml_mul_mat(ctx0, cur_permuted, all_coefs);  // [n_altup, n_embd, n_tokens]
+
+    // final shape must be the same as cur: [n_embd, n_tokens, n_altup]
+    predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3));
+    predictions = ggml_add(ctx0, predictions, cur);
+    cb(predictions, "predictions", il);
+
+    return predictions;
+}
+
+// input predictions       shape: [n_embd, n_tokens, n_altup]
+// input activated         shape: [n_embd, n_tokens]
+// output                  shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);
+    ggml_tensor * innovation        = ggml_sub(ctx0, activated, active_prediction);  // [n_embd, n_tokens]
+    cb(innovation, "innovation", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities);  // [n_altup, n_tokens]
+    all_coefs               = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f);                    // + 1.0
+    cb(all_coefs, "all_coefs", il);
+    all_coefs = ggml_transpose(ctx0, all_coefs);                                               // [n_tokens, n_altup]
+    all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup);                           // [1, n_tokens, n_altup]
+
+    innovation              = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
+    ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs);   // [n_embd, n_tokens, n_altup]
+    corrected               = ggml_add(ctx0, corrected, predictions);  // [n_embd, n_tokens, n_altup]
+    cb(corrected, "corrected", il);
+
+    return corrected;
+}
diff --git a/llama/llama.cpp/src/models/glm4-moe.cpp b/llama/llama.cpp/src/models/glm4-moe.cpp
new file mode 100644
index 00000000..33ee7070
--- /dev/null
+++ b/llama/llama.cpp/src/models/glm4-moe.cpp
@@ -0,0 +1,153 @@
+#include "models.h"
+
+llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // Only process up to last layer (skip final NextN layer)
+    // Final layer tensors are loaded but not processed in forward pass
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            }
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            }
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            }
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            // Apply Q/K norm if available (GLM-4.5 355B variant)
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+            }
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+            }
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // Post-attention norm
+        cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            // Dense FFN layer
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // Process routed experts using existing MoE infrastructure
+            ggml_tensor * routed_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    true, hparams.expert_weights_scale,
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(routed_out, "ffn_moe_out", il);
+
+            // Process shared expert on original input
+            ggml_tensor * shared_out = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(shared_out, "ffn_shexp_out", il);
+
+            // Final output: routed_output + shared_output
+            cur = ggml_add(ctx0, routed_out, shared_out);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/glm4.cpp b/llama/llama.cpp/src/models/glm4.cpp
new file mode 100644
index 00000000..f789b282
--- /dev/null
+++ b/llama/llama.cpp/src/models/glm4.cpp
@@ -0,0 +1,127 @@
+#include "models.h"
+
+
+
+llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // Post-attention norm (new!)
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Add the input (residual connection after post-attention norm)
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            // Pre-MLP norm
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // MLP
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            // Post-MLP norm
+            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_mlp_norm", il);
+        }
+        // Add residual connection after post-MLP norm
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+    // Final norm
+    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // Output projection
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/gpt2.cpp b/llama/llama.cpp/src/models/gpt2.cpp
new file mode 100644
index 00000000..60761c8e
--- /dev/null
+++ b/llama/llama.cpp/src/models/gpt2.cpp
@@ -0,0 +1,105 @@
+#include "models.h"
+
+llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * pos;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+    cb(pos, "pos_embd", -1);
+
+    inpL = ggml_add(ctx0, inpL, pos);
+    cb(inpL, "inpL", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/gptneox.cpp b/llama/llama.cpp/src/models/gptneox.cpp
new file mode 100644
index 00000000..2151b14e
--- /dev/null
+++ b/llama/llama.cpp/src/models/gptneox.cpp
@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // ffn
+        if (hparams.use_par_res) {
+            // attention and ffn are computed in parallel
+            // x = x + attn(ln1(x)) + ffn(ln2(x))
+
+            ggml_tensor * attn_out = cur;
+
+            cur = build_norm(inpL,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, inpL);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, attn_out);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        } else {
+            // attention and ffn are computed sequentially
+            // x = x + attn(ln1(x))
+            // x = x + ffn(ln2(x))
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/granite-hybrid.cpp b/llama/llama.cpp/src/models/granite-hybrid.cpp
new file mode 100644
index 00000000..f6ca4c17
--- /dev/null
+++ b/llama/llama.cpp/src/models/granite-hybrid.cpp
@@ -0,0 +1,196 @@
+#include "models.h"
+
+
+llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // Positional embeddings populated if rope enabled
+    ggml_tensor * inp_pos = nullptr;
+    if (hparams.rope_finetuned) {
+        inp_pos = build_inp_pos();
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        if (hparams.is_recurrent(il)) {
+            // ssm layer //
+            cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        } else {
+            // attention layer //
+            cur = build_attention_layer(cur, inp_pos, inp->get_attn(), model, n_embd_head, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // ffn
+        cur = build_layer_ffn(cur, inpSA, model, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // For Granite architectures - scale logits
+    if (hparams.f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_granite_hybrid::build_attention_layer(ggml_tensor *             cur,
+                                                              ggml_tensor *             inp_pos,
+                                                              llm_graph_input_attn_kv * inp_attn,
+                                                              const llama_model &       model,
+                                                              const int64_t             n_embd_head,
+                                                              const int                 il) {
+    // compute Q and K and (optionally) RoPE them
+    ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur, "Qcur", il);
+    if (model.layers[il].bq) {
+        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+        cb(Qcur, "Qcur", il);
+    }
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+    if (model.layers[il].bk) {
+        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+        cb(Kcur, "Kcur", il);
+    }
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+    if (model.layers[il].bv) {
+        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+        cb(Vcur, "Vcur", il);
+    }
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+    const bool use_rope = hparams.rope_finetuned;
+    if (use_rope) {
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+    }
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, model.layers[il].bo,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+    cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llm_build_granite_hybrid::build_layer_ffn(ggml_tensor *       cur,
+                                                        ggml_tensor *       inpSA,
+                                                        const llama_model & model,
+                                                        const int           il) {
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "ffn_inp", il);
+
+    // feed-forward network (non-MoE)
+    if (model.layers[il].ffn_gate_inp == nullptr) {
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+    } else {
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // For Granite MoE Shared
+        if (hparams.n_ff_shexp > 0) {
+            ggml_tensor * ffn_shexp =
+                build_ffn(cur,
+                    model.layers[il].ffn_up_shexp, NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+    }
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "ffn_out", il);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}
diff --git a/llama/llama.cpp/src/models/granite.cpp b/llama/llama.cpp/src/models/granite.cpp
new file mode 100644
index 00000000..18748e9c
--- /dev/null
+++ b/llama/llama.cpp/src/models/granite.cpp
@@ -0,0 +1,211 @@
+#include "models.h"
+
+
+llm_build_granite::llm_build_granite(
+    const llama_model & model,
+    const llm_graph_params & params)
+    : llm_graph_context(params) {
+
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - built only if rope enabled
+    ggml_tensor * inp_pos = nullptr;
+    if (hparams.rope_finetuned) {
+        inp_pos = build_inp_pos();
+    }
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        cur = build_attention_layer(
+            cur, inp_pos, inp_attn,
+            model, n_embd_head, il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // ffn
+        cur = build_layer_ffn(cur, inpSA, model, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // For Granite architectures - scale logits
+    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_granite::build_attention_layer(
+          ggml_tensor             * cur,
+          ggml_tensor             * inp_pos,
+          llm_graph_input_attn_kv * inp_attn,
+    const llama_model             & model,
+    const int64_t                 n_embd_head,
+    const int                     il) {
+
+    // compute Q and K and (optionally) RoPE them
+    ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur, "Qcur", il);
+    if (model.layers[il].bq) {
+        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+        cb(Qcur, "Qcur", il);
+    }
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+    if (model.layers[il].bk) {
+        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+        cb(Kcur, "Kcur", il);
+    }
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+    if (model.layers[il].bv) {
+        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+        cb(Vcur, "Vcur", il);
+    }
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+    const bool use_rope = hparams.rope_finetuned;
+    if (use_rope) {
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+        Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+        Kcur = ggml_rope_ext(
+                ctx0, Kcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+    }
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, model.layers[il].bo,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llm_build_granite::build_layer_ffn(
+          ggml_tensor       * cur,
+          ggml_tensor       * inpSA,
+    const llama_model       & model,
+    const int                 il) {
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "ffn_inp", il);
+
+    // feed-forward network (non-MoE)
+    if (model.layers[il].ffn_gate_inp == nullptr) {
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+
+    } else {
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // For Granite MoE Shared
+        if (hparams.n_ff_shexp > 0) {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                model.layers[il].ffn_up_shexp,   NULL, NULL,
+                model.layers[il].ffn_gate_shexp, NULL, NULL,
+                model.layers[il].ffn_down_shexp, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+    }
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "ffn_out", il);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}
diff --git a/llama/llama.cpp/src/models/graph-context-mamba.cpp b/llama/llama.cpp/src/models/graph-context-mamba.cpp
new file mode 100644
index 00000000..b9a363b3
--- /dev/null
+++ b/llama/llama.cpp/src/models/graph-context-mamba.cpp
@@ -0,0 +1,283 @@
+#include "models.h"
+
+llm_graph_context_mamba::llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {}
+
+ggml_tensor * llm_graph_context_mamba::build_mamba_layer(llm_graph_input_rs * inp,
+                                                         ggml_tensor *        cur,
+                                                         const llama_model &  model,
+                                                         const llama_ubatch & ubatch,
+                                                         int                  il) {
+    const auto * mctx_cur = inp->mctx;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const auto & layer = model.layers[il];
+
+    const int64_t d_conv         = hparams.ssm_d_conv;
+    const int64_t d_inner        = hparams.ssm_d_inner;
+    const int64_t d_state        = hparams.ssm_d_state;
+    const int64_t dt_rank        = hparams.ssm_dt_rank;
+    const int64_t n_head         = d_inner;
+    const int64_t head_dim       = 1;
+    const int64_t n_seqs         = ubatch.n_seqs;
+    // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers)
+    const bool    ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms;
+
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+    ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
+    conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs);
+
+    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur);
+    // split the above in two
+    // => {d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * x  = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0);
+    ggml_tensor * z =
+        ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner * ggml_element_size(xz));
+
+    // conv
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
+
+        // copy last (d_conv - 1) columns back into the state cache
+        ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2],
+                                               n_seq_tokens * (conv_x->nb[0]));
+
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, last_conv,
+                         ggml_view_1d(ctx0, conv_states_all, (d_conv - 1) * (d_inner) * (n_seqs),
+                                      kv_head * (d_conv - 1) * (d_inner) *ggml_element_size(conv_states_all))));
+
+        // 1D convolution
+        // The equivalent is to make a self-overlapping view of conv_x
+        // over d_conv columns at each stride in the 3rd dimension,
+        // then element-wise multiply that with the conv1d weight,
+        // then sum the elements of each row,
+        // (the last two steps are a dot product over rows (also doable with mul_mat))
+        // then permute away the ne[0] dimension,
+        // and then you're left with the resulting x tensor.
+        // For simultaneous sequences, all sequences need to have the same length.
+        x = ggml_ssm_conv(ctx0, conv_x, layer.ssm_conv1d);
+
+        // bias
+        x = ggml_add(ctx0, x, layer.ssm_conv1d_b);
+
+        x = ggml_silu(ctx0, x);
+    }
+
+    // ssm
+    {
+        // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}
+        ggml_tensor * x_db = build_lora_mm(layer.ssm_x, x);
+        // split
+        ggml_tensor * dt   = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0);
+        ggml_tensor * B =
+            ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1],
+                         x_db->nb[2], ggml_element_size(x_db) * dt_rank);
+        ggml_tensor * C =
+            ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1],
+                         x_db->nb[2], ggml_element_size(x_db) * (dt_rank + d_state));
+
+        // Some Mamba variants (e.g. FalconMamba, Jamba) apply RMS norm in B, C & Dt layers
+        if (ssm_dt_b_c_rms || (layer.ssm_dt_norm && layer.ssm_b_norm && layer.ssm_c_norm)) {
+            dt = build_norm(dt, layer.ssm_dt_norm, NULL, LLM_NORM_RMS, il);
+            B  = build_norm(B, layer.ssm_b_norm, NULL, LLM_NORM_RMS, il);
+            C  = build_norm(C, layer.ssm_c_norm, NULL, LLM_NORM_RMS, il);
+        }
+
+        // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
+        dt = build_lora_mm(layer.ssm_dt, dt);
+        dt = ggml_add(ctx0, dt, layer.ssm_dt_b);
+
+        cur = x;
+        x   = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs);
+
+        ggml_tensor * A = layer.ssm_a;
+
+        // use the states and the indices provided by build_recurrent_state
+        // (this is necessary in order to properly use the states before they are overwritten,
+        //  while avoiding to make unnecessary copies of the states)
+        auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+            ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
+            // Custom operator to optimize the parallel associative scan
+            // as described in the Annex D of the Mamba paper.
+            // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+            return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+        };
+
+        ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, x->nb[3] * x->ne[3]),
+                         ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs,
+                                      kv_head * d_state * d_inner * ggml_element_size(ssm_states_all))));
+
+        ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0);
+
+        // TODO: skip computing output earlier for unused tokens
+
+        y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, layer.ssm_d));
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+
+        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        cur = build_lora_mm(layer.ssm_out, y);
+    }
+
+    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
+
+    return cur;
+}
+
+ggml_tensor * llm_graph_context_mamba::build_mamba2_layer(llm_graph_input_rs * inp,
+                                                          ggml_tensor *        cur,
+                                                          const llama_model &  model,
+                                                          const llama_ubatch & ubatch,
+                                                          int                  il) const {
+    const auto * mctx_cur = inp->mctx;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const int64_t d_conv   = hparams.ssm_d_conv;
+    const int64_t d_inner  = hparams.ssm_d_inner;
+    const int64_t d_state  = hparams.ssm_d_state;
+    const int64_t n_head   = hparams.ssm_dt_rank;
+    const int64_t head_dim = d_inner / n_head;
+    const int64_t n_group  = hparams.ssm_n_group;
+    const int64_t n_seqs   = ubatch.n_seqs;
+
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+    ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
+    conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs);
+
+    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
+
+    // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
+    ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur);
+
+    // split the above in three
+    ggml_tensor * z   = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * zxBCdt->nb[0],
+                                     zxBCdt->nb[1], zxBCdt->nb[2], 0);
+    ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2 * n_group * d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1],
+                                     zxBCdt->nb[2], d_inner * ggml_element_size(zxBCdt));
+    ggml_tensor * dt  = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2],
+                                     (2 * d_inner + 2 * n_group * d_state) * ggml_element_size(zxBCdt));
+
+    // conv
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0);
+
+        // copy last (d_conv - 1) columns back into the state cache
+        ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs,
+                                               conv_x->nb[1], conv_x->nb[2], n_seq_tokens * (conv_x->nb[0]));
+
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv,
+                                               ggml_view_1d(ctx0, conv_states_all,
+                                                            (d_conv - 1) * (d_inner + 2 * n_group * d_state) * (n_seqs),
+                                                            kv_head * (d_conv - 1) * (d_inner + 2 * n_group * d_state) *
+                                                                ggml_element_size(conv_states_all))));
+
+        // 1D convolution
+        // The equivalent is to make a self-overlapping view of conv_x
+        // over d_conv columns at each stride in the 3rd dimension,
+        // then element-wise multiply that with the conv1d weight,
+        // then sum the elements of each row,
+        // (the last two steps are a dot product over rows (also doable with mul_mat))
+        // then permute away the ne[0] dimension,
+        // and then you're left with the resulting x tensor.
+        // For simultaneous sequences, all sequences need to have the same length.
+        xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
+
+        // bias
+        xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b);
+
+        xBC = ggml_silu(ctx0, xBC);
+    }
+
+    // ssm
+    {
+        // These correspond to V K Q in SSM/attention duality
+        ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], 0);
+        ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], d_inner * ggml_element_size(xBC));
+        ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], (d_inner + n_group * d_state) * ggml_element_size(xBC));
+
+        // {n_head, n_seq_tokens, n_seqs}
+        dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b);
+
+        ggml_tensor * A = model.layers[il].ssm_a;
+
+        // use the states and the indices provided by build_recurrent_state
+        // (this is necessary in order to properly use the states before they are overwritten,
+        //  while avoiding to make unnecessary copies of the states)
+        auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+            ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
+            // TODO: use semistructured matrices to implement state-space duality
+            // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+            return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+        };
+
+        ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, ggml_nelements(x) * x->nb[0]),
+                         ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs,
+                                      kv_head * d_state * d_inner * ggml_element_size(ssm_states_all))));
+
+        ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head * x->nb[1],
+                                       n_seq_tokens * n_head * x->nb[1], 0);
+
+        // TODO: skip computing output earlier for unused tokens
+
+        y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
+        cb(y, "mamba2_y_add_d", il);
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+
+        // grouped RMS norm
+        if (model.layers[il].ssm_norm) {
+            y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs);
+            y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
+        }
+
+        y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
+
+        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        cur = build_lora_mm(model.layers[il].ssm_out, y);
+    }
+
+    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
+    cb(cur, "mamba_out", il);
+
+    return cur;
+}
diff --git a/llama/llama.cpp/src/models/grok.cpp b/llama/llama.cpp/src/models/grok.cpp
new file mode 100644
index 00000000..3c54dfee
--- /dev/null
+++ b/llama/llama.cpp/src/models/grok.cpp
@@ -0,0 +1,159 @@
+#include "models.h"
+
+llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_out_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_out_norm", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // MoE branch
+        ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_GELU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        if (model.layers[il].ffn_up) {
+            ggml_tensor * ffn_out = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(ffn_out, "ffn_out", il);
+
+            cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
+
+    // final logit soft-capping
+    if (hparams.f_final_logit_softcapping) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/grovemoe.cpp b/llama/llama.cpp/src/models/grovemoe.cpp
new file mode 100644
index 00000000..56b6db9a
--- /dev/null
+++ b/llama/llama.cpp/src/models/grovemoe.cpp
@@ -0,0 +1,141 @@
+#include "models.h"
+
+
+
+llm_build_grovemoe::llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head    = hparams.n_embd_head_v;
+    const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, cur);  // [n_expert, n_tokens]
+        cb(probs, "ffn_moe_logits", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                nullptr,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il,
+                probs);
+        cb(moe_out, "ffn_moe_out", il);
+        cur = moe_out;
+
+        // TODO: Only do the expert selection and weights once
+        moe_out = build_moe_ffn(cur,
+                    nullptr,
+                    model.layers[il].ffn_up_chexps,
+                    model.layers[il].ffn_gate_chexps,
+                    model.layers[il].ffn_down_chexps,
+                    nullptr,
+                    n_chunk_expert, n_expert_used > n_chunk_expert ? n_chunk_expert : n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il,
+                    probs);
+        cb(moe_out, "ffn_adj_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ggml_scale(ctx0, moe_out, hparams.expert_group_scale));
+        cb(cur, "ffn_final_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/hunyuan-dense.cpp b/llama/llama.cpp/src/models/hunyuan-dense.cpp
new file mode 100644
index 00000000..7d5dcc78
--- /dev/null
+++ b/llama/llama.cpp/src/models/hunyuan-dense.cpp
@@ -0,0 +1,132 @@
+#include "models.h"
+
+llm_build_hunyuan_dense::llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+            Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm, nullptr,
+                        LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_norm", il);
+
+            Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm, nullptr,
+                        LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_norm", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+        // feed-forward network (non-MoE)
+        ggml_tensor * cur_mlp = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur_mlp, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur_mlp, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/hunyuan-moe.cpp b/llama/llama.cpp/src/models/hunyuan-moe.cpp
new file mode 100644
index 00000000..77e39de5
--- /dev/null
+++ b/llama/llama.cpp/src/models/hunyuan-moe.cpp
@@ -0,0 +1,154 @@
+#include "models.h"
+
+llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = build_norm(Kcur,
+                    model.layers[il].attn_k_norm, nullptr,
+                    LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_norm", il);
+
+            Qcur = build_norm(Qcur,
+                    model.layers[il].attn_q_norm, nullptr,
+                    LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_norm", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+            model.layers[il].ffn_norm, NULL,
+            LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network (non-MoE)
+        ggml_tensor * cur_mlp = build_ffn(cur,
+                model.layers[il].ffn_up_shexp,   NULL, NULL,
+                model.layers[il].ffn_gate_shexp, NULL, NULL,
+                model.layers[il].ffn_down_shexp, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur_mlp, "ffn_mlp", il);
+
+        // MoE branch
+        ggml_tensor * cur_moe = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU,
+                true, // norm_topk_prob
+                false,
+                0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur_moe, "ffn_moe_out", il);
+
+        ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp);
+        cb(ffn_out, "ffn_out", il);
+
+        cur = ggml_add(ctx0, ffn_out, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/internlm2.cpp b/llama/llama.cpp/src/models/internlm2.cpp
new file mode 100644
index 00000000..387e8211
--- /dev/null
+++ b/llama/llama.cpp/src/models/internlm2.cpp
@@ -0,0 +1,120 @@
+#include "models.h"
+
+llm_build_internlm2::llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/jais.cpp b/llama/llama.cpp/src/models/jais.cpp
new file mode 100644
index 00000000..3e3376e6
--- /dev/null
+++ b/llama/llama.cpp/src/models/jais.cpp
@@ -0,0 +1,86 @@
+#include "models.h"
+
+llm_build_jais::llm_build_jais(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*cur->nb[0]*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/jamba.cpp b/llama/llama.cpp/src/models/jamba.cpp
new file mode 100644
index 00000000..a0187772
--- /dev/null
+++ b/llama/llama.cpp/src/models/jamba.cpp
@@ -0,0 +1,106 @@
+#include "models.h"
+
+llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_hybrid = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const int64_t n_head_kv = hparams.n_head_kv(il);
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        if (n_head_kv == 0) {
+            cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);
+        } else {
+            // Attention
+
+            struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // No RoPE :)
+            cur = build_attn(inp_hybrid->get_attn(),
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        // residual
+        struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur);
+        cb(cur, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            // FFN
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, false,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+        // residual
+        cur = ggml_add(ctx0, ffn_inp, cur);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    // final rmsnorm
+    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/lfm2.cpp b/llama/llama.cpp/src/models/lfm2.cpp
new file mode 100644
index 00000000..7f805d78
--- /dev/null
+++ b/llama/llama.cpp/src/models/lfm2.cpp
@@ -0,0 +1,175 @@
+#include "models.h"
+
+#include "../llama-memory-hybrid.h"
+
+
+llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params),
+    model(model) {
+    ggml_tensor * cur = build_inp_embd(model.tok_embd);
+    cb(cur, "model.embed_tokens", -1);
+
+    ggml_build_forward_expand(gf, cur);
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    auto *        inp_hybrid  = build_inp_mem_hybrid();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);
+
+        auto * prev_cur = cur;
+        cur             = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "model.layers.{}.operator_norm", il);
+
+        cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) :
+                                         build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
+            prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
+        }
+
+        cur = ggml_add(ctx0, prev_cur, cur);
+
+        auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
+
+        ggml_tensor * ffn_out =
+            is_moe_layer ? build_moe_feed_forward(ffn_norm_out, il) : build_dense_feed_forward(ffn_norm_out, il);
+        cb(ffn_norm_out, "model.layers.{}.ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_out);
+    }
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_lfm2::build_moe_feed_forward(ggml_tensor * cur, int il) const {
+    return build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
+                        model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false, 0.0,
+                        static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func), il);
+}
+
+ggml_tensor * llm_build_lfm2::build_dense_feed_forward(ggml_tensor * cur, int il) const {
+    GGML_ASSERT(!model.layers[il].ffn_up_b);
+    GGML_ASSERT(!model.layers[il].ffn_gate_b);
+    GGML_ASSERT(!model.layers[il].ffn_down_b);
+    return build_ffn(cur,
+        model.layers[il].ffn_up, NULL, NULL,
+        model.layers[il].ffn_gate, NULL, NULL,
+        model.layers[il].ffn_down, NULL, NULL,
+        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+}
+
+ggml_tensor * llm_build_lfm2::build_attn_block(ggml_tensor *             cur,
+                                               ggml_tensor *             inp_pos,
+                                               llm_graph_input_attn_kv * inp_attn,
+                                               int                       il) const {
+    GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
+    const auto n_embd_head = hparams.n_embd_head_v;
+    const auto n_head_kv   = hparams.n_head_kv(il);
+
+    auto * q = build_lora_mm(model.layers[il].wq, cur);
+    cb(q, "model.layers.{}.self_attn.q_proj", il);
+    auto * k = build_lora_mm(model.layers[il].wk, cur);
+    cb(k, "model.layers.{}.self_attn.k_proj", il);
+    auto * v = build_lora_mm(model.layers[il].wv, cur);
+    cb(v, "model.layers.{}.self_attn.v_proj", il);
+
+    q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens);
+    k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
+    v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
+
+    // qk norm
+    q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+    cb(q, "model.layers.{}.self_attn.q_layernorm", il);
+    k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+    cb(k, "model.layers.{}.self_attn.k_layernorm", il);
+
+    // RoPE
+    q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
+                      attn_factor, beta_fast, beta_slow);
+    k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
+                      attn_factor, beta_fast, beta_slow);
+
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, NULL,
+            q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+
+    cb(cur, "model.layers.{}.self_attn.out_proj", il);
+
+    return cur;
+}
+
+ggml_tensor * llm_build_lfm2::build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il) {
+    const auto *   mctx_cur     = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
+    const uint32_t kv_head      = mctx_cur->get_head();
+    const int64_t  n_seq_tokens = ubatch.n_seq_tokens;
+    const int64_t  n_seqs       = ubatch.n_seqs;
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
+    const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
+
+    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
+    cb(bcx, "model.layers.{}.conv.in_proj", il);
+
+    constexpr auto n_chunks = 3;
+    GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
+    const auto chunk_size = bcx->ne[0] / n_chunks;
+    auto *     b          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
+                                         0 * chunk_size * ggml_element_size(bcx));
+    auto *     c          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
+                                         1 * chunk_size * ggml_element_size(bcx));
+    auto *     x          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
+                                         2 * chunk_size * ggml_element_size(bcx));
+
+    auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
+
+    // read conv state
+    auto * conv_state = mctx_cur->get_r_l(il);
+    auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
+    auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
+
+    bx = ggml_concat(ctx0, conv, bx, 0);
+    GGML_ASSERT(bx->ne[0] > conv->ne[0]);
+
+    // last d_conv columns is a new conv state
+    auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
+                                   (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
+    GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
+
+    // write new conv conv state
+    ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
+                                           ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
+                                                        kv_head * d_conv * n_embd * ggml_element_size(new_conv))));
+
+    auto * conv_kernel = model.layers[il].shortconv.conv;
+    auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);
+    cb(conv_out, "model.layers.{}.conv.conv", il);
+
+    auto * y = ggml_mul(ctx0, c, conv_out);
+    y        = build_lora_mm(model.layers[il].shortconv.out_proj, y);
+    cb(y, "model.layers.{}.conv.out_proj", il);
+    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+    y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);
+
+    return y;
+}
diff --git a/llama/llama.cpp/src/models/llada-moe.cpp b/llama/llama.cpp/src/models/llada-moe.cpp
new file mode 100644
index 00000000..5f64686f
--- /dev/null
+++ b/llama/llama.cpp/src/models/llada-moe.cpp
@@ -0,0 +1,122 @@
+#include "models.h"
+
+llm_build_llada_moe::llm_build_llada_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, false,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/llada.cpp b/llama/llama.cpp/src/models/llada.cpp
new file mode 100644
index 00000000..85703366
--- /dev/null
+++ b/llama/llama.cpp/src/models/llada.cpp
@@ -0,0 +1,99 @@
+#include "models.h"
+
+llm_build_llada::llm_build_llada(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    // LLaDA is similar to LLaMA but uses non-causal attention for diffusion
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // Non-causal attention for diffusion
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute separate Q, K, V projections without bias, matching LLaDALlamaBlock
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/llama-iswa.cpp b/llama/llama.cpp/src/models/llama-iswa.cpp
new file mode 100644
index 00000000..03f80616
--- /dev/null
+++ b/llama/llama.cpp/src/models/llama-iswa.cpp
@@ -0,0 +1,174 @@
+#include "models.h"
+
+llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // temperature tuning
+    ggml_tensor * inp_attn_scale = nullptr;
+    inp_attn_scale = build_inp_attn_scale();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
+                              (il + 1) % hparams.n_no_rope_layer_step != 0;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (use_rope) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+            } else if (inp_attn_scale) {
+                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            if (use_rope && hparams.use_kq_norm) {
+                // Llama4TextL2Norm
+                Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
+                Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
+                cb(Qcur, "Qcur_normed", il);
+                cb(Kcur, "Kcur_normed", il);
+            }
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, false,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
+                    il);
+
+            // Shared experts
+            ggml_tensor * shexp_out = build_ffn(ffn_inp_normed,
+                model.layers[il].ffn_up_shexp,   NULL, NULL,
+                model.layers[il].ffn_gate_shexp, NULL, NULL,
+                model.layers[il].ffn_down_shexp, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(shexp_out, "ffn_moe_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, shexp_out);
+            cb(cur, "ffn_moe_out_merged", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/llama.cpp b/llama/llama.cpp/src/models/llama.cpp
new file mode 100644
index 00000000..ab7fd5d0
--- /dev/null
+++ b/llama/llama.cpp/src/models/llama.cpp
@@ -0,0 +1,155 @@
+#include "models.h"
+
+llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            if (hparams.use_kq_norm) {
+                // Llama4TextL2Norm
+                Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
+                Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
+                cb(Qcur, "Qcur_normed", il);
+                cb(Kcur, "Kcur_normed", il);
+            }
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/mamba.cpp b/llama/llama.cpp/src/models/mamba.cpp
new file mode 100644
index 00000000..46819613
--- /dev/null
+++ b/llama/llama.cpp/src/models/mamba.cpp
@@ -0,0 +1,55 @@
+#include "models.h"
+
+
+llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * rs_inp = build_rs_inp();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        if (model.arch == LLM_ARCH_MAMBA2) {
+            cur = build_mamba2_layer(rs_inp, cur, model, ubatch, il);
+        } else {
+            cur = build_mamba_layer(rs_inp, cur, model, ubatch, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    // final rmsnorm
+    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
diff --git a/llama/llama.cpp/src/models/minicpm3.cpp b/llama/llama.cpp/src/models/minicpm3.cpp
new file mode 100644
index 00000000..f374a9fd
--- /dev/null
+++ b/llama/llama.cpp/src/models/minicpm3.cpp
@@ -0,0 +1,199 @@
+#include "models.h"
+
+llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    //TODO: if the model varies, these parameters need to be read from the model
+    const int64_t n_embd_base = 256;
+    const float scale_embd  = 12.0f;
+    const float scale_depth = 1.4f;
+    const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k));
+
+    const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // scale the input embeddings
+    inpL = ggml_scale(ctx0, inpL, scale_embd);
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * q = NULL;
+            // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
+            q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+            cb(q, "q", il);
+
+            q = build_norm(q,
+                    model.layers[il].attn_q_a_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(q, "q", il);
+
+            // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
+            q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+            cb(q, "q", il);
+
+            // split into {n_head * n_embd_head_qk_nope, n_tokens}
+            ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                    ggml_row_size(q->type, hparams.n_embd_head_k),
+                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_head * n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                    ggml_row_size(q->type, hparams.n_embd_head_k),
+                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
+                    kv_pe_compresseed->nb[1],
+                    0);
+            cb(kv_compressed, "kv_compressed", il);
+
+            // and {n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
+                    kv_pe_compresseed->nb[1],
+                    kv_pe_compresseed->nb[1],
+                    ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            kv_compressed = build_norm(kv_compressed,
+                    model.layers[il].attn_kv_a_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(kv_compressed, "kv_compressed", il);
+
+            // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
+            ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
+            cb(kv, "kv", il);
+
+            // split into {n_head * n_embd_head_qk_nope, n_tokens}
+            ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
+                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    0);
+            cb(k_nope, "k_nope", il);
+
+            // and {n_head * n_embd_head_v, n_tokens}
+            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope)));
+            cb(v_states, "v_states", il);
+
+            v_states = ggml_cont(ctx0, v_states);
+            cb(v_states, "v_states", il);
+
+            q_pe = ggml_rope_ext(
+                    ctx0, q_pe, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+            cb(q_pe, "q_pe", il);
+
+            // shared RoPE key
+            k_pe = ggml_rope_ext(
+                    ctx0, k_pe, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+            cb(k_pe, "k_pe", il);
+
+            ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+            cb(q_states, "q_states", il);
+
+            ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+            cb(k_states, "k_states", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // scale_res - scale the hidden states for residual connection
+        const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct?
+        cur = ggml_scale(ctx0, cur, scale_res);
+        cb(cur, "hidden_scaled", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        // scale the hidden states for residual connection
+        cur = ggml_scale(ctx0, cur, scale_res);
+        cb(cur, "hidden_scaled_ffn", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head scaling
+    const float scale_lmhead = float(n_embd_base)/float(n_embd);
+    cur = ggml_scale(ctx0, cur, scale_lmhead);
+    cb(cur, "lmhead_scaling", -1);
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/minimax-m2.cpp b/llama/llama.cpp/src/models/minimax-m2.cpp
new file mode 100644
index 00000000..f7001bad
--- /dev/null
+++ b/llama/llama.cpp/src/models/minimax-m2.cpp
@@ -0,0 +1,124 @@
+
+#include "models.h"
+
+llm_build_minimax_m2::llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    // GGML_ASSERT(n_embd_head == hparams.n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto inp_attn = build_attn_inp_kv();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = inpL;
+
+        // self_attention
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+            Kcur = ggml_rope_ext(
+                ctx0, Kcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/models.go b/llama/llama.cpp/src/models/models.go
new file mode 100644
index 00000000..c6c42b6f
--- /dev/null
+++ b/llama/llama.cpp/src/models/models.go
@@ -0,0 +1,6 @@
+package models
+
+// #cgo CXXFLAGS: -std=c++17
+// #cgo CPPFLAGS: -I${SRCDIR}/../../include -I${SRCDIR}/../../vendor
+// #cgo CPPFLAGS: -I${SRCDIR}/../../../../ml/backend/ggml/ggml/include
+import "C"
diff --git a/llama/llama.cpp/src/models/models.h b/llama/llama.cpp/src/models/models.h
new file mode 100644
index 00000000..71fea796
--- /dev/null
+++ b/llama/llama.cpp/src/models/models.h
@@ -0,0 +1,544 @@
+#pragma once
+
+#include "../llama-model.h"
+#include "../llama-graph.h"
+
+// TODO: remove in follow-up PR - move to .cpp files
+#include "../llama-memory-recurrent.h"
+#include <cmath>
+
+struct llm_graph_context_mamba : public llm_graph_context {
+    llm_graph_context_mamba(const llm_graph_params & params);
+
+    virtual ~llm_graph_context_mamba() = default;
+
+    ggml_tensor * build_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il);
+    ggml_tensor * build_mamba2_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il) const;
+
+};
+
+// Base class for RWKV-related models
+struct llm_build_rwkv6_base : public llm_graph_context {
+    const llama_model & model;
+
+    llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params);
+
+    virtual ~llm_build_rwkv6_base() = default;
+
+    ggml_tensor * build_rwkv6_channel_mix(const llama_layer * layer,
+                                          ggml_tensor *       cur,
+                                          ggml_tensor *       x_prev,
+                                          llm_arch            arch) const;
+
+    ggml_tensor * build_rwkv6_time_mix(llm_graph_input_rs * inp,
+                                       ggml_tensor *        cur,
+                                       ggml_tensor *        x_prev,
+                                       const llama_ubatch & ubatch,
+                                       int                  il) const;
+};
+
+// Base class for RWKV7-related models
+struct llm_build_rwkv7_base : public llm_graph_context {
+    const llama_model & model;
+
+    llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params);
+
+    virtual ~llm_build_rwkv7_base() = default;
+
+    // RWKV7-specific graph building methods
+    ggml_tensor * build_rwkv7_channel_mix(const llama_layer * layer,
+                                          ggml_tensor *       cur,
+                                          ggml_tensor *       x_prev,
+                                          llm_arch            arch) const;
+    ggml_tensor * build_rwkv7_time_mix(llm_graph_input_rs * inp,
+                                       ggml_tensor *        cur,
+                                       ggml_tensor *        x_prev,
+                                       ggml_tensor *&       first_layer_value,
+                                       const llama_ubatch & ubatch,
+                                       int                  il) const;
+};
+
+struct llm_build_afmoe : public llm_graph_context {
+    llm_build_afmoe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_apertus : public llm_graph_context {
+    llm_build_apertus(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_arcee : public llm_graph_context {
+    llm_build_arcee(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_arctic : public llm_graph_context {
+    llm_build_arctic(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_arwkv7 : public llm_build_rwkv7_base {
+    llm_build_arwkv7(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_baichuan : public llm_graph_context {
+    llm_build_baichuan(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_bailingmoe2 : public llm_graph_context {
+    llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_bailingmoe : public llm_graph_context {
+    llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_bert : public llm_graph_context {
+    llm_build_bert(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_bitnet : public llm_graph_context {
+    llm_build_bitnet(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_bloom : public llm_graph_context {
+    llm_build_bloom(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_chameleon : public llm_graph_context {
+    llm_build_chameleon(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_chatglm : public llm_graph_context {
+    llm_build_chatglm(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_codeshell : public llm_graph_context {
+    llm_build_codeshell(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_cogvlm : public llm_graph_context {
+    llm_build_cogvlm(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_cohere2_iswa : public llm_graph_context {
+    llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_command_r : public llm_graph_context {
+    llm_build_command_r(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_dbrx : public llm_graph_context {
+    llm_build_dbrx(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_deci : public llm_graph_context {
+    llm_build_deci(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_deepseek2 : public llm_graph_context {
+    llm_build_deepseek2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_deepseek : public llm_graph_context {
+    llm_build_deepseek(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_dots1 : public llm_graph_context {
+    llm_build_dots1(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_dream : public llm_graph_context {
+    llm_build_dream(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_ernie4_5 : public llm_graph_context {
+    llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_ernie4_5_moe : public llm_graph_context {
+    llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params);
+};
+
+template <bool iswa>
+struct llm_build_exaone4 : public llm_graph_context {
+    llm_build_exaone4(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_exaone : public llm_graph_context {
+    llm_build_exaone(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_falcon : public llm_graph_context {
+    llm_build_falcon(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_falcon_h1 : public llm_graph_context_mamba {
+    llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_gemma2_iswa : public llm_graph_context {
+    llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_gemma3_iswa : public llm_graph_context {
+    llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_gemma3n_iswa : public llm_graph_context {
+    const llama_model & model;
+
+    const int64_t n_embd_head;
+    const int64_t n_embd_altup;
+    const int64_t n_altup;
+    const int     i_altup_act;
+    const int     n_layer_sparsity = 10; // number of layers using activation sparsity
+    const float   f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95)
+
+    llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params);
+    ggml_tensor * calc_magnitude(ggml_tensor * x);
+    ggml_tensor * view_2d_slice(ggml_tensor * x, int idx);
+    ggml_tensor * get_per_layer_inputs();
+    ggml_tensor * project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer);
+    ggml_tensor * gaussian_topk(ggml_tensor * x);
+    ggml_tensor * altup_compute_router_modalities(ggml_tensor * x, int il);
+    ggml_tensor * altup_predict(ggml_tensor * cur, int il);
+    ggml_tensor * laurel(ggml_tensor * cur, int il);
+    ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il);
+};
+
+struct llm_build_gemma_embedding : public llm_graph_context {
+    llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_gemma : public llm_graph_context {
+    llm_build_gemma(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_glm4 : public llm_graph_context {
+    llm_build_glm4(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_glm4_moe : public llm_graph_context {
+    llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_gpt2 : public llm_graph_context {
+    llm_build_gpt2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_gptneox : public llm_graph_context {
+    llm_build_gptneox(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_granite : public llm_graph_context {
+    llm_build_granite(const llama_model & model, const llm_graph_params & params);
+
+private:
+    ggml_tensor * build_attention_layer(
+              ggml_tensor             * cur,
+              ggml_tensor             * inp_pos,
+              llm_graph_input_attn_kv * inp_attn,
+        const llama_model             & model,
+        const int64_t                 n_embd_head,
+        const int                     il);
+
+    ggml_tensor * build_layer_ffn(
+              ggml_tensor       * cur,
+              ggml_tensor       * inpSA,
+        const llama_model       & model,
+        const int                 il);
+};
+
+struct llm_build_granite_hybrid : public llm_graph_context_mamba {
+    llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params);
+    ggml_tensor * build_layer_ffn(ggml_tensor * cur, ggml_tensor * inpSA, const llama_model & model, const int il);
+    ggml_tensor * build_attention_layer(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn,
+        const llama_model & model,const int64_t n_embd_head, const int il);
+};
+
+struct llm_build_grok : public llm_graph_context {
+    llm_build_grok(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_grovemoe : public llm_graph_context {
+    llm_build_grovemoe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_hunyuan_dense : public llm_graph_context {
+    llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_hunyuan_moe : public llm_graph_context {
+    llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_internlm2 : public llm_graph_context {
+    llm_build_internlm2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_jais : public llm_graph_context {
+    llm_build_jais(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_jamba : public llm_graph_context_mamba {
+    llm_build_jamba(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_lfm2 : public llm_graph_context {
+    const llama_model & model;
+
+    llm_build_lfm2(const llama_model & model, const llm_graph_params & params);
+    ggml_tensor * build_moe_feed_forward(ggml_tensor * cur, int il) const;
+    ggml_tensor * build_dense_feed_forward(ggml_tensor * cur, int il) const;
+    ggml_tensor * build_attn_block(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, int il) const;
+    ggml_tensor * build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il);
+
+};
+
+struct llm_build_llada : public llm_graph_context {
+    llm_build_llada(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_llada_moe : public llm_graph_context {
+    llm_build_llada_moe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_llama : public llm_graph_context {
+    llm_build_llama(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_llama_iswa : public llm_graph_context {
+    llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_mamba : public llm_graph_context_mamba {
+    llm_build_mamba(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_minicpm3 : public llm_graph_context {
+    llm_build_minicpm3(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_minimax_m2 : public llm_graph_context {
+    llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_mpt : public llm_graph_context {
+    llm_build_mpt(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_nemotron : public llm_graph_context {
+    llm_build_nemotron(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_nemotron_h : public llm_graph_context_mamba {
+    llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params);
+    ggml_tensor * build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il);
+    ggml_tensor * build_attention_layer(ggml_tensor * cur, llm_graph_input_attn_kv * inp_attn,
+        const llama_model & model, const int64_t n_embd_head, const int il);
+};
+
+struct llm_build_neo_bert : public llm_graph_context {
+    llm_build_neo_bert(const llama_model & model, const llm_graph_params & params);
+};
+
+template <bool iswa>
+struct llm_build_olmo2 : public llm_graph_context {
+    llm_build_olmo2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_olmoe : public llm_graph_context {
+    llm_build_olmoe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_olmo : public llm_graph_context {
+    llm_build_olmo(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_openai_moe_iswa : public llm_graph_context {
+    llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_openelm : public llm_graph_context {
+    llm_build_openelm(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_orion : public llm_graph_context {
+    llm_build_orion(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_pangu_embedded : public llm_graph_context {
+    llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_phi2 : public llm_graph_context {
+    llm_build_phi2(const llama_model & model, const llm_graph_params & params);
+};
+
+template<bool iswa>
+struct llm_build_phi3 : public llm_graph_context {
+    llm_build_phi3(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_plamo2 : public llm_graph_context_mamba {
+    llm_build_plamo2(const llama_model & model, const llm_graph_params & params);
+    private:
+        ggml_tensor * build_plamo2_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il);
+        ggml_tensor * build_plamo2_attn_layer(llm_graph_input_attn_kv * inp, ggml_tensor * inp_pos, ggml_tensor * cur,
+                                                const llama_model & model, int il);
+};
+
+struct llm_build_plamo : public llm_graph_context {
+    llm_build_plamo(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_plm : public llm_graph_context {
+    llm_build_plm(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_qwen2 : public llm_graph_context {
+    llm_build_qwen2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_qwen2moe : public llm_graph_context {
+    llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_qwen2vl : public llm_graph_context {
+    llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_qwen3 : public llm_graph_context {
+    llm_build_qwen3(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_qwen3moe : public llm_graph_context {
+    llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_qwen3vl : public llm_graph_context {
+    llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_qwen3vlmoe : public llm_graph_context {
+    llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params);
+};
+struct llm_build_qwen3next : public llm_graph_context_mamba {
+    llm_build_qwen3next(const llama_model & model, const llm_graph_params & params);
+private:
+    ggml_tensor * build_layer_attn(
+    llm_graph_input_attn_kv * inp_attn,
+                ggml_tensor * cur,
+                ggml_tensor * inp_pos,
+                        int   il);
+
+    ggml_tensor * build_layer_attn_linear(
+         llm_graph_input_rs * inp,
+                ggml_tensor * cur,
+                ggml_tensor * causal_mask,
+                ggml_tensor * identity,
+                        int   il);
+
+    ggml_tensor * build_layer_ffn(
+                ggml_tensor * cur,
+                        int   il);
+
+    ggml_tensor * build_delta_net_recurrent(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * beta,
+                ggml_tensor * state,
+                ggml_tensor * causal_mask,
+                ggml_tensor * identity,
+                        int   il);
+
+    ggml_tensor * build_delta_net_chunking(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * beta,
+                ggml_tensor * state,
+                ggml_tensor * causal_mask,
+                ggml_tensor * identity,
+                        int   il);
+
+    ggml_tensor * build_norm_gated(
+                ggml_tensor * input,
+                ggml_tensor * weights,
+                ggml_tensor * gate,
+                        int   layer);
+
+    const llama_model & model;
+};
+
+struct llm_build_qwen : public llm_graph_context {
+    llm_build_qwen(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_refact : public llm_graph_context {
+    llm_build_refact(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_rnd1 : public llm_graph_context {
+    llm_build_rnd1(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_rwkv6 : public llm_build_rwkv6_base {
+    llm_build_rwkv6(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base {
+    llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_rwkv7 : public llm_build_rwkv7_base {
+    llm_build_rwkv7(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_seed_oss : public llm_graph_context {
+    llm_build_seed_oss(const llama_model & model, const llm_graph_params & params);
+};
+
+template <bool iswa>
+struct llm_build_smallthinker : public llm_graph_context {
+    llm_build_smallthinker(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_smollm3 : public llm_graph_context {
+    llm_build_smollm3(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_solar : public llm_graph_context {
+    llm_build_solar(const llama_model & model, const llm_graph_params & params);
+};
+
+
+struct llm_build_stablelm : public llm_graph_context {
+    llm_build_stablelm(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_starcoder2 : public llm_graph_context {
+    llm_build_starcoder2(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_starcoder : public llm_graph_context {
+    llm_build_starcoder(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_t5_dec : public llm_graph_context {
+    llm_build_t5_dec(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_t5_enc : public llm_graph_context {
+    llm_build_t5_enc(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_wavtokenizer_dec : public llm_graph_context {
+    llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params);
+};
+
+struct llm_build_xverse : public llm_graph_context {
+    llm_build_xverse(const llama_model & model, const llm_graph_params & params);
+};
diff --git a/llama/llama.cpp/src/models/mpt.cpp b/llama/llama.cpp/src/models/mpt.cpp
new file mode 100644
index 00000000..2328e027
--- /dev/null
+++ b/llama/llama.cpp/src/models/mpt.cpp
@@ -0,0 +1,126 @@
+#include "models.h"
+
+
+
+llm_build_mpt::llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * pos;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    if (model.pos_embd) {
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+        pos                   = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+        cb(pos, "pos_embd", -1);
+
+        inpL = ggml_add(ctx0, inpL, pos);
+        cb(inpL, "inpL", -1);
+    }
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * attn_norm;
+
+        attn_norm = build_norm(inpL, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, il);
+        cb(attn_norm, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = attn_norm;
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            if (model.layers[il].bqkv) {
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+            }
+
+            if (hparams.f_clamp_kqv > 0.0f) {
+                cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                cb(cur, "wqkv_clamped", il);
+            }
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 0 * sizeof(float) * (n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+
+            // Q/K Layernorm
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
+                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed forward
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                NULL, NULL, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                model.layers[il].ffn_act, LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/nemotron-h.cpp b/llama/llama.cpp/src/models/nemotron-h.cpp
new file mode 100644
index 00000000..54143488
--- /dev/null
+++ b/llama/llama.cpp/src/models/nemotron-h.cpp
@@ -0,0 +1,121 @@
+#include "models.h"
+
+
+
+llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+    ggml_build_forward_expand(gf, inpL);
+
+    auto * inp = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        if (hparams.is_recurrent(il)) {
+            // ssm layer //
+            cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        } else if (hparams.n_ff(il) == 0) {
+            // attention layer //
+            cur = build_attention_layer(cur, inp->get_attn(), model, n_embd_head, il);
+        } else {
+            cur = build_ffn_layer(cur, model, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // add residual
+        cur = ggml_add(ctx0, cur, inpSA);
+        cb(cur, "nemotron_h_block_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *             cur,
+                                                          llm_graph_input_attn_kv * inp_attn,
+                                                          const llama_model &       model,
+                                                          const int64_t             n_embd_head,
+                                                          const int                 il) {
+    // compute Q and K and (optionally) RoPE them
+    ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur, "Qcur", il);
+    if (model.layers[il].bq) {
+        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+        cb(Qcur, "Qcur", il);
+    }
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+    if (model.layers[il].bk) {
+        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+        cb(Kcur, "Kcur", il);
+    }
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+    if (model.layers[il].bv) {
+        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+        cb(Vcur, "Vcur", il);
+    }
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, model.layers[il].bo,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+    cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
+    cur = build_ffn(cur,
+            model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+            NULL, NULL, NULL,
+            model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+            NULL, LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+    cb(cur, "ffn_out", il);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}
diff --git a/llama/llama.cpp/src/models/nemotron.cpp b/llama/llama.cpp/src/models/nemotron.cpp
new file mode 100644
index 00000000..fcead041
--- /dev/null
+++ b/llama/llama.cpp/src/models/nemotron.cpp
@@ -0,0 +1,122 @@
+#include "models.h"
+
+llm_build_nemotron::llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    //GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm,
+                model.layers[il].ffn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                NULL,                      NULL,                        NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/neo-bert.cpp b/llama/llama.cpp/src/models/neo-bert.cpp
new file mode 100644
index 00000000..7c32bfca
--- /dev/null
+++ b/llama/llama.cpp/src/models/neo-bert.cpp
@@ -0,0 +1,104 @@
+#include "models.h"
+
+llm_build_neo_bert::llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // construct input embeddings (token, type, position)
+    inpL = build_inp_embd(model.tok_embd);
+    cb(inpL, "inp_embd", -1);
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * cur = inpL;
+
+        // pre-norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+
+        {
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            // self-attention
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            // RoPE
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            cb(cur, "kqv_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        // re-add the layer input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        ggml_tensor * ffn_inp = cur;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // pre-norm
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,
+                NULL, NULL, NULL, NULL, NULL,
+                model.layers[il].ffn_down,
+                NULL, NULL, NULL,
+                LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+
+        // attentions bypass the intermediate layer
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm_enc, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_embd", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/olmo.cpp b/llama/llama.cpp/src/models/olmo.cpp
new file mode 100644
index 00000000..bbd623f1
--- /dev/null
+++ b/llama/llama.cpp/src/models/olmo.cpp
@@ -0,0 +1,121 @@
+#include "models.h"
+
+llm_build_olmo::llm_build_olmo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                NULL, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (hparams.f_clamp_kqv > 0.0f) {
+                Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (hparams.f_clamp_kqv > 0.0f) {
+                Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (hparams.f_clamp_kqv > 0.0f) {
+                Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                NULL, NULL,
+                LLM_NORM, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            NULL, NULL,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/olmo2.cpp b/llama/llama.cpp/src/models/olmo2.cpp
new file mode 100644
index 00000000..713552da
--- /dev/null
+++ b/llama/llama.cpp/src/models/olmo2.cpp
@@ -0,0 +1,150 @@
+#include "models.h"
+
+template <bool iswa>
+llm_build_olmo2<iswa>::llm_build_olmo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = inpL;
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            const bool is_swa = hparams.is_swa(il);
+
+            if (is_swa) {
+                // For sliding window layers, Olmo3 use regular rope with no yarn rope scaling.
+                // This is achieved here by setting freq_scale and attn_factor to 1.
+                // We also set ext_factor to 0 to avoid a few unnecessary computations.
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
+                    0.0, 1.0, beta_fast, beta_slow
+                    );
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
+                    0.0, 1.0, beta_fast, beta_slow
+                    );
+            } else {
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_ffn(ffn_inp,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_olmo2<false>;
+template struct llm_build_olmo2<true>;
diff --git a/llama/llama.cpp/src/models/olmoe.cpp b/llama/llama.cpp/src/models/olmoe.cpp
new file mode 100644
index 00000000..b8b6988f
--- /dev/null
+++ b/llama/llama.cpp/src/models/olmoe.cpp
@@ -0,0 +1,124 @@
+#include "models.h"
+
+llm_build_olmoe::llm_build_olmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, false,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/openai-moe-iswa.cpp b/llama/llama.cpp/src/models/openai-moe-iswa.cpp
new file mode 100644
index 00000000..96596709
--- /dev/null
+++ b/llama/llama.cpp/src/models/openai-moe-iswa.cpp
@@ -0,0 +1,124 @@
+#include "models.h"
+
+llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, nullptr,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1) {
+            // skip computing output for unused tokens
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = ffn_inp;
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, nullptr,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        // MoE branch
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,  model.layers[il].ffn_gate_inp_b,
+                model.layers[il].ffn_up_exps,   model.layers[il].ffn_up_exps_b,
+                model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b,
+                model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SWIGLU_OAI_MOE, false,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/openelm.cpp b/llama/llama.cpp/src/models/openelm.cpp
new file mode 100644
index 00000000..ee46a337
--- /dev/null
+++ b/llama/llama.cpp/src/models/openelm.cpp
@@ -0,0 +1,124 @@
+#include "models.h"
+
+llm_build_openelm::llm_build_openelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const int64_t n_head    = hparams.n_head(il);
+        const int64_t n_head_kv = hparams.n_head_kv(il);
+        const int64_t n_head_qkv = 2*n_head_kv + n_head;
+
+        cur = inpL;
+        ggml_tensor * residual = cur;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, cur->nb[1], cur->nb[2], 0);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv)));
+            cb(Vcur, "Vcur", il);
+
+            Qcur = build_norm(Qcur,
+                    model.layers[il].attn_q_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur", il);
+
+            Kcur = build_norm(Kcur,
+                    model.layers[il].attn_k_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, NULL,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, NULL,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Qcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+            cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        inpL = cur;
+    }
+    cur = inpL;
+
+    // norm
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/orion.cpp b/llama/llama.cpp/src/models/orion.cpp
new file mode 100644
index 00000000..bb02273b
--- /dev/null
+++ b/llama/llama.cpp/src/models/orion.cpp
@@ -0,0 +1,123 @@
+#include "models.h"
+
+llm_build_orion::llm_build_orion(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            // if (model.layers[il].bq) {
+            //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            //     cb(Qcur, "Qcur", il);
+            // }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            // if (model.layers[il].bk) {
+            //     Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            //     cb(Kcur, "Kcur", il);
+            // }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            // if (model.layers[il].bv) {
+            //     Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            //     cb(Vcur, "Vcur", il);
+            // }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/pangu-embedded.cpp b/llama/llama.cpp/src/models/pangu-embedded.cpp
new file mode 100644
index 00000000..664572a5
--- /dev/null
+++ b/llama/llama.cpp/src/models/pangu-embedded.cpp
@@ -0,0 +1,121 @@
+#include "models.h"
+
+
+llm_build_pangu_embedded::llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (model.output_b != nullptr) {
+        cur = ggml_add(ctx0, cur, model.output_b);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/phi2.cpp b/llama/llama.cpp/src/models/phi2.cpp
new file mode 100644
index 00000000..22dbf610
--- /dev/null
+++ b/llama/llama.cpp/src/models/phi2.cpp
@@ -0,0 +1,121 @@
+#include "models.h"
+
+
+llm_build_phi2::llm_build_phi2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * attn_norm_output;
+    ggml_tensor * ffn_output;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        attn_norm_output = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(attn_norm_output, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv) {
+                cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+            } else {
+                Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
+                Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
+                Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            }
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // with phi2, we scale the Q to avoid precision issues
+            // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
+            Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
+            inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
+            attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
+        }
+        // FF
+        {
+            ffn_output = build_ffn(attn_norm_output,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(ffn_output, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_output);
+        cur = ggml_add(ctx0, cur, inpL);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output_no_bias", -1);
+
+    cur = ggml_add(ctx0, cur, model.output_b);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/phi3.cpp b/llama/llama.cpp/src/models/phi3.cpp
new file mode 100644
index 00000000..c8e5da33
--- /dev/null
+++ b/llama/llama.cpp/src/models/phi3.cpp
@@ -0,0 +1,152 @@
+#include "models.h"
+
+template<bool iswa>
+llm_build_phi3<iswa>::llm_build_phi3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        auto * residual = inpL;
+
+        // self-attention
+        {
+            // rope freq factors for 128k context
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            ggml_tensor* attn_norm_output = build_norm(inpL,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM_RMS, il);
+            cb(attn_norm_output, "attn_norm", il);
+
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv) {
+                cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output);
+                cb(cur, "wqkv", il);
+
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head * sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+                }
+                else {
+                Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
+                Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
+                Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            }
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
+            cb(Qcur, "Qcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
+            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+        }
+        cur = ggml_add(ctx0, cur, residual);
+        residual = cur;
+
+        cur = build_norm(cur,
+                model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+        cur = ggml_add(ctx0, residual, cur);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    if (model.output_b != nullptr) {
+        cb(cur, "result_output_no_bias", -1);
+        cur = ggml_add(ctx0, cur, model.output_b);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_phi3<false>;
+template struct llm_build_phi3<true>;
diff --git a/llama/llama.cpp/src/models/plamo.cpp b/llama/llama.cpp/src/models/plamo.cpp
new file mode 100644
index 00000000..04ff709f
--- /dev/null
+++ b/llama/llama.cpp/src/models/plamo.cpp
@@ -0,0 +1,110 @@
+#include "models.h"
+
+llm_build_plamo::llm_build_plamo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * sa_inp = cur;
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur    = ggml_get_rows(ctx0,    cur, inp_out_ids);
+            sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids);
+            inpL   = ggml_get_rows(ctx0,   inpL, inp_out_ids);
+        }
+        ggml_tensor * sa_out = cur;
+
+        cur = sa_inp;
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, sa_out);
+        cur = ggml_add(ctx0, cur, inpL);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/plamo2.cpp b/llama/llama.cpp/src/models/plamo2.cpp
new file mode 100644
index 00000000..31115a08
--- /dev/null
+++ b/llama/llama.cpp/src/models/plamo2.cpp
@@ -0,0 +1,316 @@
+#include "models.h"
+
+llm_build_plamo2::llm_build_plamo2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+    cb(inpL, "embedding_output", -1);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_hybrid = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * residual = inpL;
+
+        // ggml_graph_add_node(gf, model.layers[il].attn_norm);
+        // cb(model.layers[il].attn_norm, "attn_norm", il);
+
+        // pre_mixer_norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+
+        // check if this layer is Mamba or Attention
+        bool is_mamba_layer = hparams.is_recurrent(il);
+
+        if (is_mamba_layer) {
+            // PLaMo-2 Mamba layer
+            cur = build_plamo2_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);
+        } else {
+            // PLaMo-2 Attention layer
+            cur = build_plamo2_attn_layer(inp_hybrid->get_attn(), inp_pos, cur, model, il);
+        }
+
+        // post_mixer_norm
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        // residual connection
+        cur = ggml_add(ctx0, cur, residual);
+        cb(cur, "attn_residual", il);
+        residual = cur;
+
+        // pre-ffn norm
+        cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_pre_norm", il);
+
+        // feed-forward network
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                NULL, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        // post ffn norm
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
+            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+        }
+
+        // residual connection
+        cur = ggml_add(ctx0, cur, residual);
+        cb(cur, "ffn_residual", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // final norm
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+
+    // Explicitly mark as output tensor to ensure proper backend assignment
+    ggml_set_output(cur);
+
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_plamo2::build_plamo2_attn_layer(llm_graph_input_attn_kv * inp,
+                                                        ggml_tensor *             inp_pos,
+                                                        ggml_tensor *             cur,
+                                                        const llama_model &       model,
+                                                        int                       il) {
+    // self-attention
+    {
+        // PLaMo-2 uses combined QKV tensor
+        ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
+        cb(qkv, "wqkv", il);
+
+        // split QKV tensor into Q, K, V
+        const int64_t n_embd_head_q = hparams.n_embd_head_k;
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+        const int64_t n_embd_head_v = hparams.n_embd_head_v;
+        int32_t       n_head        = hparams.n_head(il);
+        int32_t       n_head_kv     = hparams.n_head_kv(il);
+
+        const int64_t q_offset = 0;
+        const int64_t k_offset = n_embd_head_q * n_head;
+        const int64_t v_offset = k_offset + n_embd_head_k * n_head_kv;
+
+        ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head, n_tokens, n_embd_head_q * sizeof(float),
+                                          qkv->nb[1], q_offset * ggml_element_size(qkv));
+        ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float),
+                                          qkv->nb[1], k_offset * ggml_element_size(qkv));
+        ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv, n_tokens, n_embd_head_v * sizeof(float),
+                                          qkv->nb[1], v_offset * ggml_element_size(qkv));
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+        cb(Qcur, "Qcur_normed", il);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+        cb(Kcur, "Kcur_normed", il);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        cur = build_attn(inp,
+            model.layers[il].wo, NULL,
+            Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f / sqrtf(float(n_embd_head_v)), il);
+    }
+
+    cb(cur, "attn_out", il);
+
+    return cur;
+}
+
+ggml_tensor * llm_build_plamo2::build_plamo2_mamba_layer(llm_graph_input_rs * inp,
+                                                         ggml_tensor *        cur,
+                                                         const llama_model &  model,
+                                                         const llama_ubatch & ubatch,
+                                                         int                  il) {
+    const auto * mctx_cur = inp->mctx;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const int64_t d_conv   = hparams.ssm_d_conv;
+    const int64_t d_inner  = hparams.ssm_d_inner;
+    const int64_t d_state  = hparams.ssm_d_state;
+    const int64_t n_heads  = hparams.ssm_dt_rank;
+    const int64_t head_dim = d_inner / n_heads;
+    const int64_t n_group  = hparams.ssm_n_group;
+    const int64_t n_seqs   = ubatch.n_seqs;
+
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+    ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
+    conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs);
+
+    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    // in_proj: {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * zx = build_lora_mm(model.layers[il].ssm_in, cur);
+    cb(zx, "mamba_in_proj", il);
+    // {8192, 5, 1, 1} -> {8192, 1, 5, 1}
+    zx = ggml_permute(ctx0, zx, 0, 2, 1, 3);
+    zx = ggml_cont_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs);
+    cb(zx, "mamba_in_proj_out", il);
+
+    // split into z and x
+    // => {head_dim * n_heads, n_seq_tokens, n_seqs}
+    ggml_tensor * x = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3],
+                                   head_dim * ggml_element_size(zx));
+    x               = ggml_cont_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs);
+    // x = ggml_permute(ctx0, x, 0, 2, 1, 3);
+    cb(x, "mamba_x_split", il);
+
+    ggml_tensor * z =
+        ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], 0);
+    cb(z, "mamba_z_split", il);
+
+    // conv1d
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
+        cb(conv_x, "mamba_conv1d_input", il);
+
+        // copy last (d_conv - 1) columns back into the state cache
+        ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2],
+                                               n_seq_tokens * (conv_x->nb[0]));
+
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv,
+                                               ggml_view_1d(ctx0, conv_states_all,
+                                                            (d_conv - 1) * (d_inner + 2 * n_group * d_state) * (n_seqs),
+                                                            kv_head * (d_conv - 1) * (d_inner + 2 * n_group * d_state) *
+                                                                ggml_element_size(conv_states_all))));
+        cb(conv_states_all, "mamba_conv1d_state", il);
+
+        // 1D convolution
+        x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
+        cb(x, "mamba_conv1d", il);
+
+        x = ggml_silu(ctx0, x);
+        cb(x, "mamba_conv1d_silu", il);
+    }
+
+    // SSM
+    {
+        // bcdt_proj: {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}
+        ggml_tensor * x_bcdt = build_lora_mm(model.layers[il].ssm_x, x);
+        cb(x_bcdt, "mamba_bcdt_proj", il);
+
+        // split into dt, B, C
+        const int64_t dt_dim = std::max(64, int(hparams.n_embd / 16));
+        ggml_tensor * B  = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], 0);
+        ggml_tensor * C  = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2],
+                                        ggml_element_size(x_bcdt) * d_state);
+        ggml_tensor * dt = ggml_view_3d(ctx0, x_bcdt, dt_dim, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2],
+                                        ggml_element_size(x_bcdt) * (2 * d_state));
+        cb(B, "mamba_B_raw", il);
+        cb(C, "mamba_C_raw", il);
+        cb(dt, "mamba_dt_raw", il);
+
+        // Apply RMS norm to dt, B, C (PLaMo-2 specific)
+        B  = build_norm(B, model.layers[il].ssm_b_norm, NULL, LLM_NORM_RMS, il);
+        C  = build_norm(C, model.layers[il].ssm_c_norm, NULL, LLM_NORM_RMS, il);
+        dt = build_norm(dt, model.layers[il].ssm_dt_norm, NULL, LLM_NORM_RMS, il);
+        cb(B, "mamba_B_normed", il);
+        cb(C, "mamba_C_normed", il);
+        cb(dt, "mamba_dt_normed", il);
+
+        // dt_proj: {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
+        dt = build_lora_mm(model.layers[il].ssm_dt, dt);
+        dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b);
+        cb(dt, "mamba_dt_proj", il);
+
+        ggml_tensor * A = ggml_reshape_2d(ctx0, model.layers[il].ssm_a, 1, n_heads);
+        cb(A, "mamba_A", il);
+
+        x = ggml_view_4d(ctx0, x, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x),
+                         head_dim * n_heads * ggml_element_size(x),
+                         head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
+        B = ggml_view_4d(ctx0, B, d_state, 1, n_seq_tokens, n_seqs, d_state * B->nb[0], B->nb[1], B->nb[2], 0);
+        C = ggml_view_4d(ctx0, C, d_state, 1, n_seq_tokens, n_seqs, d_state * C->nb[0], C->nb[1], C->nb[2], 0);
+
+        // use the states and the indices provided by build_recurrent_state
+        // (this is necessary in order to properly use the states before they are overwritten,
+        //  while avoiding to make unnecessary copies of the states)
+        auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+            ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_heads, mctx_cur->get_size());
+
+            // Custom operator to optimize the parallel associative scan
+            // as described in the Annex D of the Mamba paper.
+            // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+            return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+        };
+
+        ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+        cb(y_ssm, "mamba_ssm_scan", il);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(
+                    ctx0,
+                    ggml_view_1d(ctx0, y_ssm, n_heads * head_dim * d_state * n_seqs,
+                                 n_heads * head_dim * n_seq_tokens * n_seqs * ggml_element_size(y_ssm)),
+                    ggml_view_1d(ctx0, ssm_states_all, n_heads * head_dim * d_state * n_seqs,
+                                 kv_head * n_seqs * n_heads * head_dim * d_state * ggml_element_size(ssm_states_all))));
+        cb(ssm_states_all, "mamba_ssm_states", il);
+
+        ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs,
+                                       head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x),
+                                       head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
+        cb(y, "mamba_y_view", il);
+
+        // Add D parameter and apply gating with z
+        // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs}
+        ggml_tensor * D = ggml_reshape_2d(ctx0, model.layers[il].ssm_d, 1, n_heads);
+        y               = ggml_add(ctx0, y, ggml_mul(ctx0, x, D));
+        cb(y, "mamba_y_add_d", il);
+
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+        cb(y, "mamba_y_swiglu_z", il);
+
+        // out_proj: {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        y   = ggml_view_3d(ctx0, y, head_dim * n_heads, n_seq_tokens, n_seqs, y->nb[2], y->nb[3], 0);
+        cur = build_lora_mm(model.layers[il].ssm_out, y);
+        cb(cur, "mamba_out_proj", il);
+    }
+
+    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
+    cb(cur, "mamba_out", il);
+
+    return cur;
+}
diff --git a/llama/llama.cpp/src/models/plm.cpp b/llama/llama.cpp/src/models/plm.cpp
new file mode 100644
index 00000000..481cbba6
--- /dev/null
+++ b/llama/llama.cpp/src/models/plm.cpp
@@ -0,0 +1,168 @@
+#include "models.h"
+
+llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k));
+
+    const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * q = NULL;
+            q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            cb(q, "q", il);
+
+            // split into {n_head * n_embd_head_qk_nope, n_tokens}
+            ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                    ggml_row_size(q->type, hparams.n_embd_head_k),
+                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_head * n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                    ggml_row_size(q->type, hparams.n_embd_head_k),
+                    ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                    ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
+                    kv_pe_compresseed->nb[1],
+                    0);
+            cb(kv_compressed, "kv_compressed", il);
+
+            // and {n_embd_head_qk_rope, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
+                    kv_pe_compresseed->nb[1],
+                    kv_pe_compresseed->nb[1],
+                    ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            kv_compressed = build_norm(kv_compressed,
+                    model.layers[il].attn_kv_a_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(kv_compressed, "kv_compressed", il);
+
+            // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
+            ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
+            cb(kv, "kv", il);
+
+            // split into {n_head * n_embd_head_qk_nope, n_tokens}
+            ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                    ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
+                    ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    0);
+            cb(k_nope, "k_nope", il);
+
+            // and {n_head * n_embd_head_v, n_tokens}
+            ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+                    ggml_row_size(kv->type, (n_embd_head_qk_nope)));
+            cb(v_states, "v_states", il);
+
+            v_states = ggml_cont(ctx0, v_states);
+            cb(v_states, "v_states", il);
+
+            v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
+                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
+                    0);
+            cb(v_states, "v_states", il);
+
+            q_pe = ggml_rope_ext(
+                    ctx0, q_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+            cb(q_pe, "q_pe", il);
+
+            // shared RoPE key
+            k_pe = ggml_rope_ext(
+                    ctx0, k_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+            cb(k_pe, "k_pe", il);
+
+            ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+            cb(q_states, "q_states", il);
+
+            ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+            cb(k_states, "k_states", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                NULL, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/qwen.cpp b/llama/llama.cpp/src/models/qwen.cpp
new file mode 100644
index 00000000..31fd9b73
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen.cpp
@@ -0,0 +1,108 @@
+#include "models.h"
+
+
+llm_build_qwen::llm_build_qwen(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 2*sizeof(float)*(n_embd));
+
+            // using mode = 2 for neox mode
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward forward
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/qwen2.cpp b/llama/llama.cpp/src/models/qwen2.cpp
new file mode 100644
index 00000000..587a9324
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen2.cpp
@@ -0,0 +1,117 @@
+#include "models.h"
+
+llm_build_qwen2::llm_build_qwen2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (model.output_b != nullptr) {
+        cur = ggml_add(ctx0, cur, model.output_b);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/qwen2moe.cpp b/llama/llama.cpp/src/models/qwen2moe.cpp
new file mode 100644
index 00000000..49142b71
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen2moe.cpp
@@ -0,0 +1,151 @@
+#include "models.h"
+
+llm_build_qwen2moe::llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, false,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // FFN shared expert
+        {
+            ggml_tensor * cur_gate_inp = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
+            cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
+
+            // sigmoid
+            ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
+            cb(cur_gate, "ffn_shexp_gate", il);
+
+            ggml_tensor * cur_ffn = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur_ffn, "ffn_shexp", il);
+
+            ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
+            cb(ffn_shexp_out, "ffn_shexp_out", il);
+
+            moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
+            cb(moe_out, "ffn_out", il);
+
+            cur = moe_out;
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/qwen2vl.cpp b/llama/llama.cpp/src/models/qwen2vl.cpp
new file mode 100644
index 00000000..9be38675
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen2vl.cpp
@@ -0,0 +1,117 @@
+#include "models.h"
+
+llm_build_qwen2vl::llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_multi(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_multi(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/qwen3.cpp b/llama/llama.cpp/src/models/qwen3.cpp
new file mode 100644
index 00000000..a5cfffa5
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen3.cpp
@@ -0,0 +1,117 @@
+#include "models.h"
+
+llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/qwen3moe.cpp b/llama/llama.cpp/src/models/qwen3moe.cpp
new file mode 100644
index 00000000..888534fb
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen3moe.cpp
@@ -0,0 +1,124 @@
+#include "models.h"
+
+llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+        cur = moe_out;
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/qwen3next.cpp b/llama/llama.cpp/src/models/qwen3next.cpp
new file mode 100644
index 00000000..c8f1b5ec
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen3next.cpp
@@ -0,0 +1,1042 @@
+#include "ggml.h"
+#include "models.h"
+
+#define CHUNK_SIZE 64
+
+llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params), model(model) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+    cb(inpL, "model.embed_tokens", -1);
+
+    auto * inp = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    ggml_tensor * causal_mask =
+        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens, ubatch.n_seq_tokens), 1.0f),
+                    GGML_TRI_TYPE_LOWER);
+
+    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens), 1.0f));
+
+    ggml_build_forward_expand(gf, causal_mask);
+    ggml_build_forward_expand(gf, identity);
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // Determine layer type and build appropriate attention mechanism
+        if (hparams.is_recurrent(il)) {
+            // Linear attention layer (gated delta net)
+            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, il);
+        } else {
+            // Full attention layer
+            cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // Residual connection
+        cur = ggml_add(ctx0, cur, inpSA);
+        cb(cur, "attn_residual", il);
+
+        // Save the tensor before post-attention norm for residual connection
+        ggml_tensor * ffn_residual = cur;
+
+        // Post-attention norm
+        ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);
+        cb(attn_post_norm, "attn_post_norm", il);
+
+        // FFN layer (MoE or dense) - without residual connection
+        cur = build_layer_ffn(attn_post_norm, il);
+        cb(cur, "ffn_out", il);
+
+        // Residual connection for FFN - add to the tensor from before post_attention_layernorm
+        cur = ggml_add(ctx0, cur, ffn_residual);
+        cb(cur, "post_moe", il);
+
+        // Input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    // Final norm
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // LM head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
+        ggml_tensor * q,
+        ggml_tensor * k,
+        ggml_tensor * v,
+        ggml_tensor * g,
+        ggml_tensor * beta,
+        ggml_tensor * state,
+        ggml_tensor * causal_mask,
+        ggml_tensor * identity,
+        int           il) {
+    GGML_ASSERT(ggml_is_contiguous(q));
+    GGML_ASSERT(ggml_is_contiguous(k));
+    GGML_ASSERT(ggml_is_contiguous(v));
+    GGML_ASSERT(ggml_is_contiguous(g));
+    GGML_ASSERT(ggml_is_contiguous(beta));
+    GGML_ASSERT(ggml_is_contiguous(state));
+
+    const int64_t S_k      = q->ne[0];
+    const int64_t H_k      = q->ne[1];
+    const int64_t n_tokens = q->ne[2];
+    const int64_t n_seqs   = q->ne[3];
+
+    const int64_t S_v = v->ne[0];
+    const int64_t H_v = v->ne[1];
+
+    GGML_ASSERT(v->ne[2] == n_tokens);
+    GGML_ASSERT(k->ne[2] == n_tokens);
+    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
+    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
+    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
+
+    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
+    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
+
+    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
+
+    // TODO: can this ever be false?
+    const bool use_qk_l2norm = true;
+
+    if (use_qk_l2norm) {
+        const float eps_norm = hparams.f_norm_rms_eps;
+
+        q = ggml_l2_norm(ctx0, q, eps_norm);
+        k = ggml_l2_norm(ctx0, k, eps_norm);
+    }
+
+    const float scale = 1.0f / sqrtf(S_v);
+
+    q = ggml_scale(ctx0, q, scale);
+
+    beta = ggml_sigmoid(ctx0, beta);
+
+    ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity);
+
+    cb(q, "q_in", il);
+    cb(k, "k_in", il);
+    cb(v, "v_in", il);
+    cb(beta, "beta_in", il);
+    cb(g, "g_in", il);
+
+    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
+    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
+    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
+    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
+
+    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
+    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
+
+    cb(q, "q_perm", il);
+    cb(k, "k_perm", il);
+    cb(v, "v_perm", il);
+    cb(beta, "beta_perm", il);
+    cb(g, "g_perm", il);
+    cb(state, "state_in", il);
+
+    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
+    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
+    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
+    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
+
+    // Do padding
+    const int64_t chunk_size = CHUNK_SIZE;
+
+    const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
+    const int64_t n_chunks = (n_tokens + pad) / chunk_size;
+
+    q = ggml_pad(ctx0, q, 0, pad, 0, 0);
+    k = ggml_pad(ctx0, k, 0, pad, 0, 0);
+    v = ggml_pad(ctx0, v, 0, pad, 0, 0);
+    g = ggml_pad(ctx0, g, pad, 0, 0, 0);
+    beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
+
+    cb(q, "q_pad", il);
+    cb(k, "k_pad", il);
+    cb(v, "v_pad", il);
+    cb(beta, "beta_pad", il);
+    cb(g, "g_pad", il);
+
+    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
+    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
+
+    cb(v_beta, "v_beta", il);
+    cb(k_beta, "k_beta", il);
+
+    ggml_tensor * chunked_mask =
+        ggml_view_4d(ctx0, causal_mask, chunk_size,
+                chunk_size,         causal_mask->ne[2], causal_mask->ne[3],
+                causal_mask->nb[1], causal_mask->nb[2], causal_mask->nb[3], 0);
+
+    ggml_tensor * chunked_diag_mask =
+        ggml_view_4d(ctx0, causal_diag_mask, chunk_size,
+                chunk_size,              causal_diag_mask->ne[2], causal_diag_mask->ne[3],
+                causal_diag_mask->nb[1], causal_diag_mask->nb[2], causal_diag_mask->nb[3], 0);
+
+    ggml_tensor * chunked_identity =
+        ggml_view_4d(ctx0, identity, chunk_size,
+            chunk_size,      identity->ne[2], identity->ne[3],
+            identity->nb[1], identity->nb[2], identity->nb[3], 0);
+
+    q      = ggml_cont_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
+    k      = ggml_cont_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
+    k_beta = ggml_cont_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
+    v      = ggml_cont_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
+    v_beta = ggml_cont_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
+
+    g    = ggml_cont_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
+    beta = ggml_cont_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
+
+    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
+
+    cb(g_cumsum, "g_cumsum", il);
+
+    ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
+    ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
+
+    ggml_tensor * gcs_j_broadcast =
+        ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
+
+    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
+
+    cb(decay_mask, "decay_mask", il);
+
+    decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask);
+    decay_mask = ggml_exp(ctx0, decay_mask);
+    decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask);
+
+    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
+
+    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
+    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, chunked_mask));
+
+    cb(attn, "attn_pre_solve", il);
+
+    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, chunked_mask);
+    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, chunked_identity, attn_lower), attn_lower);
+
+    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
+    attn                     = ggml_mul(ctx0, lin_solve, chunked_mask);
+    attn                     = ggml_add(ctx0, attn, chunked_identity);
+
+    cb(attn, "attn_solved", il);
+
+    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
+
+    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
+    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
+
+    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
+
+    cb(kbeta_gexp, "kbeta_gexp", il);
+
+    ggml_tensor * k_cumdecay =
+        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
+
+    cb(k_cumdecay, "k_cumdecay", il);
+
+    ggml_tensor * core_attn_out = nullptr;
+    ggml_tensor * new_state = ggml_dup(ctx0, state);
+
+    cb(new_state, "new_state", il);
+
+    for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
+        auto chunkify = [=](ggml_tensor * t) {
+            return ggml_cont(ctx0, ggml_view_4d(ctx0, t, t->ne[0], chunk_size, 1, t->ne[3],
+                t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk));
+        };
+
+        auto chunkify_g = [=](ggml_tensor * t) {
+            return ggml_cont(ctx0, ggml_view_4d(ctx0, t, chunk_size, t->ne[1], 1, t->ne[3],
+                t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk));
+        };
+
+        ggml_tensor * k_chunk = chunkify(k);
+        ggml_tensor * q_chunk = chunkify(q);
+        ggml_tensor * v_chunk = chunkify(v);
+
+        ggml_tensor * g_cs_chunk = chunkify_g(g_cumsum);
+        ggml_tensor * g_cs_chunk_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cs_chunk));
+
+        ggml_tensor * decay_mask_chunk = chunkify(decay_mask);
+        ggml_tensor * k_cumdecay_chunk = chunkify(k_cumdecay);
+
+        ggml_tensor * gexp_chunk = ggml_exp(ctx0, g_cs_chunk_t);
+
+        // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
+        attn = ggml_mul_mat(ctx0, k_chunk, q_chunk);
+        attn = ggml_mul(ctx0, attn, decay_mask_chunk);
+        attn = ggml_mul(ctx0, attn, ggml_add(ctx0, chunked_identity, chunked_mask));
+
+        ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
+
+        // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
+        ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
+
+        // v_new = v_i - v_prime
+        ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
+        ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
+
+        // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
+        ggml_tensor * q_g_exp    = ggml_mul(ctx0, q_chunk, gexp_chunk);
+        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
+
+        // core_attn_out[:, :, i] = attn_inter + attn @ v_new
+        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn);
+
+        ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
+
+        core_attn_out = core_attn_out == nullptr ? core_attn_out_chunk : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 1);
+
+        // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
+        // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
+        // key_gdiff = key * g_diff.unsqueeze(-1)
+        // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
+        // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
+
+        ggml_tensor * g_cum_last =
+            ggml_cont(ctx0, ggml_view_4d(ctx0, g_cs_chunk_t, g_cs_chunk_t->ne[0], 1, g_cs_chunk_t->ne[2], g_cs_chunk_t->ne[3],
+                                        g_cs_chunk_t->nb[1], g_cs_chunk_t->nb[2], g_cs_chunk_t->nb[3],
+                                        g_cs_chunk_t->nb[0] * (g_cs_chunk_t->ne[1] - 1)));
+
+        ggml_tensor * gexp_last =
+            ggml_reshape_4d(ctx0, ggml_exp(ctx0, g_cum_last), 1, 1, g_cum_last->ne[0] * g_cum_last->ne[2], g_cum_last->ne[3]);
+
+        ggml_tensor * g_cum_last_3d =
+            ggml_reshape_3d(ctx0, g_cum_last, g_cum_last->ne[0], g_cum_last->ne[2], g_cum_last->ne[3]);
+
+        ggml_tensor * g_cumsum_3d = ggml_reshape_3d(ctx0, g_cs_chunk, g_cs_chunk->ne[0], g_cs_chunk->ne[2], g_cs_chunk->ne[3]);
+
+        ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum_3d, g_cum_last_3d));
+
+        ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
+
+        ggml_tensor * key_gdiff = ggml_mul(ctx0, k_chunk,
+                                        ggml_reshape_4d(ctx0, g_diff_exp, 1, g_diff_exp->ne[0], g_diff_exp->ne[1],
+                                                        g_diff_exp->ne[2] * g_diff_exp->ne[3]));
+
+        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)));
+
+        new_state = ggml_add(ctx0,
+            ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last, gexp_last->ne[0], gexp_last->ne[1], H_v, n_seqs)),
+            ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
+    }
+
+    core_attn_out = ggml_cont_4d(ctx0, core_attn_out, S_v, chunk_size * n_chunks, H_v, n_seqs);
+
+    ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, S_v, n_tokens, H_v, n_seqs, core_attn_out->nb[1], core_attn_out->nb[2], core_attn_out->nb[3], 0);
+    cb(output_tokens, "output_tokens", il);
+
+    // flatten output
+    ggml_tensor * flat_output =
+        ggml_cont_1d(ctx0, ggml_permute(ctx0, output_tokens, 0, 2, 1, 3), S_v * H_v * n_tokens * n_seqs);
+
+    ggml_tensor * flat_state = ggml_cont_1d(ctx0, new_state, S_v * S_v * H_v * n_seqs);
+
+    return ggml_concat(ctx0, flat_output, flat_state, 0);
+}
+
+ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent(
+        ggml_tensor * q,
+        ggml_tensor * k,
+        ggml_tensor * v,
+        ggml_tensor * g,
+        ggml_tensor * beta,
+        ggml_tensor * state,
+        ggml_tensor * causal_mask,
+        ggml_tensor * identity,
+        int           il) {
+    GGML_ASSERT(ggml_is_contiguous(q));
+    GGML_ASSERT(ggml_is_contiguous(k));
+    GGML_ASSERT(ggml_is_contiguous(v));
+    GGML_ASSERT(ggml_is_contiguous(g));
+    GGML_ASSERT(ggml_is_contiguous(beta));
+    GGML_ASSERT(ggml_is_contiguous(state));
+
+    const int64_t S_k      = q->ne[0];
+    const int64_t H_k      = q->ne[1];
+    const int64_t n_tokens = q->ne[2];
+    const int64_t n_seqs   = q->ne[3];
+
+    const int64_t S_v = v->ne[0];
+    const int64_t H_v = v->ne[1];
+
+    GGML_ASSERT(v->ne[2] == n_tokens);
+    GGML_ASSERT(k->ne[2] == n_tokens);
+    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
+    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
+    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
+
+    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
+    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
+
+    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
+
+    // TODO: can this ever be false?
+    const bool use_qk_l2norm = true;
+
+    if (use_qk_l2norm) {
+        const float eps_norm = hparams.f_norm_rms_eps;
+
+        q = ggml_l2_norm(ctx0, q, eps_norm);
+        k = ggml_l2_norm(ctx0, k, eps_norm);
+    }
+
+    const float scale = 1.0f / sqrtf(S_v);
+
+    q = ggml_scale(ctx0, q, scale);
+
+    beta = ggml_sigmoid(ctx0, beta);
+
+    ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity);
+
+    cb(q, "q_in", il);
+    cb(k, "k_in", il);
+    cb(v, "v_in", il);
+    cb(beta, "beta_in", il);
+    cb(g, "g_in", il);
+
+    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
+    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
+    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
+    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
+
+    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
+    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
+
+    cb(q, "q_perm", il);
+    cb(k, "k_perm", il);
+    cb(v, "v_perm", il);
+    cb(beta, "beta_perm", il);
+    cb(g, "g_perm", il);
+    cb(state, "state_in", il);
+
+    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
+    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
+    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
+    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
+
+    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
+    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
+
+    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
+
+    cb(k_beta, "k_beta", il);
+    cb(v_beta, "v_beta", il);
+    cb(g_cumsum, "g_cumsum", il);
+
+    ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, n_tokens, 1, H_v, n_seqs);  // [chunk_size, 1, n_tokens, n_seqs]
+    ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, n_tokens, H_v, n_seqs);  // [1, chunk_size, n_tokens, n_seqs]
+
+    // Broadcast both tensors to [chunk_size, chunk_size, H_v, n_seqs]
+    // ggml_tensor * gcs_i_broadcast =
+    //     ggml_repeat_4d(ctx0, gcs_i, GGML_DELTA_NET_CHUNK, GGML_DELTA_NET_CHUNK, num_chunks * H_v,
+    //                     n_seqs);  // [chunk_size, 1, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs]
+    // Don't need this, this one will get auto-broadcast
+    ggml_tensor * gcs_j_broadcast =
+        ggml_repeat_4d(ctx0, gcs_j, n_tokens, n_tokens, H_v, n_seqs);  // [1, chunk_size, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs]
+
+    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
+
+    // Apply lower triangular mask to ensure attention is causal (only past tokens influence current)
+    decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask);
+    // Apply exponential to get the decay mask values
+    decay_mask = ggml_exp(ctx0, decay_mask);
+    // Apply lower triangular mask again to ensure only lower triangular values remain
+    decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask);
+
+    cb(decay_mask, "decay_mask", il);
+
+    // attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0)
+    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
+
+    cb(kmulkbeta, "kmulkbeta", il);
+
+    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
+    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
+
+    cb(attn, "attn_pre_rec", il);
+
+    // for i in range(1, chunk_size):
+    //          row = attn[..., i, :i].clone()
+    //          sub = attn[..., :i, :i].clone()
+    //          attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
+    // attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
+    //
+    // We reduce this to a linear triangular solve: AX = B, where B = attn, A = I - tril(A)
+    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
+    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
+
+    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
+    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
+    attn                     = ggml_add(ctx0, attn, identity);
+
+    // value = attn @ v_beta
+    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
+
+    cb(v, "value_beta", il);
+
+    // k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1))
+    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
+    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
+
+    cb(gexp, "g_cum_exp", il);
+
+    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
+
+    cb(kbeta_gexp, "kbeta_gexp", il);
+
+    ggml_tensor * k_cumdecay =
+        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
+
+    cb(k_cumdecay, "k_cumdecay", il);
+
+    // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
+    attn = ggml_mul_mat(ctx0, k, q);
+    attn = ggml_mul(ctx0, attn, decay_mask);
+    attn = ggml_mul(ctx0, attn, ggml_add(ctx0, identity, causal_mask));
+
+    cb(attn, "attn_decay_key", il);
+
+    ggml_tensor * state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state));
+
+    // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
+    ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay);
+
+    cb(v_prime, "v_prime", il);
+
+    // v_new = v_i - v_prime
+    ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v, v_prime), v_prime);
+
+    ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
+
+    cb(v_new, "v_new", il);
+
+    // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
+    ggml_tensor * q_g_exp    = ggml_mul(ctx0, q, gexp);
+    ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
+
+    cb(attn_inter, "attn_inter", il);
+
+    // core_attn_out[:, :, i] = attn_inter + attn @ v_new
+    ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn);
+
+    cb(v_attn, "v_attn", il);
+
+    ggml_tensor * core_attn_out = ggml_add(ctx0, attn_inter, v_attn);
+
+    cb(core_attn_out, "core_attn_out", il);
+
+    // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
+    // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
+    // key_gdiff = key * g_diff.unsqueeze(-1)
+    // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
+    // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
+
+    ggml_tensor * g_cum_last =
+        ggml_cont(ctx0, ggml_view_4d(ctx0, g_cumsum_t, g_cumsum_t->ne[0], 1, g_cumsum_t->ne[2], g_cumsum_t->ne[3],
+                                    g_cumsum_t->nb[1], g_cumsum_t->nb[2], g_cumsum_t->nb[3],
+                                    g_cumsum_t->nb[0] * (g_cumsum_t->ne[1] - 1)));
+
+    cb(g_cum_last, "g_cum_last", il);
+
+    ggml_tensor * gexp_last =
+        ggml_reshape_4d(ctx0, ggml_exp(ctx0, g_cum_last), 1, 1, g_cum_last->ne[0] * g_cum_last->ne[2], g_cum_last->ne[3]);
+
+    cb(gexp_last, "gexp_last", il);
+
+    ggml_tensor * g_cum_last_3d =
+        ggml_reshape_3d(ctx0, g_cum_last, g_cum_last->ne[0], g_cum_last->ne[2], g_cum_last->ne[3]);
+
+    cb(g_cum_last_3d, "g_cum_last_3d", il);
+
+    ggml_tensor * g_cumsum_3d = ggml_reshape_3d(ctx0, g_cumsum, g_cumsum->ne[0], g_cumsum->ne[2], g_cumsum->ne[3]);
+
+    cb(g_cumsum_3d, "g_cumsum_3d", il);
+
+    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum_3d, g_cum_last_3d));
+
+    cb(g_diff, "g_diff", il);
+
+    ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
+
+    cb(g_diff_exp, "g_diff_exp", il);
+
+    ggml_tensor * key_gdiff = ggml_mul(ctx0, k,
+                                    ggml_reshape_4d(ctx0, g_diff_exp, 1, g_diff_exp->ne[0], g_diff_exp->ne[1],
+                                                    g_diff_exp->ne[2] * g_diff_exp->ne[3]));
+
+    cb(key_gdiff, "key_gdiff", il);
+
+    ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)));
+
+    cb(kgdmulvnew, "kgdmulvnew", il);
+
+    state = ggml_add(ctx0, ggml_mul(ctx0, state, gexp_last), kgdmulvnew);
+
+    cb(state, "new_state", il);
+
+    // flatten output
+    ggml_tensor * flat_output =
+        ggml_cont_1d(ctx0, ggml_permute(ctx0, core_attn_out, 0, 2, 1, 3), S_v * H_v * n_tokens * n_seqs);
+
+    ggml_tensor * flat_state = ggml_cont_1d(ctx0, state, S_v * S_v * H_v * n_seqs);
+
+    return ggml_concat(ctx0, flat_output, flat_state, 0);
+}
+
+ggml_tensor * llm_build_qwen3next::build_norm_gated(
+        ggml_tensor * input,
+        ggml_tensor * weights,
+        ggml_tensor * gate,
+        int           layer) {
+    ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);
+    ggml_tensor * gated_silu = ggml_silu(ctx0, gate);
+
+    return ggml_mul(ctx0, normalized, gated_silu);
+}
+
+ggml_tensor * llm_build_qwen3next::build_layer_attn(
+        llm_graph_input_attn_kv * inp,
+        ggml_tensor *             cur,
+        ggml_tensor *             inp_pos,
+        int                       il) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
+
+    // Qwen3Next uses a single Q projection that outputs query + gate
+    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur_full, "Qcur_full", il);
+
+    Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1);
+
+    // Split Q projection into query and gate
+    // The split should be along dimension 0 (the feature dimension)
+    ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
+                                             Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);
+    ggml_tensor * gate =
+        ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
+                     Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));
+    cb(Qcur, "Qcur", il);
+    cb(gate, "gate", il);
+
+    // Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention
+    Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+    cb(Qcur, "Qcur_reshaped", il);
+
+    // Apply Q normalization
+    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
+    cb(Qcur, "Qcur_normed", il);
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+
+    // Apply K normalization
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+    Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
+    cb(Kcur, "Kcur_normed", il);
+
+    // Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads)
+    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
+    cb(gate, "gate_reshaped", il);
+
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+    // Apply RoPE
+    Qcur = ggml_rope_ext(
+            ctx0, Qcur, inp_pos, nullptr,
+            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+
+    Kcur = ggml_rope_ext(
+            ctx0, Kcur, inp_pos, nullptr,
+            n_rot, rope_type, n_ctx_orig, freq_base,
+            freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    // Attention computation
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    cur = build_attn(inp,
+                nullptr, nullptr,
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+    cb(cur, "attn_pregate", il);
+
+    ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate);
+    cb(gate_sigmoid, "gate_sigmoid", il);
+
+    cur = ggml_mul(ctx0, cur, gate_sigmoid);
+    cb(cur, "attn_gated", il);
+
+    cur = build_lora_mm(model.layers[il].wo, cur);
+    cb(cur, "attn_output", il);
+
+    return cur;
+}
+
+ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
+        llm_graph_input_rs * inp,
+        ggml_tensor *        cur,
+        ggml_tensor *        causal_mask,
+        ggml_tensor *        identity,
+        int                  il) {
+    const auto * mctx_cur = inp->mctx;
+
+    const int64_t d_inner      = hparams.ssm_d_inner;
+    const int64_t n_seqs       = ubatch.n_seqs;
+    const int64_t head_k_dim   = hparams.ssm_d_state;
+    const int64_t num_k_heads  = hparams.ssm_n_group;
+    const int64_t num_v_heads  = hparams.ssm_dt_rank;
+    const int64_t head_v_dim   = d_inner / num_v_heads;
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    // Input projections
+    ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, cur);
+    cb(mixed_qkvz, "linear_attn_mixed_qkvz", il);
+
+    ggml_tensor * mixed_ba = build_lora_mm(model.layers[il].ssm_beta_alpha, cur);
+    cb(mixed_ba, "linear_attn_mixed_ba", il);
+
+    int64_t       qkvz_new_dim        = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads);
+    ggml_tensor * mixed_qkvz_reshaped = ggml_cont_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);
+
+    // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads]
+    int64_t       ba_new_dim        = 2 * num_v_heads / num_k_heads;
+    ggml_tensor * mixed_ba_reshaped = ggml_cont_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);
+
+    // Split mixed_ba into b and a (beta and alpha parameters)
+    int64_t split_sizes_ba[2] = {
+        num_v_heads / num_k_heads,  // beta size
+        num_v_heads / num_k_heads   // alpha size
+    };
+
+    ggml_tensor * b = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[0], num_k_heads, n_seq_tokens, n_seqs,
+                                   mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3], 0);
+    cb(b, "b", il);
+
+    ggml_tensor * a = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[1], num_k_heads, n_seq_tokens, n_seqs,
+                                   mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3],
+                                   split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));
+    cb(a, "a", il);
+
+    // Reshape b and a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]
+    ggml_tensor * beta  = ggml_cont_3d(ctx0, b, num_v_heads, n_seq_tokens, n_seqs);
+    ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
+
+    GGML_ASSERT(ggml_nelements(beta) + ggml_nelements(alpha) == ggml_nelements(mixed_ba));
+
+    ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
+    ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
+    cb(alpha_softplus, "a_softplus", il);
+    ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus
+    cb(gate, "gate", il);
+
+    // Split mixed_qkvz into query, key, value, z
+    int64_t split_sizes_qkvz[4] = {
+        head_k_dim,                              // query size
+        head_k_dim,                              // key size
+        head_v_dim * num_v_heads / num_k_heads,  // value size
+        head_v_dim * num_v_heads / num_k_heads   // z size
+    };
+
+    ggml_tensor * query =
+        ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[0], num_k_heads, n_seq_tokens, n_seqs,
+                     mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], 0);
+    cb(query, "q", il);
+
+    ggml_tensor * key = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[1], num_k_heads, n_seq_tokens, n_seqs,
+                                     mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],
+                                     split_sizes_qkvz[0] * sizeof(float));
+    cb(key, "k", il);
+
+    ggml_tensor * value =
+        ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[2], num_k_heads, n_seq_tokens, n_seqs,
+                     mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],
+                     (split_sizes_qkvz[0] + split_sizes_qkvz[1]) * sizeof(float));
+    cb(value, "v", il);
+
+    ggml_tensor * z = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[3], num_k_heads, n_seq_tokens, n_seqs,
+                                   mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],
+                                   (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * sizeof(float));
+    cb(z, "z", il);
+
+    GGML_ASSERT(ggml_nelements(query) + ggml_nelements(key) + ggml_nelements(value) + ggml_nelements(z) ==
+                ggml_nelements(mixed_qkvz));
+
+    // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions
+    // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]
+    ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);
+    cb(query_flat, "query_flat", il);
+
+    // key: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]
+    ggml_tensor * key_flat = ggml_cont_3d(ctx0, key, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);
+    cb(key_flat, "key_flat", il);
+
+    // value_reshaped: [head_v_dim, num_v_heads, n_tokens, n_seqs] -> [head_v_dim * num_v_heads, n_tokens, n_seqs]
+    ggml_tensor * value_flat = ggml_cont_3d(ctx0, value, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
+    cb(value_flat, "value_flat", il);
+
+    // Get convolution states from cache
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+    // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
+
+    // Build the convolution states tensor
+    ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
+    cb(conv_states, "conv_states", il);
+
+    // Now concatenate along the feature dimension (dim 0) to get [conv_dim, n_tokens, n_seqs]
+    ggml_tensor * qkv_mixed = ggml_concat(ctx0, query_flat, key_flat, 0);
+    qkv_mixed               = ggml_concat(ctx0, qkv_mixed, value_flat, 0);
+    cb(qkv_mixed, "qkv_mixed", il);
+
+    qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
+    cb(qkv_mixed, "qkv_mixed_permuted", il);
+
+    // Calculate the total conv dimension
+    int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;
+
+    // Calculate convolution kernel size
+    ggml_tensor * conv_kernel      = model.layers[il].ssm_conv1d;
+    const int64_t conv_kernel_size = conv_kernel->ne[0];
+    const int64_t conv_channels    = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
+    conv_states                    = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
+    cb(conv_states, "conv_states_reshaped", il);
+
+    ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
+    cb(conv_input, "conv_input", il);
+
+    // Update convolution state cache
+    // Extract the last (conv_kernel_size - 1) states from conv_input
+    ggml_tensor * last_conv_states =
+        ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1],
+                     conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input));
+    cb(last_conv_states, "last_conv_states", il);
+
+    ggml_tensor * state_update_target =
+        ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs,
+                     kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all));
+    cb(state_update_target, "state_update_target", il);
+
+    ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
+    cb(conv_states_all, "conv_states_updated", il);
+
+    // Apply SSM convolution
+    ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
+    cb(conv_output_proper, "conv_output_raw", il);
+
+    conv_output_proper = ggml_cont(ctx0, ggml_transpose(ctx0, conv_output_proper));
+    cb(conv_output_proper, "conv_output_pre_silu", il);
+
+    ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);
+    cb(conv_output_silu, "conv_output_silu", il);
+
+    ggml_tensor * conv_qkv_mix =
+        ggml_cont_2d(ctx0, ggml_transpose(ctx0, conv_output_silu), qkv_dim, n_seq_tokens * n_seqs);
+    cb(conv_qkv_mix, "conv_qkv_mix", il);
+
+    // Extract the convolved Q, K, V from conv_output
+    ggml_tensor * q_conv =
+        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, conv_qkv_mix->nb[1], 0);
+    cb(q_conv, "q_conv", il);
+    ggml_tensor * k_conv =
+        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, conv_qkv_mix->nb[1],
+                     head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
+    cb(k_conv, "k_conv", il);
+    ggml_tensor * v_conv =
+        ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, conv_qkv_mix->nb[1],
+                     2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
+    cb(v_conv, "v_conv", il);
+
+    // Unsqueeze them
+    q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+
+    beta = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs);
+
+    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
+    state               = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
+    cb(state, "state_predelta", il);
+
+    // if head keys and value keys are different, repeat to force tensors into matching shapes
+    if (num_k_heads != num_v_heads) {
+        GGML_ASSERT(num_v_heads % num_k_heads == 0);
+        int64_t repeat_factor = num_v_heads / num_k_heads;
+
+        // repeat interleave: reshape to (repeat part, 1, remaining part), do repeat, then reshape back
+        ggml_tensor * q_reshaped = ggml_reshape_3d(ctx0, q_conv, head_k_dim, 1, num_k_heads * n_seq_tokens * n_seqs);
+        ggml_tensor * k_reshaped = ggml_reshape_3d(ctx0, k_conv, head_k_dim, 1, num_k_heads * n_seq_tokens * n_seqs);
+
+        // Repeat along the third dimension (the new dimension with size 1)
+        ggml_tensor * q_repeated =
+            ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads * n_seq_tokens * n_seqs, 1);
+        ggml_tensor * k_repeated =
+            ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads * n_seq_tokens * n_seqs, 1);
+
+        // Reshape back to merge the head and repeat dimensions
+        // From [head_dim, num_k_heads, repeat_factor, n_seq_tokens * n_seqs]
+        // Back to [head_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs]
+        q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);
+        k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);
+    }
+
+    cb(q_conv, "q_conv_predelta", il);
+    cb(k_conv, "k_conv_predelta", il);
+    cb(v_conv, "v_conv_predelta", il);
+
+    // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
+    ggml_tensor * attn_out = n_seq_tokens > CHUNK_SIZE ?
+        build_delta_net_chunking (q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il) :
+        build_delta_net_recurrent(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il);
+    cb(attn_out, "attn_out", il);
+
+    // The tensors were concatenated 1d, so we need to extract them 1d as well
+    const int64_t output_flat_size = head_v_dim * num_v_heads * n_seq_tokens * n_seqs;
+    ggml_tensor * attn_out_1d      = ggml_view_1d(ctx0, attn_out, output_flat_size, 0);
+    cb(attn_out_1d, "attn_out_1d", il);
+
+    ggml_tensor * attn_out_final = ggml_cont_4d(ctx0, attn_out_1d, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+    cb(attn_out_final, "attn_out_reshaped", il);
+
+    // Extract the state part (second part of the concatenated tensor)
+    // State starts after n_tokens elements along dimension 1
+    const int64_t state_flat_size = head_v_dim * head_v_dim * num_v_heads * n_seqs;
+
+    ggml_tensor * state_1d =
+        ggml_view_1d(ctx0, attn_out, state_flat_size, output_flat_size * ggml_element_size(attn_out));
+    cb(state_1d, "state_1d", il);
+
+    // Update the recurrent states
+    ggml_build_forward_expand(gf,
+                              ggml_cpy(ctx0, state_1d,
+                                       ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
+                                                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
+
+    GGML_ASSERT(ggml_nelements(attn_out_1d) + ggml_nelements(state_1d) == ggml_nelements(attn_out));
+
+    // Reshape both attn_out_final and z to 2D tensors for normalization
+    // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
+    ggml_tensor * attn_out_2d_final =
+        ggml_cont_2d(ctx0, attn_out_final, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+
+    // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
+    ggml_tensor * z_2d = ggml_cont_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+
+    // Apply gated normalization: self.norm(core_attn_out, z)
+    ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
+
+    // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
+    ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
+    cb(final_output, "final_output", il);
+
+    // Output projection
+    cur = build_lora_mm(model.layers[il].ssm_out, final_output);
+    cb(cur, "linear_attn_out", il);
+
+    // Reshape back to original dimensions
+    cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+    return cur;
+}
+
+ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int il) {
+    // Check if this is an MoE layer
+    if (model.layers[il].ffn_gate_inp != nullptr) {
+        // MoE branch
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used, LLM_FFN_SILU,
+                true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // Add shared experts if present - following Qwen3Next reference implementation
+        if (model.layers[il].ffn_up_shexp != nullptr) {
+            ggml_tensor * ffn_shexp =
+                build_ffn(cur,
+                    model.layers[il].ffn_up_shexp, NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            // Apply shared expert gating as in the reference implementation
+            // The shared expert has its own gate that is sigmoided
+            // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token)
+            ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
+            cb(shared_gate, "shared_expert_gate", il);
+
+            // Apply sigmoid to the gate
+            shared_gate = ggml_sigmoid(ctx0, shared_gate);
+            cb(shared_gate, "shared_expert_gate_sigmoid", il);
+
+            // The gate needs to be broadcast to match the dimensions of ffn_shexp
+            // ffn_shexp is [n_embd, n_tokens, 1, 1] and shared_gate is [1, n_tokens, 1, 1]
+            // We need to repeat the gate along the feature dimension
+            shared_gate = ggml_repeat(ctx0, shared_gate, ffn_shexp);
+            cb(shared_gate, "shared_expert_gate_broadcast", il);
+
+            // Apply the gate to the shared expert output
+            ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
+            cb(ffn_shexp, "ffn_shexp_gated", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+    } else {
+        // Dense FFN branch (not currently used I believe)
+        cur = build_ffn(cur,
+            model.layers[il].ffn_up, NULL, NULL,
+            model.layers[il].ffn_gate, NULL, NULL,
+            model.layers[il].ffn_down, NULL, NULL,
+            NULL,
+            LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+    }
+    return cur;
+}
diff --git a/llama/llama.cpp/src/models/qwen3vl-moe.cpp b/llama/llama.cpp/src/models/qwen3vl-moe.cpp
new file mode 100644
index 00000000..f72f80a8
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen3vl-moe.cpp
@@ -0,0 +1,149 @@
+#include "models.h"
+
+llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const size_t n_deepstack_layers = hparams.n_deepstack_layers;
+    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
+
+    if (ubatch.embd) {
+        // Image input: split main embd and deepstack embds
+        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
+        for (size_t i = 0; i < n_deepstack_layers; i++) {
+            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
+        }
+        inpL = inpL_main;
+    }
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_multi(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_multi(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+        cur = moe_out;
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
+            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+            cb(cur, "deepstack_out", il);
+        }
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
diff --git a/llama/llama.cpp/src/models/qwen3vl.cpp b/llama/llama.cpp/src/models/qwen3vl.cpp
new file mode 100644
index 00000000..0bae5223
--- /dev/null
+++ b/llama/llama.cpp/src/models/qwen3vl.cpp
@@ -0,0 +1,141 @@
+#include "models.h"
+
+llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const size_t n_deepstack_layers = hparams.n_deepstack_layers;
+    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
+
+    if (ubatch.embd) {
+        // Image input: split main embd and deepstack embds
+        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
+        for (size_t i = 0; i < n_deepstack_layers; i++) {
+            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
+        }
+        inpL = inpL_main;
+    }
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_multi(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_multi(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
+            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+            cb(cur, "deepstack_out", il);
+        }
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/refact.cpp b/llama/llama.cpp/src/models/refact.cpp
new file mode 100644
index 00000000..ff5eb284
--- /dev/null
+++ b/llama/llama.cpp/src/models/refact.cpp
@@ -0,0 +1,94 @@
+#include "models.h"
+
+llm_build_refact::llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/rnd1.cpp b/llama/llama.cpp/src/models/rnd1.cpp
new file mode 100644
index 00000000..46b3dc3e
--- /dev/null
+++ b/llama/llama.cpp/src/models/rnd1.cpp
@@ -0,0 +1,126 @@
+#include "models.h"
+
+// RND1 is a Qwen3Moe AR model converted to diffusion model.
+llm_build_rnd1::llm_build_rnd1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // Non-causal attention for diffusion
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+        cur = moe_out;
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/rwkv6-base.cpp b/llama/llama.cpp/src/models/rwkv6-base.cpp
new file mode 100644
index 00000000..7beed2da
--- /dev/null
+++ b/llama/llama.cpp/src/models/rwkv6-base.cpp
@@ -0,0 +1,162 @@
+#include "models.h"
+
+llm_build_rwkv6_base::llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params),
+    model(model) {}
+
+ggml_tensor * llm_build_rwkv6_base::build_rwkv6_channel_mix(const llama_layer * layer,
+                                                            ggml_tensor *       cur,
+                                                            ggml_tensor *       x_prev,
+                                                            llm_arch            arch) const {
+    ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);
+    switch (arch) {
+        case LLM_ARCH_RWKV6:
+            {
+                ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur);
+                ggml_tensor * xr = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_r), cur);
+
+                ggml_tensor * r = ggml_sigmoid(ctx0, build_lora_mm(layer->channel_mix_receptance, xr));
+                ggml_tensor * k = ggml_sqr(ctx0, ggml_relu(ctx0, build_lora_mm(layer->channel_mix_key, xk)));
+                cur             = ggml_mul(ctx0, r, build_lora_mm(layer->channel_mix_value, k));
+            }
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+    return cur;
+}
+
+ggml_tensor * llm_build_rwkv6_base::build_rwkv6_time_mix(llm_graph_input_rs * inp,
+                                                         ggml_tensor *        cur,
+                                                         ggml_tensor *        x_prev,
+                                                         const llama_ubatch & ubatch,
+                                                         int                  il) const {
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+
+    const auto n_tokens     = ubatch.n_tokens;
+    const auto n_seqs       = ubatch.n_seqs;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_embd       = hparams.n_embd;
+    const auto head_size    = hparams.wkv_head_size;
+    const auto n_head       = n_embd / head_size;
+    const auto n_head_kv    = hparams.n_head_kv(il);
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const auto & layer = model.layers[il];
+
+    bool is_qrwkv = layer.time_mix_first == nullptr;
+
+    ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);
+
+    sx  = ggml_reshape_2d(ctx0, sx, n_embd, n_tokens);
+    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
+
+    ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_x), cur);
+
+    xxx = ggml_reshape_4d(ctx0, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xxx)),
+                          layer.time_mix_w1->ne[1] / 5, 1, 5, n_tokens);
+
+    xxx = ggml_cont(ctx0, ggml_permute(ctx0, xxx, 0, 1, 3, 2));
+
+    xxx = ggml_mul_mat(
+        ctx0, ggml_reshape_4d(ctx0, layer.time_mix_w2, layer.time_mix_w2->ne[0], layer.time_mix_w2->ne[1], 1, 5), xxx);
+
+    ggml_tensor *xw, *xk, *xv, *xr, *xg;
+    if (layer.time_mix_lerp_fused) {
+        // fusing these weights makes some performance improvement
+        sx  = ggml_reshape_3d(ctx0, sx, n_embd, 1, n_tokens);
+        cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
+        xxx = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xxx, layer.time_mix_lerp_fused), sx), cur);
+        xw  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);
+        xk  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
+        xv  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
+        xr  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
+        xg  = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
+    } else {
+        // for backward compatibility
+        xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);
+        xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
+        xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
+        xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
+        xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
+
+        xw = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xw, layer.time_mix_lerp_w), sx), cur);
+        xk = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xk, layer.time_mix_lerp_k), sx), cur);
+        xv = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xv, layer.time_mix_lerp_v), sx), cur);
+        xr = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xr, layer.time_mix_lerp_r), sx), cur);
+        xg = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xg, layer.time_mix_lerp_g), sx), cur);
+    }
+    ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr);
+    ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk);
+    ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv);
+    if (layer.time_mix_receptance_b) {
+        r = ggml_add(ctx0, r, layer.time_mix_receptance_b);
+    }
+    if (layer.time_mix_key_b) {
+        k = ggml_add(ctx0, k, layer.time_mix_key_b);
+    }
+    if (layer.time_mix_value_b) {
+        v = ggml_add(ctx0, v, layer.time_mix_value_b);
+    }
+    ggml_tensor * g = build_lora_mm(layer.time_mix_gate, xg);
+    if (is_qrwkv) {
+        g = ggml_sigmoid(ctx0, g);
+    } else {
+        g = ggml_silu(ctx0, g);
+    }
+    if (n_head_kv != 0 && n_head_kv != n_head) {
+        GGML_ASSERT(n_head % n_head_kv == 0);
+        k                 = ggml_reshape_4d(ctx0, k, head_size, 1, n_head_kv, n_tokens);
+        v                 = ggml_reshape_4d(ctx0, v, head_size, 1, n_head_kv, n_tokens);
+        ggml_tensor * tmp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_size, n_head / n_head_kv, n_head_kv, n_tokens);
+        k                 = ggml_repeat(ctx0, k, tmp);
+        v                 = ggml_repeat(ctx0, v, tmp);
+    }
+    k = ggml_reshape_3d(ctx0, k, head_size, n_head, n_tokens);
+    v = ggml_reshape_3d(ctx0, v, head_size, n_head, n_tokens);
+    r = ggml_reshape_3d(ctx0, r, head_size, n_head, n_tokens);
+
+    ggml_tensor * w =
+        ggml_mul_mat(ctx0, layer.time_mix_decay_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_decay_w1, xw)));
+
+    w = ggml_add(ctx0, w, layer.time_mix_decay);
+    w = ggml_exp(ctx0, ggml_neg(ctx0, ggml_exp(ctx0, w)));
+    w = ggml_reshape_3d(ctx0, w, head_size, n_head, n_tokens);
+
+    if (is_qrwkv) {
+        // k = k * (1 - w)
+        k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w));
+    }
+    ggml_tensor * wkv_state = build_rs(inp, mctx_cur->get_s_l(il), hparams.n_embd_s(), n_seqs);
+
+    ggml_tensor * wkv_output;
+    if (is_qrwkv) {
+        wkv_output = ggml_gated_linear_attn(ctx0, k, v, r, w, wkv_state, pow(head_size, -0.5f));
+    } else {
+        wkv_output = ggml_rwkv_wkv6(ctx0, k, v, r, layer.time_mix_first, w, wkv_state);
+    }
+    cur       = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0);
+    wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));
+
+    ggml_build_forward_expand(
+        gf, ggml_cpy(ctx0, wkv_state,
+                     ggml_view_1d(ctx0, mctx_cur->get_s_l(il), hparams.n_embd_s() * n_seqs,
+                                  hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il)))));
+
+    if (!is_qrwkv) {
+        // group norm with head_count groups
+        cur = ggml_reshape_3d(ctx0, cur, n_embd / n_head, n_head, n_tokens);
+        cur = ggml_norm(ctx0, cur, 64e-5f);
+
+        // Convert back to regular vectors.
+        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b);
+    } else {
+        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
+    }
+    cur = ggml_mul(ctx0, cur, g);
+    cur = build_lora_mm(layer.time_mix_output, cur);
+
+    return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs);
+}
diff --git a/llama/llama.cpp/src/models/rwkv6.cpp b/llama/llama.cpp/src/models/rwkv6.cpp
new file mode 100644
index 00000000..15453fbf
--- /dev/null
+++ b/llama/llama.cpp/src/models/rwkv6.cpp
@@ -0,0 +1,94 @@
+#include "models.h"
+
+llm_build_rwkv6::llm_build_rwkv6(const llama_model & model, const llm_graph_params & params) :
+    llm_build_rwkv6_base(model, params) {
+    GGML_ASSERT(hparams.token_shift_count == 2);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd       = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs       = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL                      = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_shift =
+            ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
+        ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1],
+                                               token_shift->nb[2], n_embd * ggml_element_size(token_shift));
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+            ctx0, att_shift,
+            ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), 1);
+
+        cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il);
+        cb(ffn_norm, "ffn_norm", il);
+
+        x_prev = ggml_concat(
+            ctx0, ffn_shift,
+            ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), 1);
+
+        token_shift = ggml_concat(ctx0,
+                                  ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2],
+                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(att_norm)),
+                                  ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2],
+                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(ffn_norm)),
+                                  1);
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+        ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
+        x_prev   = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens);
+        cur      = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            ffn_inp  = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+            ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids);
+            x_prev   = ggml_get_rows(ctx0, x_prev, inp_out_ids);
+            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
+        }
+        cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6);
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) {
+            cur = ggml_scale(ctx0, cur, 0.5F);
+        }
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/rwkv6qwen2.cpp b/llama/llama.cpp/src/models/rwkv6qwen2.cpp
new file mode 100644
index 00000000..e84e5973
--- /dev/null
+++ b/llama/llama.cpp/src/models/rwkv6qwen2.cpp
@@ -0,0 +1,86 @@
+#include "models.h"
+
+llm_build_rwkv6qwen2::llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv6_base(model, params) {
+    GGML_ASSERT(n_embd == hparams.n_embd_r());
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+                ctx0,
+                token_shift,
+                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
+                1
+                );
+
+        cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il);
+
+        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/rwkv7-base.cpp b/llama/llama.cpp/src/models/rwkv7-base.cpp
new file mode 100644
index 00000000..cda44653
--- /dev/null
+++ b/llama/llama.cpp/src/models/rwkv7-base.cpp
@@ -0,0 +1,135 @@
+#include "models.h"
+
+llm_build_rwkv7_base::llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params),
+    model(model) {}
+
+ggml_tensor * llm_build_rwkv7_base::build_rwkv7_channel_mix(const llama_layer * layer,
+                                                            ggml_tensor *       cur,
+                                                            ggml_tensor *       x_prev,
+                                                            llm_arch            arch) const {
+    ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur);
+    switch (arch) {
+        case LLM_ARCH_RWKV7:
+            {
+                ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur);
+
+                ggml_tensor * k = ggml_sqr(ctx0, ggml_relu(ctx0, build_lora_mm(layer->channel_mix_key, xk)));
+
+                cur = build_lora_mm(layer->channel_mix_value, k);
+            }
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+    return cur;
+}
+
+ggml_tensor * llm_build_rwkv7_base::build_rwkv7_time_mix(llm_graph_input_rs * inp,
+                                                         ggml_tensor *        cur,
+                                                         ggml_tensor *        x_prev,
+                                                         ggml_tensor *&       first_layer_value,
+                                                         const llama_ubatch & ubatch,
+                                                         int                  il) const {
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+
+    const auto n_tokens     = ubatch.n_tokens;
+    const auto n_seqs       = ubatch.n_seqs;
+    const auto n_embd       = hparams.n_embd;
+    const auto head_size    = hparams.wkv_head_size;
+    const auto head_count   = n_embd / head_size;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const auto & layer = model.layers[il];
+
+    bool has_gating = layer.time_mix_g1 && layer.time_mix_g2;
+
+    ggml_tensor * sx    = ggml_sub(ctx0, x_prev, cur);
+    ggml_tensor * dummy = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_embd, n_seq_tokens, n_seqs, has_gating ? 6 : 5);
+    sx                  = ggml_repeat(ctx0, sx, dummy);
+
+    ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_fused), cur);
+
+    ggml_tensor * xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0);
+    ggml_tensor * xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
+    ggml_tensor * xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
+    ggml_tensor * xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
+    ggml_tensor * xa = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
+    ggml_tensor * xg =
+        has_gating ? ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 5 * sizeof(float)) :
+                     nullptr;
+
+    ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr);
+    ggml_tensor * w = ggml_add(
+        ctx0, ggml_mul_mat(ctx0, layer.time_mix_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xw))),
+        layer.time_mix_w0);
+    w = ggml_exp(ctx0, ggml_scale(ctx0, ggml_sigmoid(ctx0, w), -0.606531));
+
+    ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk);
+    ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv);
+    if (first_layer_value == nullptr) {
+        first_layer_value = v;
+    } else {
+        // Add the first layer value as a residual connection.
+        v = ggml_add(ctx0, v,
+                     ggml_mul(ctx0, ggml_sub(ctx0, first_layer_value, v),
+                              ggml_sigmoid(ctx0, ggml_add(ctx0,
+                                                          ggml_mul_mat(ctx0, layer.time_mix_v2,
+                                                                       ggml_mul_mat(ctx0, layer.time_mix_v1, xv)),
+                                                          layer.time_mix_v0))));
+    }
+    ggml_tensor * g = nullptr;
+    if (layer.time_mix_g1 && layer.time_mix_g2) {
+        g = ggml_mul_mat(ctx0, layer.time_mix_g2, ggml_sigmoid(ctx0, ggml_mul_mat(ctx0, layer.time_mix_g1, xg)));
+    }
+    ggml_tensor * a = ggml_sigmoid(
+        ctx0, ggml_add(ctx0, ggml_mul_mat(ctx0, layer.time_mix_a2, ggml_mul_mat(ctx0, layer.time_mix_a1, xa)),
+                       layer.time_mix_a0));
+
+    ggml_tensor * kk = ggml_reshape_3d(ctx0, ggml_mul(ctx0, k, layer.time_mix_k_k), head_size, head_count, n_tokens);
+    kk               = ggml_l2_norm(ctx0, kk, 1e-12);
+
+    ggml_tensor * ka = ggml_mul(ctx0, k, layer.time_mix_k_a);
+    k                = ggml_add(ctx0, k, ggml_sub(ctx0, ggml_mul(ctx0, a, ka), ka));
+
+    r = ggml_reshape_3d(ctx0, r, head_size, head_count, n_tokens);
+    w = ggml_reshape_3d(ctx0, w, head_size, head_count, n_tokens);
+    k = ggml_reshape_3d(ctx0, k, head_size, head_count, n_tokens);
+    v = ggml_reshape_3d(ctx0, v, head_size, head_count, n_tokens);
+    a = ggml_reshape_3d(ctx0, a, head_size, head_count, n_tokens);
+
+    ggml_tensor * wkv_state = build_rs(inp, mctx_cur->get_s_l(il), hparams.n_embd_s(), n_seqs);
+
+    ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx0, r, w, k, v, ggml_neg(ctx0, kk), ggml_mul(ctx0, kk, a), wkv_state);
+    cur                      = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0);
+    wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));
+
+    ggml_build_forward_expand(
+        gf, ggml_cpy(ctx0, wkv_state,
+                     ggml_view_1d(ctx0, mctx_cur->get_s_l(il), hparams.n_embd_s() * n_seqs,
+                                  hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il)))));
+
+    if (layer.time_mix_ln && layer.time_mix_ln_b) {
+        // group norm with head_count groups
+        cur = ggml_reshape_3d(ctx0, cur, n_embd / head_count, head_count, n_tokens);
+        cur = ggml_norm(ctx0, cur, 64e-5f);
+
+        // Convert back to regular vectors.
+        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b);
+    } else {
+        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
+    }
+    ggml_tensor * rk = ggml_sum_rows(
+        ctx0, ggml_mul(ctx0, ggml_mul(ctx0, k, r), ggml_reshape_2d(ctx0, layer.time_mix_r_k, head_size, head_count)));
+    cur = ggml_add(ctx0, cur, ggml_reshape_2d(ctx0, ggml_mul(ctx0, v, rk), n_embd, n_tokens));
+
+    if (has_gating) {
+        cur = ggml_mul(ctx0, cur, g);
+    }
+    cur = build_lora_mm(layer.time_mix_output, cur);
+
+    return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs);
+}
diff --git a/llama/llama.cpp/src/models/rwkv7.cpp b/llama/llama.cpp/src/models/rwkv7.cpp
new file mode 100644
index 00000000..5caf6553
--- /dev/null
+++ b/llama/llama.cpp/src/models/rwkv7.cpp
@@ -0,0 +1,90 @@
+#include "models.h"
+
+llm_build_rwkv7::llm_build_rwkv7(const llama_model & model, const llm_graph_params & params) :
+    llm_build_rwkv7_base(model, params) {
+    GGML_ASSERT(hparams.token_shift_count == 2);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * v_first = nullptr;
+
+    inpL = build_inp_embd(model.tok_embd);
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd       = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs       = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL                      = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_shift =
+            ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
+        ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1],
+                                               token_shift->nb[2], n_embd * ggml_element_size(token_shift));
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+            ctx0, att_shift,
+            ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), 1);
+
+        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il);
+        cb(ffn_norm, "ffn_norm", il);
+
+        x_prev = ggml_concat(
+            ctx0, ffn_shift,
+            ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), 1);
+
+        token_shift = ggml_concat(ctx0,
+                                  ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2],
+                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(att_norm)),
+                                  ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2],
+                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(ffn_norm)),
+                                  1);
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+        ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
+        x_prev   = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            ffn_inp  = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+            ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids);
+            x_prev   = ggml_get_rows(ctx0, x_prev, inp_out_ids);
+        }
+        cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7);
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/seed-oss.cpp b/llama/llama.cpp/src/models/seed-oss.cpp
new file mode 100644
index 00000000..0dc33c50
--- /dev/null
+++ b/llama/llama.cpp/src/models/seed-oss.cpp
@@ -0,0 +1,124 @@
+#include "models.h"
+
+llm_build_seed_oss::llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/smallthinker.cpp b/llama/llama.cpp/src/models/smallthinker.cpp
new file mode 100644
index 00000000..277eec29
--- /dev/null
+++ b/llama/llama.cpp/src/models/smallthinker.cpp
@@ -0,0 +1,120 @@
+#include "models.h"
+
+template <bool iswa>
+llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA  = inpL;
+        ggml_tensor * probs  = nullptr;
+
+        probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
+        cb(probs, "ffn_moe_logits", il);
+
+        // norm
+        cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (hparams.n_no_rope_layer_step == n_layer || il % hparams.n_no_rope_layer_step != 0) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            probs = ggml_get_rows(ctx0, probs, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * ffn_out =
+            build_moe_ffn(cur,
+                    nullptr,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_RELU, true,
+                    false, 0.0,
+                    static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
+                    il, probs);
+
+        cb(ffn_out, "ffn_out", il);
+        cur = ffn_out;
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_smallthinker<false>;
+template struct llm_build_smallthinker<true>;
diff --git a/llama/llama.cpp/src/models/smollm3.cpp b/llama/llama.cpp/src/models/smollm3.cpp
new file mode 100644
index 00000000..97c30dee
--- /dev/null
+++ b/llama/llama.cpp/src/models/smollm3.cpp
@@ -0,0 +1,128 @@
+#include "models.h"
+
+llm_build_smollm3::llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (use_rope) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/solar.cpp b/llama/llama.cpp/src/models/solar.cpp
new file mode 100644
index 00000000..97383928
--- /dev/null
+++ b/llama/llama.cpp/src/models/solar.cpp
@@ -0,0 +1,158 @@
+#include "models.h"
+
+llm_build_solar::llm_build_solar(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        auto * inp_attn = build_attn_inp_kv();
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+        struct ggml_tensor * bskcn_1;
+        struct ggml_tensor * bskcn_2;
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            if (hparams.n_bskcn(0, il)) {
+                bskcn_1 = inpSA;
+            }
+
+            if (hparams.n_bskcn(1, il)) {
+                bskcn_2 = inpSA;
+            }
+
+            if (hparams.n_bskcn(2, il)) {
+                inpSA = ggml_add(
+                   ctx0,
+                   ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
+                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
+            }
+
+            if (hparams.n_bskcn(3, il)) {
+                inpSA = ggml_add(
+                   ctx0,
+                   ggml_mul(ctx0, bskcn_2, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
+                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
+            }
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                   cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/stablelm.cpp b/llama/llama.cpp/src/models/stablelm.cpp
new file mode 100644
index 00000000..bed1915c
--- /dev/null
+++ b/llama/llama.cpp/src/models/stablelm.cpp
@@ -0,0 +1,146 @@
+#include "models.h"
+
+llm_build_stablelm::llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * inpSA = cur;
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm,
+                        NULL,
+                        LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm,
+                        NULL,
+                        LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpL  = ggml_get_rows(ctx0,  inpL, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            if (model.layers[il].ffn_norm) {
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, il);
+                cb(cur, "ffn_norm", il);
+            } else {
+                // parallel residual
+                cur = inpSA;
+            }
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/starcoder.cpp b/llama/llama.cpp/src/models/starcoder.cpp
new file mode 100644
index 00000000..e197af4a
--- /dev/null
+++ b/llama/llama.cpp/src/models/starcoder.cpp
@@ -0,0 +1,100 @@
+#include "models.h"
+
+llm_build_starcoder::llm_build_starcoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+    cb(pos, "pos_embd", -1);
+
+    inpL = ggml_add(ctx0, inpL, pos);
+    cb(inpL, "inpL", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/starcoder2.cpp b/llama/llama.cpp/src/models/starcoder2.cpp
new file mode 100644
index 00000000..e40ef2cb
--- /dev/null
+++ b/llama/llama.cpp/src/models/starcoder2.cpp
@@ -0,0 +1,121 @@
+#include "models.h"
+
+llm_build_starcoder2::llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                NULL,                      NULL,                        NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_GELU, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/t5-dec.cpp b/llama/llama.cpp/src/models/t5-dec.cpp
new file mode 100644
index 00000000..297e450d
--- /dev/null
+++ b/llama/llama.cpp/src/models/t5-dec.cpp
@@ -0,0 +1,166 @@
+#include "models.h"
+
+llm_build_t5_dec::llm_build_t5_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    //const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * embd_enc       = build_inp_cross_embd();
+    ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec();
+
+    const int64_t n_outputs_enc = embd_enc->ne[1];
+
+    auto * inp_attn_self  = build_attn_inp_kv();
+    auto * inp_attn_cross = build_attn_inp_cross();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const int64_t dec_n_layer = hparams.dec_n_layer;
+
+    for (int il = 0; il < dec_n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
+            ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b);
+
+            cur = build_attn(inp_attn_self,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
+            cb(cur, "kqv_out", il);
+        }
+        cur = ggml_add(ctx0, cur, inpSA);
+        cb(cur, "cross_inp", il);
+
+        ggml_tensor * inpCA = cur;
+
+        // norm
+        cur = build_norm(cur,
+                model.layers[il].attn_norm_cross, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm_cross", il);
+
+        // cross-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc);
+
+            cur = build_attn(inp_attn_cross,
+                    model.layers[il].wo_cross, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+            cb(cur, "kqv_out", il);
+
+            //ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+            //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+            //cb(kq, "kq", il);
+
+            //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
+            //cb(kq, "kq_soft_max_ext", il);
+
+            //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
+            //cb(v, "v", il);
+
+            //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
+            //cb(kqv, "kqv", il);
+
+            //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+            //cb(kqv_merged, "kqv_merged", il);
+
+            //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+            //cb(cur, "kqv_merged_cont", il);
+
+            //ggml_build_forward_expand(gf, cur);
+
+            //cur = build_lora_mm(model.layers[il].wo_cross, cur);
+            //cb(cur, "kqv_out", il);
+        }
+        if (il == dec_n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // T5 uses relu, flan-T5 uses gelu-gated
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU,
+                    model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ,
+                    il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cb(cur, "result_embd", -1);
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/t5-enc.cpp b/llama/llama.cpp/src/models/t5-enc.cpp
new file mode 100644
index 00000000..70e1d80d
--- /dev/null
+++ b/llama/llama.cpp/src/models/t5-enc.cpp
@@ -0,0 +1,96 @@
+#include "models.h"
+
+llm_build_t5_enc::llm_build_t5_enc(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm_enc, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
+            ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo_enc, nullptr,
+                    Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
+            cb(cur, "kqv_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm_enc, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // T5 uses relu, flan-T5 uses gelu-gated
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up_enc,   NULL, NULL,
+                    model.layers[il].ffn_gate_enc, NULL, NULL,
+                    model.layers[il].ffn_down_enc, NULL, NULL,
+                    NULL,
+                    model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
+                    model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
+                    il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cb(cur, "result_embd", -1);
+
+    cur = build_norm(cur,
+            model.output_norm_enc, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/wavtokenizer-dec.cpp b/llama/llama.cpp/src/models/wavtokenizer-dec.cpp
new file mode 100644
index 00000000..537a0d41
--- /dev/null
+++ b/llama/llama.cpp/src/models/wavtokenizer-dec.cpp
@@ -0,0 +1,149 @@
+#include "models.h"
+
+llm_build_wavtokenizer_dec::llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL));
+
+    cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1);
+    cur = ggml_add(ctx0, cur, model.conv1d_b);
+
+    // posnet
+    for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) {
+        const auto & layer = model.layers[il].posnet;
+
+        inpL = cur;
+
+        switch (il) {
+            case 0:
+            case 1:
+            case 3:
+            case 4:
+                {
+                    cur = build_norm(cur,
+                            layer.norm1,
+                            layer.norm1_b,
+                            LLM_NORM_GROUP, 0);
+
+                    cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur);
+
+                    cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1);
+                    cur = ggml_add(ctx0, cur, layer.conv1_b);
+
+                    cur = build_norm(cur,
+                            layer.norm2,
+                            layer.norm2_b,
+                            LLM_NORM_GROUP, 0);
+
+                    cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur);
+
+                    cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1);
+                    cur = ggml_add(ctx0, cur, layer.conv2_b);
+
+                    cur = ggml_add(ctx0, cur, inpL);
+                } break;
+            case 2:
+                {
+                    cur = build_norm(cur,
+                            layer.attn_norm,
+                            layer.attn_norm_b,
+                            LLM_NORM_GROUP, 0);
+
+                    ggml_tensor * q;
+                    ggml_tensor * k;
+                    ggml_tensor * v;
+
+                    q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1);
+                    k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1);
+                    v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1);
+
+                    q = ggml_add(ctx0, q, layer.attn_q_b);
+                    k = ggml_add(ctx0, k, layer.attn_k_b);
+                    v = ggml_add(ctx0, v, layer.attn_v_b);
+
+                    q = ggml_cont(ctx0, ggml_transpose(ctx0, q));
+                    k = ggml_cont(ctx0, ggml_transpose(ctx0, k));
+
+                    ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+
+                    kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f);
+
+                    cur = ggml_mul_mat(ctx0, kq, v);
+
+                    cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1);
+                    cur = ggml_add(ctx0, cur, layer.attn_o_b);
+
+                    cur = ggml_add(ctx0, cur, inpL);
+                } break;
+            case 5:
+                {
+                    cur = build_norm(cur,
+                            layer.norm,
+                            layer.norm_b,
+                            LLM_NORM_GROUP, 0);
+                } break;
+            default: GGML_ABORT("unknown posnet layer");
+        };
+    }
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+    cur = build_norm(cur,
+            model.tok_norm,
+            model.tok_norm_b,
+            LLM_NORM, -1);
+
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+    inpL = cur;
+
+    // convnext
+    for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) {
+        const auto & layer = model.layers[il].convnext;
+
+        cur = inpL;
+
+        cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1);
+        cur = ggml_add(ctx0, cur, layer.dw_b);
+
+        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+        cur = build_norm(cur,
+                layer.norm,
+                layer.norm_b,
+                LLM_NORM, -1);
+
+        cur = build_ffn(cur,
+                layer.pw1, layer.pw1_b, NULL,
+                NULL,      NULL,        NULL,
+                layer.pw2, layer.pw2_b, NULL,
+                NULL,
+                LLM_FFN_GELU, LLM_FFN_SEQ, il);
+
+        cur = ggml_mul(ctx0, cur, layer.gamma);
+
+        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+        inpL = ggml_add(ctx0, cur, inpL);
+    }
+    cur = inpL;
+
+    cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+    cur = build_norm(cur,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cur = ggml_add(ctx0, cur, model.output_b);
+
+    cb(cur, "result_embd", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/models/xverse.cpp b/llama/llama.cpp/src/models/xverse.cpp
new file mode 100644
index 00000000..364797dd
--- /dev/null
+++ b/llama/llama.cpp/src/models/xverse.cpp
@@ -0,0 +1,108 @@
+#include "models.h"
+
+llm_build_xverse::llm_build_xverse(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/llama/llama.cpp/src/unicode.cpp b/llama/llama.cpp/src/unicode.cpp
index ce336a22..040518e1 100644
--- a/llama/llama.cpp/src/unicode.cpp
+++ b/llama/llama.cpp/src/unicode.cpp
@@ -750,6 +750,80 @@ static std::vector<size_t> unicode_regex_split_custom_kimi_k2(const std::string
     return bpe_offsets;
 }
 
+// AFMOE digit handling: splits digits with leading 1-2 based on total length modulo 3
+static std::vector<size_t> unicode_regex_split_custom_afmoe(const std::string & text, const std::vector<size_t> & offsets) {
+    std::vector<size_t> bpe_offsets;
+    bpe_offsets.reserve(offsets.size());
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; ) {
+            const auto flags = _get_flags(pos);
+
+            // Handle digit sequences with special splitting logic
+            if (flags.is_number) {
+                size_t digit_start = pos;
+                size_t digit_count = 0;
+
+                // Count consecutive digits
+                while (_get_flags(pos).is_number && pos < offset_end) {
+                    digit_count++;
+                    pos++;
+                }
+
+                // Split based on total length modulo 3
+                size_t remainder = digit_count % 3;
+                size_t current = digit_start;
+
+                // Emit leading 1-2 digits if needed
+                if (remainder > 0) {
+                    _add_token(current + remainder);
+                    current += remainder;
+                }
+
+                // Emit groups of 3
+                while (current < digit_start + digit_count) {
+                    _add_token(current + 3);
+                    current += 3;
+                }
+                continue;
+            }
+
+            // For non-digits, just move forward
+            pos++;
+        }
+
+        // Add any remaining content
+        if (_prev_end < offset_end) {
+            _add_token(offset_end);
+        }
+    }
+
+    return bpe_offsets;
+}
+
 static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
     std::vector<size_t> bpe_offsets;
 
@@ -763,6 +837,9 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
     } else if (regex_expr == "\\p{Han}+") {
         // K2's first pattern - handle all K2 patterns together
         bpe_offsets = unicode_regex_split_custom_kimi_k2(text, offsets);
+    } else if (regex_expr == "\\p{AFMoE_digits}") {
+        // AFMOE digit pattern - use custom implementation for proper splitting
+        bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
     }
 
     return bpe_offsets;
diff --git a/llama/llama.cpp/tools/mtmd/clip-impl.h b/llama/llama.cpp/tools/mtmd/clip-impl.h
index 1669fad9..cd47865b 100644
--- a/llama/llama.cpp/tools/mtmd/clip-impl.h
+++ b/llama/llama.cpp/tools/mtmd/clip-impl.h
@@ -39,6 +39,7 @@
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
 #define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
 #define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
+#define KEY_IS_DEEPSTACK_LAYERS "clip.vision.is_deepstack_layers"
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -63,6 +64,7 @@
 #define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
 #define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
 #define TN_PATCH_BIAS      "v.patch_embd.bias"
+#define TN_ATTN_QKV        "%s.blk.%d.attn_qkv.%s"
 #define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
@@ -93,6 +95,9 @@
 #define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 #define TN_TOK_GLM_BOI     "adapter.boi"                // glm-edge (these embeddings are not in text model)
 #define TN_TOK_GLM_EOI     "adapter.eoi"                // glm-edge (these embeddings are not in text model)
+#define TN_DEEPSTACK_NORM  "v.deepstack.%d.norm.%s"     // qwen3vl deepstack
+#define TN_DEEPSTACK_FC1   "v.deepstack.%d.fc1.%s"      // qwen3vl deepstack
+#define TN_DEEPSTACK_FC2   "v.deepstack.%d.fc2.%s"      // qwen3vl deepstack
 
 // mimicpmv
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
@@ -116,6 +121,14 @@
 #define TN_MM_NORM_PRE  "mm.a.norm_pre.%s"
 #define TN_MM_NORM_MID  "mm.a.norm_mid.%s"
 
+// cogvlm
+#define TN_MM_POST_FC_NORM "mm.post_fc_norm.%s"
+#define TN_MM_H_TO_4H      "mm.up.%s"
+#define TN_MM_GATE         "mm.gate.%s"
+#define TN_MM_4H_TO_H      "mm.down.%s"
+#define TN_TOK_BOI         "v.boi"
+#define TN_TOK_EOI         "v.eoi"
+
 // align x to upper multiple of n
 #define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
 
@@ -127,6 +140,7 @@ enum projector_type {
     PROJECTOR_TYPE_MINICPMV,
     PROJECTOR_TYPE_GLM_EDGE,
     PROJECTOR_TYPE_QWEN2VL,
+    PROJECTOR_TYPE_QWEN3VL,
     PROJECTOR_TYPE_GEMMA3,
     PROJECTOR_TYPE_IDEFICS3,
     PROJECTOR_TYPE_PIXTRAL,
@@ -139,6 +153,9 @@ enum projector_type {
     PROJECTOR_TYPE_VOXTRAL,
     PROJECTOR_TYPE_LFM2,
     PROJECTOR_TYPE_KIMIVL,
+    PROJECTOR_TYPE_LIGHTONOCR,
+    PROJECTOR_TYPE_COGVLM,
+    PROJECTOR_TYPE_JANUS_PRO,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -150,6 +167,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
     { PROJECTOR_TYPE_QWEN2VL,   "qwen2vl_merger"},
     { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
+    { PROJECTOR_TYPE_QWEN3VL,   "qwen3vl_merger"},
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
     { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
     { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
@@ -161,6 +179,9 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
     { PROJECTOR_TYPE_LFM2,      "lfm2"},
     { PROJECTOR_TYPE_KIMIVL,    "kimivl"},
+    { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},
+    { PROJECTOR_TYPE_COGVLM,    "cogvlm"},
+    { PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -203,7 +224,6 @@ static void clip_log_callback_default(enum ggml_log_level level, const char * te
 }
 
 struct clip_logger_state {
-    ggml_log_level verbosity_thold;
     ggml_log_callback log_callback;
     void * log_callback_user_data;
 };
@@ -237,17 +257,11 @@ static void clip_log_internal(enum ggml_log_level level, const char * format, ..
     va_end(args);
 }
 
-#define LOG_TMPL(level, ...) \
-    do { \
-        if ((level) >= g_logger_state.verbosity_thold) { \
-            clip_log_internal((level), __VA_ARGS__); \
-        } \
-    } while (0)
-#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  __VA_ARGS__)
+#define LOG_INF(...) clip_log_internal(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define LOG_WRN(...) clip_log_internal(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
+#define LOG_ERR(...) clip_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LOG_DBG(...) clip_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LOG_CNT(...) clip_log_internal(GGML_LOG_LEVEL_CONT,  __VA_ARGS__)
 
 //
 // cpp wrappers
diff --git a/llama/llama.cpp/tools/mtmd/clip.cpp b/llama/llama.cpp/tools/mtmd/clip.cpp
index c984e628..3334ff25 100644
--- a/llama/llama.cpp/tools/mtmd/clip.cpp
+++ b/llama/llama.cpp/tools/mtmd/clip.cpp
@@ -6,7 +6,6 @@
 #include "clip-impl.h"
 #include "ggml.h"
 #include "ggml-cpp.h"
-#include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 #include "gguf.h"
@@ -17,15 +16,12 @@
 #include <cstring>
 #include <fstream>
 #include <map>
-#include <regex>
 #include <stdexcept>
 #include <unordered_set>
 #include <vector>
-#include <sstream>
 #include <cinttypes>
 #include <limits>
 #include <array>
-#include <numeric>
 #include <functional>
 
 #if defined(_WIN32)
@@ -41,7 +37,7 @@
 #endif
 #endif
 
-struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
+struct clip_logger_state g_logger_state = {clip_log_callback_default, NULL};
 
 enum ffn_op_type {
     FFN_GELU,
@@ -176,16 +172,18 @@ enum patch_merge_type {
 };
 
 struct clip_hparams {
-    int32_t image_size;
-    int32_t patch_size;
-    int32_t n_embd;
-    int32_t n_ff;
-    int32_t projection_dim;
-    int32_t n_head;
-    int32_t n_layer;
+    int32_t image_size = 0;
+    int32_t patch_size = 0;
+    int32_t n_embd = 0;
+    int32_t n_ff = 0;
+    int32_t projection_dim = 0;
+    int32_t n_head = 0;
+    int32_t n_layer = 0;
     // idefics3
-    int32_t preproc_image_size = 0;
-    int32_t proj_scale_factor = 0;
+    int32_t image_longest_edge = 0;
+    int32_t image_min_pixels = -1;
+    int32_t image_max_pixels = -1;
+    int32_t n_merge = 0; // number of patch merges **per-side**
 
     float image_mean[3];
     float image_std[3];
@@ -207,7 +205,6 @@ struct clip_hparams {
     std::unordered_set<int32_t> vision_feature_layer;
     int32_t attn_window_size = 0;
     int32_t n_wa_pattern = 0;
-    int32_t spatial_merge_size = 0;
 
     // audio
     int32_t n_mel_bins = 0; // whisper preprocessor
@@ -217,6 +214,26 @@ struct clip_hparams {
     bool has_llava_projector = false;
     int minicpmv_version = 0;
     int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
+
+    // custom value provided by user, can be undefined if not set
+    int32_t custom_image_min_tokens = -1;
+    int32_t custom_image_max_tokens = -1;
+
+    void set_limit_image_tokens(int n_tokens_min, int n_tokens_max) {
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        const int patch_area = patch_size * patch_size * cur_merge * cur_merge;
+        image_min_pixels = (custom_image_min_tokens > 0 ? custom_image_min_tokens : n_tokens_min) * patch_area;
+        image_max_pixels = (custom_image_max_tokens > 0 ? custom_image_max_tokens : n_tokens_max) * patch_area;
+        warmup_image_size = static_cast<int>(std::sqrt(image_max_pixels));
+    }
+
+    void set_warmup_n_tokens(int n_tokens) {
+        int n_tok_per_side = static_cast<int>(std::sqrt(n_tokens));
+        GGML_ASSERT(n_tok_per_side * n_tok_per_side == n_tokens && "n_tokens must be n*n");
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        warmup_image_size = n_tok_per_side * patch_size * cur_merge;
+        // TODO: support warmup size for custom token numbers
+    }
 };
 
 struct clip_layer {
@@ -227,6 +244,8 @@ struct clip_layer {
     ggml_tensor * q_b = nullptr;
     ggml_tensor * v_w = nullptr;
     ggml_tensor * v_b = nullptr;
+    ggml_tensor * qkv_w = nullptr;
+    ggml_tensor * qkv_b = nullptr;
 
     ggml_tensor * o_w = nullptr;
     ggml_tensor * o_b = nullptr;
@@ -252,6 +271,18 @@ struct clip_layer {
     // layer scale (no bias)
     ggml_tensor * ls_1_w = nullptr;
     ggml_tensor * ls_2_w = nullptr;
+
+    // qwen3vl deepstack merger
+    ggml_tensor * deepstack_norm_w = nullptr;
+    ggml_tensor * deepstack_norm_b = nullptr;
+    ggml_tensor * deepstack_fc1_w = nullptr;
+    ggml_tensor * deepstack_fc1_b = nullptr;
+    ggml_tensor * deepstack_fc2_w = nullptr;
+    ggml_tensor * deepstack_fc2_b = nullptr;
+
+    bool has_deepstack() const {
+        return deepstack_fc1_w != nullptr;
+    }
 };
 
 struct clip_model {
@@ -271,6 +302,8 @@ struct clip_model {
 
     std::vector<clip_layer> layers;
 
+    int32_t n_deepstack_layers = 0; // used by Qwen3-VL, calculated from clip_layer
+
     ggml_tensor * post_ln_w;
     ggml_tensor * post_ln_b;
 
@@ -299,8 +332,6 @@ struct clip_model {
     // GLMV-Edge projection
     ggml_tensor * mm_model_adapter_conv_w = nullptr;
     ggml_tensor * mm_model_adapter_conv_b = nullptr;
-    ggml_tensor * mm_glm_tok_boi = nullptr;
-    ggml_tensor * mm_glm_tok_eoi = nullptr;
 
     // MobileVLM projection
     ggml_tensor * mm_model_mlp_1_w = nullptr;
@@ -372,6 +403,15 @@ struct clip_model {
     ggml_tensor * mm_norm_pre_w = nullptr;
     ggml_tensor * mm_norm_mid_w = nullptr;
 
+    // cogvlm
+    ggml_tensor * mm_post_fc_norm_w = nullptr;
+    ggml_tensor * mm_post_fc_norm_b = nullptr;
+    ggml_tensor * mm_h_to_4h_w = nullptr;
+    ggml_tensor * mm_gate_w = nullptr;
+    ggml_tensor * mm_4h_to_h_w = nullptr;
+    ggml_tensor * mm_boi = nullptr;
+    ggml_tensor * mm_eoi = nullptr;
+
     bool audio_has_avgpool() const {
         return proj_type == PROJECTOR_TYPE_QWEN2A
             || proj_type == PROJECTOR_TYPE_VOXTRAL;
@@ -400,12 +440,14 @@ struct clip_ctx {
 
     int max_nodes = 8192;
     ggml_backend_sched_ptr sched;
+    clip_flash_attn_type flash_attn_type = CLIP_FLASH_ATTN_TYPE_AUTO;
 
     // for debugging
     bool debug_graph = false;
     std::vector<ggml_tensor *> debug_print_tensors;
 
     clip_ctx(clip_context_params & ctx_params) {
+        flash_attn_type = ctx_params.flash_attn_type;
         debug_graph = std::getenv("MTMD_DEBUG_GRAPH") != nullptr;
         backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
         if (!backend_cpu) {
@@ -434,6 +476,13 @@ struct clip_ctx {
             LOG_INF("%s: CLIP using CPU backend\n", __func__);
         }
 
+        if (ctx_params.image_min_tokens > 0) {
+            model.hparams.custom_image_min_tokens = ctx_params.image_min_tokens;
+        }
+        if (ctx_params.image_max_tokens > 0) {
+            model.hparams.custom_image_max_tokens = ctx_params.image_max_tokens;
+        }
+
         backend_ptrs.push_back(backend_cpu);
         backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu));
 
@@ -522,7 +571,7 @@ struct clip_graph {
             const int batch_size = 1;
             GGML_ASSERT(n_patches_x == n_patches_y);
             const int patches_per_image = n_patches_x;
-            const int kernel_size = hparams.proj_scale_factor;
+            const int kernel_size = hparams.n_merge;
 
             cur = ggml_transpose(ctx0, cur);
             cur = ggml_cont_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
@@ -544,13 +593,13 @@ struct clip_graph {
         } else if (ctx->proj_type() == PROJECTOR_TYPE_IDEFICS3) {
             // pixel_shuffle
             // https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578
-            const int scale_factor = model.hparams.proj_scale_factor;
+            const int scale_factor = model.hparams.n_merge;
             cur = build_patch_merge_permute(cur, scale_factor);
             cur = ggml_mul_mat(ctx0, model.projection, cur);
 
         } else if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
             // pixel unshuffle block
-            const int scale_factor = model.hparams.proj_scale_factor;
+            const int scale_factor = model.hparams.n_merge;
             cur = build_patch_merge_permute(cur, scale_factor);
 
             // projection
@@ -563,6 +612,15 @@ struct clip_graph {
             cur = ggml_gelu(ctx0, cur);
             cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
             cur = ggml_add(ctx0, cur, model.mm_2_b);
+
+        } else if (ctx->proj_type() == PROJECTOR_TYPE_JANUS_PRO) {
+            cur = build_ffn(cur,
+                model.mm_0_w, model.mm_0_b,
+                nullptr, nullptr,
+                model.mm_1_w, model.mm_1_b,
+                hparams.ffn_op,
+                -1);
+
         } else {
             GGML_ABORT("SigLIP: Unsupported projector type");
         }
@@ -574,7 +632,7 @@ struct clip_graph {
     }
 
     ggml_cgraph * build_pixtral() {
-        const int n_merge = hparams.spatial_merge_size;
+        const int n_merge = hparams.n_merge;
 
         // 2D input positions
         ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
@@ -600,7 +658,7 @@ struct clip_graph {
         // mistral small 3.1 patch merger
         // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67
         if (model.mm_patch_merger_w) {
-            GGML_ASSERT(hparams.spatial_merge_size > 0);
+            GGML_ASSERT(hparams.n_merge > 0);
 
             cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w);
 
@@ -634,7 +692,7 @@ struct clip_graph {
         }
 
         // arrangement of the [IMG_BREAK] token
-        {
+        if (model.token_embd_img_break) {
             // not efficient, but works
             // the trick is to view the embeddings as a 3D tensor with shape [n_embd, n_patches_per_row, n_rows]
             // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
@@ -727,6 +785,15 @@ struct clip_graph {
             ggml_set_name(window_mask, "window_mask");
             ggml_set_input(window_mask);
 
+            // if flash attn is used, we need to pad the mask and cast to f16
+            if (ctx->flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
+                int n_pad = GGML_PAD(window_mask->ne[1], GGML_KQ_MASK_PAD) - window_mask->ne[1];
+                if (n_pad > 0) {
+                    window_mask = ggml_pad(ctx0, window_mask, 0, n_pad, 0, 0);
+                }
+                window_mask = ggml_cast(ctx0, window_mask, GGML_TYPE_F16);
+            }
+
             // inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size]
             GGML_ASSERT(batch_size == 1);
             inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4);
@@ -844,17 +911,216 @@ struct clip_graph {
         return gf;
     }
 
-    ggml_cgraph * build_minicpmv() {
-        const int batch_size = 1;
-
+    // Qwen3VL
+    ggml_cgraph * build_qwen3vl() {
+        GGML_ASSERT(model.patch_bias != nullptr);
+        GGML_ASSERT(model.position_embeddings != nullptr);
         GGML_ASSERT(model.class_embedding == nullptr);
-        const int n_pos = n_patches;
+
+        const int batch_size       = 1;
+        const int n_pos            = n_patches;
+        const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
+
+        norm_type norm_t = NORM_TYPE_NORMAL;
+
+        int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+
+        ggml_tensor * inp_raw = build_inp_raw();
+        ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+        GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+        GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+        // second conv dimension
+        {
+            auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+            inp = ggml_add(ctx0, inp, inp_1);
+
+            inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+            inp = ggml_cont_4d(
+                ctx0, inp,
+                n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+            inp = ggml_reshape_4d(
+                ctx0, inp,
+                n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+            inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+            inp = ggml_cont_3d(
+                ctx0, inp,
+                n_embd, n_patches_x * n_patches_y, batch_size);
+        }
+
+        // add patch bias
+        if (model.patch_bias != nullptr) {
+            inp = ggml_add(ctx0, inp, model.patch_bias);
+            cb(inp, "patch_bias", -1);
+        }
+
+        // calculate absolute position embedding and apply
+        ggml_tensor * learned_pos_embd = resize_position_embeddings();
+        learned_pos_embd = ggml_cont_4d(
+            ctx0, learned_pos_embd,
+            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+        learned_pos_embd = ggml_reshape_4d(
+            ctx0, learned_pos_embd,
+            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+        learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
+        learned_pos_embd = ggml_cont_3d(
+            ctx0, learned_pos_embd,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+        inp = ggml_add(ctx0, inp, learned_pos_embd);
+        cb(inp, "inp_pos_emb", -1);
+
+        ggml_tensor * inpL = inp;
+
+        ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
+        ggml_set_name(positions, "positions");
+        ggml_set_input(positions);
+
+        // pre-layernorm
+        if (model.pre_ln_w) {
+            inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
+        }
+
+        // deepstack features (stack along the feature dimension), [n_embd * len(deepstack_layers), n_patches_x * n_patches_y, batch_size]
+        ggml_tensor * deepstack_features = nullptr;
+        const int merge_factor = hparams.n_merge > 0 ? hparams.n_merge * hparams.n_merge : 4; // default 2x2=4 for qwen3vl
+
+        // loop over layers
+        for (int il = 0; il < n_layer; il++) {
+            auto & layer = model.layers[il];
+
+            ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+            // layernorm1
+            cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
+            cb(cur, "ln1", il);
+
+            // self-attention
+            {
+                cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
+                cur = ggml_add(ctx0, cur, layer.qkv_b);
+
+                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                        /* nb1    */ ggml_row_size(cur->type, d_head),
+                        /* nb2    */ cur->nb[1],
+                        /* offset */ 0);
+
+                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                        /* nb1    */ ggml_row_size(cur->type, d_head),
+                        /* nb2    */ cur->nb[1],
+                        /* offset */ ggml_row_size(cur->type, n_embd));
+
+                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                        /* nb1    */ ggml_row_size(cur->type, d_head),
+                        /* nb2    */ cur->nb[1],
+                        /* offset */ ggml_row_size(cur->type, 2 * n_embd));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                // apply M-RoPE
+                Qcur = ggml_rope_multi(
+                    ctx0, Qcur, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+                Kcur = ggml_rope_multi(
+                    ctx0, Kcur, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+
+                cb(Qcur, "Qcur_rope", il);
+                cb(Kcur, "Kcur_rope", il);
+
+                cur = build_attn(layer.o_w, layer.o_b,
+                    Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            // re-add the layer input, e.g., residual
+            cur = ggml_add(ctx0, cur, inpL);
+
+            inpL = cur; // inpL = residual, cur = hidden_states
+
+            cb(cur, "ffn_inp", il);
+
+            // layernorm2
+            cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
+            cb(cur, "ffn_inp_normed", il);
+
+            // ffn
+            cur = build_ffn(cur,
+                layer.ff_up_w, layer.ff_up_b,
+                layer.ff_gate_w, layer.ff_gate_b,
+                layer.ff_down_w, layer.ff_down_b,
+                hparams.ffn_op, il);
+
+            cb(cur, "ffn_out", il);
+
+            // residual 2
+            cur = ggml_add(ctx0, inpL, cur);
+            cb(cur, "layer_out", il);
+
+            if (layer.has_deepstack()) {
+                ggml_tensor * feat = ggml_reshape_3d(ctx0, cur, n_embd * merge_factor, n_pos / merge_factor, batch_size);
+                feat = build_norm(feat, layer.deepstack_norm_w, layer.deepstack_norm_b, norm_t, eps, il);
+                feat = build_ffn(feat,
+                    layer.deepstack_fc1_w, layer.deepstack_fc1_b,
+                    nullptr, nullptr,
+                    layer.deepstack_fc2_w, layer.deepstack_fc2_b,
+                    ffn_op_type::FFN_GELU, il);
+
+                if(!deepstack_features) {
+                    deepstack_features = feat;
+                } else {
+                    // concat along the feature dimension
+                    deepstack_features = ggml_concat(ctx0, deepstack_features, feat, 0);
+                }
+            }
+
+            inpL = cur;
+        }
+
+        // post-layernorm
+        if (model.post_ln_w) {
+            inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
+        }
+
+        // multimodal projection
+        ggml_tensor * embeddings = inpL;
+        embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
+
+        embeddings = build_ffn(embeddings,
+            model.mm_0_w, model.mm_0_b,
+            nullptr, nullptr,
+            model.mm_1_w, model.mm_1_b,
+            ffn_op_type::FFN_GELU, -1);
+
+        embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension
+
+        // build the graph
+        ggml_build_forward_expand(gf, embeddings);
+
+        return gf;
+    }
+
+    ggml_cgraph * build_minicpmv() {
+        GGML_ASSERT(model.class_embedding == nullptr);
+        const int n_pos       = n_patches;
+        const int n_embd_proj = clip_n_mmproj_embd(ctx);
 
         // position embeddings for the projector (not for ViT)
-        int n_output_dim = clip_n_mmproj_embd(ctx);
-        ggml_tensor * pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_output_dim, n_pos, batch_size);
-        ggml_set_name(pos_embed, "pos_embed");
-        ggml_set_input(pos_embed);
+        // see: https://huggingface.co/openbmb/MiniCPM-o-2_6/blob/main/resampler.py#L70
+        // base frequency omega
+        ggml_tensor * omega = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_embd_proj / 4);
+        ggml_set_name(omega, "omega");
+        ggml_set_input(omega);
+
+        // 2D input positions (using float for sinusoidal embeddings)
+        ggml_tensor * pos_h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+        ggml_set_name(pos_h, "pos_h");
+        ggml_set_input(pos_h);
+        ggml_tensor * pos_w = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+        ggml_set_name(pos_w, "pos_w");
+        ggml_set_input(pos_w);
 
         // for selecting learned pos embd, used by ViT
         struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
@@ -865,7 +1131,7 @@ struct clip_graph {
 
         ggml_tensor * inp = build_inp();
         ggml_tensor * embeddings = build_vit(
-                                inp, n_patches,
+                                inp, n_pos,
                                 NORM_TYPE_NORMAL,
                                 hparams.ffn_op,
                                 learned_pos_embd,
@@ -877,17 +1143,39 @@ struct clip_graph {
         ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
 
         // norm
-        q = build_norm(q, model.mm_model_ln_q_w, model.mm_model_ln_q_b, NORM_TYPE_NORMAL, eps, -1);
+        q = build_norm(q, model.mm_model_ln_q_w,  model.mm_model_ln_q_b,  NORM_TYPE_NORMAL, eps, -1);
         v = build_norm(v, model.mm_model_ln_kv_w, model.mm_model_ln_kv_b, NORM_TYPE_NORMAL, eps, -1);
 
+        // calculate sinusoidal pos embd
+        ggml_tensor * pos_embed = nullptr;
+        {
+            // outer product
+            ggml_tensor * omega_b = ggml_repeat_4d(ctx0, omega, omega->ne[0], n_pos, 1, 1); // n_pos rows
+            ggml_tensor * theta_x = ggml_mul(ctx0, omega_b, pos_w);
+            ggml_tensor * theta_y = ggml_mul(ctx0, omega_b, pos_h);
+            // sin and cos
+            ggml_tensor * pos_embd_x = ggml_concat(
+                ctx0,
+                ggml_sin(ctx0, theta_x),
+                ggml_cos(ctx0, theta_x),
+                0 // concat on first dim
+            );
+            ggml_tensor * pos_embd_y = ggml_concat(
+                ctx0,
+                ggml_sin(ctx0, theta_y),
+                ggml_cos(ctx0, theta_y),
+                0 // concat on first dim
+            );
+            pos_embed = ggml_concat(ctx0, pos_embd_x, pos_embd_y, 0);
+        }
+
         // k = v + pos_embed
         ggml_tensor * k = ggml_add(ctx0, v, pos_embed);
 
         // attention
         {
-            int n_embd = clip_n_mmproj_embd(ctx);
             const int d_head = 128;
-            int n_head = n_embd/d_head;
+            int n_head = n_embd_proj/d_head;
             // Use actual config value if available, otherwise fall back to hardcoded values
             int num_query = ctx->model.hparams.minicpmv_query_num;
             ggml_tensor * Q = ggml_add(ctx0,
@@ -908,10 +1196,11 @@ struct clip_graph {
             cb(K, "resampler_K", -1);
             cb(V, "resampler_V", -1);
 
+            float resampler_kq_scale = 1.0f/ sqrtf(float(d_head));
             embeddings = build_attn(
                 model.mm_model_attn_o_w,
                 model.mm_model_attn_o_b,
-                Q, K, V, nullptr, kq_scale, -1);
+                Q, K, V, nullptr, resampler_kq_scale, -1);
             cb(embeddings, "resampler_attn_out", -1);
         }
         // layernorm
@@ -956,7 +1245,7 @@ struct clip_graph {
 
         // pixel shuffle
         {
-            const int scale_factor = model.hparams.proj_scale_factor;
+            const int scale_factor = model.hparams.n_merge;
             const int bsz    = 1; // batch size, always 1 for now since we don't support batching
             const int height = n_patches_y;
             const int width  = n_patches_x;
@@ -1046,7 +1335,7 @@ struct clip_graph {
         // based on Llama4VisionPixelShuffleMLP
         // https://github.com/huggingface/transformers/blob/2932f318a20d9e54cc7aea052e040164d85de7d6/src/transformers/models/llama4/modeling_llama4.py#L1151
         {
-            const int scale_factor = model.hparams.proj_scale_factor;
+            const int scale_factor = model.hparams.n_merge;
             const int bsz = 1; // batch size, always 1 for now since we don't support batching
             GGML_ASSERT(scale_factor > 0);
             GGML_ASSERT(n_patches_x == n_patches_y); // llama4 only supports square images
@@ -1118,7 +1407,7 @@ struct clip_graph {
 
         {
             // patch_merger
-            const int scale_factor = model.hparams.proj_scale_factor;
+            const int scale_factor = model.hparams.n_merge;
             cur = build_patch_merge_permute(cur, scale_factor);
 
             // projection norm
@@ -1507,8 +1796,8 @@ struct clip_graph {
             // note: these embeddings are not present in text model, hence we cannot process them as text tokens
             // see: https://huggingface.co/THUDM/glm-edge-v-2b/blob/main/siglip.py#L53
             {
-                embeddings = ggml_concat(ctx0, model.mm_glm_tok_boi, embeddings, 1); // BOI
-                embeddings = ggml_concat(ctx0, embeddings, model.mm_glm_tok_eoi, 1); // EOI
+                embeddings = ggml_concat(ctx0, model.mm_boi, embeddings, 1); // BOI
+                embeddings = ggml_concat(ctx0, embeddings, model.mm_eoi, 1); // EOI
             }
         }
 
@@ -1521,7 +1810,6 @@ struct clip_graph {
 
         return gf;
     }
-
     // whisper encoder with custom projector
     ggml_cgraph * build_whisper_enc() {
         const int n_frames = img.nx;
@@ -1626,6 +1914,104 @@ struct clip_graph {
         return gf;
     }
 
+    // cogvlm vision encoder
+    ggml_cgraph * build_cogvlm() {
+        GGML_ASSERT(model.class_embedding != nullptr);
+        GGML_ASSERT(model.position_embeddings != nullptr);
+
+        const int n_pos = n_patches + 1; // +1 for [CLS]
+
+        // build input and concatenate class embedding
+        ggml_tensor * inp = build_inp();
+        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+        inp = ggml_add(ctx0, inp, model.position_embeddings);
+        cb(inp, "inp_pos", -1);
+
+        ggml_tensor * inpL = inp;
+
+        for (int il = 0; il < n_layer; il++) {
+            auto & layer = model.layers[il];
+            ggml_tensor * cur = inpL;
+
+            cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
+
+            cur = ggml_add(ctx0, cur, layer.qkv_b);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+                cur->nb[1], 0);
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+                cur->nb[1], n_embd * sizeof(float));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+                cur->nb[1], 2 * n_embd * sizeof(float));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+
+            cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+            cb(cur, "attn_post_norm", il);
+
+            cur = ggml_add(ctx0, cur, inpL);
+            inpL = cur;
+
+            cur = build_ffn(cur,
+                layer.ff_up_w, layer.ff_up_b,
+                layer.ff_gate_w, layer.ff_gate_b,
+                layer.ff_down_w, layer.ff_down_b,
+                hparams.ffn_op, il);
+
+            cb(cur, "ffn_out", il);
+
+            cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+            cb(cur, "ffn_post_norm", il);
+
+            cur = ggml_add(ctx0, cur, inpL);
+            cb(cur, "layer_out", il);
+            inpL = cur;
+
+        }
+
+        // remove CLS token (like build_llama4 does)
+        ggml_tensor * cur = ggml_view_2d(ctx0, inpL,
+            n_embd, n_patches,
+            ggml_row_size(inpL->type, n_embd), 0);
+
+        // Multiply with mm_model_proj
+        cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
+
+        // Apply layernorm, weight, bias
+        cur = build_norm(cur, model.mm_post_fc_norm_w, model.mm_post_fc_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);
+
+        // Apply GELU
+        cur = ggml_gelu_inplace(ctx0, cur);
+
+        // Branch 1: multiply with mm_h_to_4h_w
+        ggml_tensor * h_to_4h = ggml_mul_mat(ctx0, model.mm_h_to_4h_w, cur);
+
+        // Branch 2: multiply with mm_gate_w
+        ggml_tensor * gate = ggml_mul_mat(ctx0, model.mm_gate_w, cur);
+
+        // Apply silu
+        gate = ggml_swiglu_split(ctx0, gate, h_to_4h);
+
+        // Apply mm_4h_to_h_w
+        cur = ggml_mul_mat(ctx0, model.mm_4h_to_h_w, gate);
+
+        // Concatenate with boi and eoi
+        cur = ggml_concat(ctx0, model.mm_boi, cur, 1);
+        cur = ggml_concat(ctx0, cur, model.mm_eoi, 1);
+
+        // build the graph
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
 private:
     //
     // utility functions
@@ -1953,17 +2339,25 @@ private:
         ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
         //cb(k, "k", il);
 
-        ggml_tensor * v = ggml_permute(ctx0, v_cur, 1, 2, 0, 3);
-        v = ggml_cont(ctx0, v);
-        //cb(k, "v", il);
-
         ggml_tensor * cur;
 
-        // TODO @ngxson : support flash attention
-        {
+        if (ctx->flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
+            ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
+
+            k = ggml_cast(ctx0, k, GGML_TYPE_F16);
+            v = ggml_cast(ctx0, v, GGML_TYPE_F16);
+
+            cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, 0.0f, 0.0f);
+            ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
+
+            cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
+
+        } else {
+            ggml_tensor * v = ggml_permute(ctx0, v_cur, 1, 2, 0, 3);
+            v = ggml_cont(ctx0, v);
+
             const auto n_tokens = q->ne[1];
             const auto n_head   = q->ne[2];
-            // const auto n_kv     = k->ne[1]; // for flash attention
 
             ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
             // F32 may not needed for vision encoders?
@@ -2108,6 +2502,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
                 res = graph.build_siglip();
             } break;
         case PROJECTOR_TYPE_PIXTRAL:
+        case PROJECTOR_TYPE_LIGHTONOCR:
             {
                 res = graph.build_pixtral();
             } break;
@@ -2116,6 +2511,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 res = graph.build_qwen2vl();
             } break;
+        case PROJECTOR_TYPE_QWEN3VL:
+            {
+                res = graph.build_qwen3vl();
+            } break;
         case PROJECTOR_TYPE_MINICPMV:
             {
                 res = graph.build_minicpmv();
@@ -2138,6 +2537,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 res = graph.build_kimivl();
             } break;
+        case PROJECTOR_TYPE_JANUS_PRO:
+            {
+                res = graph.build_siglip();
+            } break;
+        case PROJECTOR_TYPE_COGVLM:
+            {
+                res = graph.build_cogvlm();
+            } break;
         default:
             {
                 res = graph.build_llava();
@@ -2241,6 +2648,10 @@ struct clip_model_loader {
             if (proj_type.empty()) {
                 if (modality == CLIP_MODALITY_VISION) {
                     get_string(KEY_VISION_PROJ_TYPE, proj_type, false);
+                    if (proj_type.empty()) {
+                        // Assume MLP if no projector type listed
+                        proj_type = "mlp";
+                    }
                 } else if (modality == CLIP_MODALITY_AUDIO) {
                     get_string(KEY_AUDIO_PROJ_TYPE, proj_type, false);
                 } else {
@@ -2277,7 +2688,6 @@ struct clip_model_loader {
 
             if (is_vision) {
                 get_u32(KEY_IMAGE_SIZE, hparams.image_size);
-                get_u32(KEY_PREPROC_IMAGE_SIZE, hparams.preproc_image_size, false);
                 get_u32(KEY_PATCH_SIZE, hparams.patch_size);
                 get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
                 get_i32(KEY_MINICPMV_VERSION, hparams.minicpmv_version, false); // legacy
@@ -2298,6 +2708,9 @@ struct clip_model_loader {
                 }
             } else if (is_audio) {
                 get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);
+                // some hparams are unused, but still need to set to avoid issues
+                hparams.image_size = 0;
+                hparams.patch_size = 1;
 
             } else {
                 GGML_ASSERT(false && "unknown modality");
@@ -2386,58 +2799,69 @@ struct clip_model_loader {
                             hparams.minicpmv_version = 2; // default to 2 if not set
                         }
                     } break;
-                case PROJECTOR_TYPE_IDEFICS3:
-                case PROJECTOR_TYPE_LFM2:
                 case PROJECTOR_TYPE_INTERNVL:
                     {
-                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
+                    } break;
+                case PROJECTOR_TYPE_IDEFICS3:
+                    {
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
+                        get_u32(KEY_PREPROC_IMAGE_SIZE, hparams.image_longest_edge, false);
+                    } break;
+                case PROJECTOR_TYPE_LFM2:
+                    {
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
+                        // ref: https://huggingface.co/LiquidAI/LFM2-VL-3B/blob/main/preprocessor_config.json
+                        hparams.set_limit_image_tokens(64, 256);
                     } break;
                 case PROJECTOR_TYPE_PIXTRAL:
+                case PROJECTOR_TYPE_LIGHTONOCR:
                     {
+                        // ref: https://huggingface.co/mistral-community/pixtral-12b/blob/main/preprocessor_config.json
+                        // TODO: verify the image_min_tokens
+                        hparams.n_merge = 1; // the original pixtral does not use patch merging
                         hparams.rope_theta = 10000.0f;
-                        hparams.warmup_image_size = hparams.patch_size * 8;
-                        // Mistral Small 2506 needs 1024x1024 image size cap to prevent OOM
-                        // ref: https://github.com/ggml-org/llama.cpp/issues/14310
-                        hparams.image_size = 1024;
-                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
+                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
+                        hparams.set_limit_image_tokens(8, 1024);
+                        hparams.set_warmup_n_tokens(256); // avoid OOM on warmup
                     } break;
                 case PROJECTOR_TYPE_KIMIVL:
                     {
                         hparams.rope_theta = 10000.0f;
-                        hparams.warmup_image_size = hparams.patch_size * 8;
-                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
+                        // TODO: check kimivl preprocessor for exact values
+                        hparams.set_limit_image_tokens(8, 1024);
+                        hparams.set_warmup_n_tokens(256); // avoid OOM on warmup
                     } break;
                 case PROJECTOR_TYPE_GEMMA3:
                     {
                         // default value (used by all model sizes in gemma 3 family)
                         // number of patches for each **side** is reduced by a factor of 4
-                        hparams.proj_scale_factor = 4;
+                        hparams.n_merge = 4;
                         // test model (tinygemma3) has a different value, we optionally read it
-                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                     } break;
                 case PROJECTOR_TYPE_QWEN2VL:
-                    {
-                        // max image size = sqrt(max_pixels) = 3584
-                        // ref: https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/blob/main/preprocessor_config.json
-                        // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable
-                        // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10
-                        hparams.image_size = 1024;
-                        hparams.warmup_image_size = hparams.patch_size * 8;
-                    } break;
                 case PROJECTOR_TYPE_QWEN25VL:
+                case PROJECTOR_TYPE_QWEN3VL:
                     {
-                        // max image size = sqrt(max_pixels)
-                        // https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json
-                        // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable
-                        // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10
-                        hparams.image_size = 1024;
-                        hparams.warmup_image_size = hparams.patch_size * 8;
-                        get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern);
+                        hparams.n_merge = 2; // default value for Qwen 2 and 2.5
+                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
+                        get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, model.proj_type == PROJECTOR_TYPE_QWEN25VL); // only 2.5 requires it
+                        // ref: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json
+                        hparams.set_limit_image_tokens(8, 4096);
+                        hparams.set_warmup_n_tokens(46*46); // avoid OOM on warmup
+                        const int warn_min_pixels = 1024 * hparams.n_merge * hparams.n_merge * hparams.patch_size * hparams.patch_size;
+                        if (hparams.image_min_pixels < warn_min_pixels) {
+                            LOG_WRN("%s: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks\n", __func__);
+                            LOG_WRN("%s: if you encounter problems with accuracy, try adding --image-min-tokens 1024\n", __func__);
+                            LOG_WRN("%s: more info: https://github.com/ggml-org/llama.cpp/issues/16842\n\n", __func__);
+                        }
                     } break;
                 case PROJECTOR_TYPE_LLAMA4:
                     {
                         hparams.rope_theta = 10000.0f;
-                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor);
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                         set_llava_uhd_res_candidates(model, 3);
                     } break;
                 case PROJECTOR_TYPE_ULTRAVOX:
@@ -2457,6 +2881,13 @@ struct clip_model_loader {
                     break;
             }
 
+            // sanity check
+            {
+                if (hparams.image_max_pixels < hparams.image_min_pixels) {
+                    throw std::runtime_error(string_format("%s: image_max_pixels (%d) is less than image_min_pixels (%d)\n", __func__, hparams.image_max_pixels, hparams.image_min_pixels));
+                }
+            }
+
             LOG_INF("%s: projector:          %s\n", __func__, proj_type.c_str());
             LOG_INF("%s: n_embd:             %d\n", __func__, hparams.n_embd);
             LOG_INF("%s: n_head:             %d\n", __func__, hparams.n_head);
@@ -2470,8 +2901,14 @@ struct clip_model_loader {
                 LOG_INF("%s: patch_size:         %d\n", __func__, hparams.patch_size);
                 LOG_INF("%s: has_llava_proj:     %d\n", __func__, hparams.has_llava_projector);
                 LOG_INF("%s: minicpmv_version:   %d\n", __func__, hparams.minicpmv_version);
-                LOG_INF("%s: proj_scale_factor:  %d\n", __func__, hparams.proj_scale_factor);
+                LOG_INF("%s: n_merge:            %d\n", __func__, hparams.n_merge);
                 LOG_INF("%s: n_wa_pattern:       %d\n", __func__, hparams.n_wa_pattern);
+                if (hparams.image_min_pixels > 0) {
+                    LOG_INF("%s: image_min_pixels:   %d%s\n", __func__, hparams.image_min_pixels, hparams.custom_image_min_tokens > 0 ? " (custom value)" : "");
+                }
+                if (hparams.image_max_pixels > 0) {
+                    LOG_INF("%s: image_max_pixels:   %d%s\n", __func__, hparams.image_max_pixels, hparams.custom_image_max_tokens > 0 ? " (custom value)" : "");
+                }
             } else if (is_audio) {
                 LOG_INF("\n--- audio hparams ---\n");
                 LOG_INF("%s: n_mel_bins:         %d\n", __func__, hparams.n_mel_bins);
@@ -2543,10 +2980,11 @@ struct clip_model_loader {
         model.layers.resize(hparams.n_layer);
         for (int il = 0; il < hparams.n_layer; ++il) {
             auto & layer = model.layers[il];
-            layer.k_w    = get_tensor(string_format(TN_ATTN_K,      prefix, il, "weight"));
-            layer.q_w    = get_tensor(string_format(TN_ATTN_Q,      prefix, il, "weight"));
-            layer.v_w    = get_tensor(string_format(TN_ATTN_V,      prefix, il, "weight"));
+            layer.k_w    = get_tensor(string_format(TN_ATTN_K,      prefix, il, "weight"), false);
+            layer.q_w    = get_tensor(string_format(TN_ATTN_Q,      prefix, il, "weight"), false);
+            layer.v_w    = get_tensor(string_format(TN_ATTN_V,      prefix, il, "weight"), false);
             layer.o_w    = get_tensor(string_format(TN_ATTN_OUTPUT, prefix, il, "weight"));
+            layer.qkv_w  = get_tensor(string_format(TN_ATTN_QKV,    prefix, il, "weight"), false);
             layer.k_norm = get_tensor(string_format(TN_ATTN_K_NORM, prefix, il, "weight"), false);
             layer.q_norm = get_tensor(string_format(TN_ATTN_Q_NORM, prefix, il, "weight"), false);
             layer.ln_1_w = get_tensor(string_format(TN_LN_1,        prefix, il, "weight"), false);
@@ -2558,6 +2996,7 @@ struct clip_model_loader {
             layer.q_b    = get_tensor(string_format(TN_ATTN_Q,      prefix, il, "bias"), false);
             layer.v_b    = get_tensor(string_format(TN_ATTN_V,      prefix, il, "bias"), false);
             layer.o_b    = get_tensor(string_format(TN_ATTN_OUTPUT, prefix, il, "bias"), false);
+            layer.qkv_b  = get_tensor(string_format(TN_ATTN_QKV,    prefix, il, "bias"), false);
             layer.ln_1_b = get_tensor(string_format(TN_LN_1,        prefix, il, "bias"), false);
             layer.ln_2_b = get_tensor(string_format(TN_LN_2,        prefix, il, "bias"), false);
 
@@ -2569,6 +3008,18 @@ struct clip_model_loader {
             layer.ff_down_w = get_tensor(string_format(TN_FFN_DOWN, prefix, il, "weight"));
             layer.ff_down_b = get_tensor(string_format(TN_FFN_DOWN, prefix, il, "bias"),   false);
 
+
+            // qwen3vl deepstack layer
+            layer.deepstack_norm_w = get_tensor(string_format(TN_DEEPSTACK_NORM, il, "weight"), false);
+            layer.deepstack_norm_b = get_tensor(string_format(TN_DEEPSTACK_NORM, il, "bias"), false);
+            layer.deepstack_fc1_w  = get_tensor(string_format(TN_DEEPSTACK_FC1,  il, "weight"), false);
+            layer.deepstack_fc1_b  = get_tensor(string_format(TN_DEEPSTACK_FC1,  il, "bias"), false);
+            layer.deepstack_fc2_w  = get_tensor(string_format(TN_DEEPSTACK_FC2,  il, "weight"), false);
+            layer.deepstack_fc2_b  = get_tensor(string_format(TN_DEEPSTACK_FC2,  il, "bias"), false);
+            if (layer.has_deepstack()) {
+                model.n_deepstack_layers++;
+            }
+
             // some models already exported with legacy (incorrect) naming which is quite messy, let's fix it here
             // note: Qwen model converted from the old surgery script has n_ff = 0, so we cannot use n_ff to check!
             bool is_ffn_swapped = (
@@ -2693,8 +3144,8 @@ struct clip_model_loader {
                     model.mm_model_mlp_1_w = get_tensor(string_format(TN_GLM_ADAPTER_D_H_2_4H, "weight"));
                     model.mm_model_mlp_2_w = get_tensor(string_format(TN_GLM_ADAPTER_GATE, "weight"));
                     model.mm_model_mlp_3_w = get_tensor(string_format(TN_GLM_ADAPTER_D_4H_2_H, "weight"));
-                    model.mm_glm_tok_boi = get_tensor(string_format(TN_TOK_GLM_BOI, "weight"));
-                    model.mm_glm_tok_eoi = get_tensor(string_format(TN_TOK_GLM_EOI, "weight"));
+                    model.mm_boi = get_tensor(string_format(TN_TOK_GLM_BOI, "weight"));
+                    model.mm_eoi = get_tensor(string_format(TN_TOK_GLM_EOI, "weight"));
                 } break;
             case PROJECTOR_TYPE_QWEN2VL:
             case PROJECTOR_TYPE_QWEN25VL:
@@ -2704,6 +3155,13 @@ struct clip_model_loader {
                     model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
                     model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
                 } break;
+            case PROJECTOR_TYPE_QWEN3VL:
+                {
+                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));
+                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_GEMMA3:
                 {
                     model.mm_input_proj_w = get_tensor(TN_MM_INP_PROJ);
@@ -2735,6 +3193,15 @@ struct clip_model_loader {
                     model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
                     model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
                 } break;
+            case PROJECTOR_TYPE_LIGHTONOCR:
+                {
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false);
+                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false);
+                    model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
+                    model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
+                } break;
             case PROJECTOR_TYPE_ULTRAVOX:
                 {
                     model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
@@ -2779,6 +3246,24 @@ struct clip_model_loader {
                     model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
                     model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 2, "weight"));
                 } break;
+            case PROJECTOR_TYPE_COGVLM:
+                {
+                    model.mm_model_proj     = get_tensor(TN_MM_PROJECTOR);
+                    model.mm_post_fc_norm_w = get_tensor(string_format(TN_MM_POST_FC_NORM, "weight"));
+                    model.mm_post_fc_norm_b = get_tensor(string_format(TN_MM_POST_FC_NORM, "bias"));
+                    model.mm_h_to_4h_w      = get_tensor(string_format(TN_MM_H_TO_4H,      "weight"));
+                    model.mm_gate_w         = get_tensor(string_format(TN_MM_GATE,         "weight"));
+                    model.mm_4h_to_h_w      = get_tensor(string_format(TN_MM_4H_TO_H,      "weight"));
+                    model.mm_boi            = get_tensor(TN_TOK_BOI);
+                    model.mm_eoi            = get_tensor(TN_TOK_EOI);
+                } break;
+            case PROJECTOR_TYPE_JANUS_PRO:
+                {
+                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));
+                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
+                } break;
             default:
                 GGML_ASSERT(false && "unknown projector type");
         }
@@ -2846,7 +3331,87 @@ struct clip_model_loader {
         }
     }
 
-    void alloc_compute_meta(clip_ctx & ctx_clip) {
+    struct support_info_op {
+        ggml_tensor * op;
+
+        // true if the op runs on the accelerated ctx_clip.backend
+        bool is_accel = true;
+    };
+
+    struct support_info_graph {
+        // whether the clip_ctx.backend supports flash attention
+        bool fattn = true;
+        ggml_tensor * fattn_op = nullptr; // for debugging
+
+        std::vector<support_info_op> ops;
+    };
+
+    static void warmup(clip_ctx & ctx_clip) {
+        support_info_graph info;
+
+        if (ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_AUTO) {
+            // try to enable flash attention to see if it's supported
+            ctx_clip.flash_attn_type = CLIP_FLASH_ATTN_TYPE_ENABLED;
+            info = alloc_compute_meta(ctx_clip);
+            if (!info.fattn && info.fattn_op) {
+                auto op = info.fattn_op;
+                LOG_WRN("%s: *****************************************************************\n", __func__);
+                LOG_WRN("%s: WARNING: flash attention not supported by %s, memory usage will increase\n", __func__, ggml_backend_name(ctx_clip.backend));
+                LOG_WRN("%s: op params: \n", __func__);
+                static auto print_shape = [](const char * fn, const char * name, ggml_tensor * t) {
+                    LOG_WRN("%s:   %s: type = %s, ne = [%d %d %d %d], nb = [%d %d %d %d]\n", fn,
+                            name, ggml_type_name(t->type),
+                            t->ne[0], t->ne[1], t->ne[2], t->ne[3],
+                            t->nb[0], t->nb[1], t->nb[2], t->nb[3]);
+                };
+                print_shape(__func__, " dst", op);
+                print_shape(__func__, "src0", op->src[0]);
+                print_shape(__func__, "src1", op->src[1]);
+                print_shape(__func__, "src2", op->src[2]);
+                LOG_WRN("%s: please report this on github as an issue\n", __func__);
+                LOG_WRN("%s: *****************************************************************\n", __func__);
+                ctx_clip.flash_attn_type = CLIP_FLASH_ATTN_TYPE_DISABLED;
+                alloc_compute_meta(ctx_clip);
+            }
+        } else {
+            info = alloc_compute_meta(ctx_clip);
+            if (!info.fattn && ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
+                LOG_WRN("%s: flash attention is not supported by the current backend; falling back to CPU (performance will be degraded)\n", __func__);
+            }
+        }
+
+        LOG_INF("%s: flash attention is %s\n", __func__,
+            (ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) ? "enabled" : "disabled");
+
+        // print ops that are not supported by the GPU backend (if there is one)
+        if (ctx_clip.backend && ctx_clip.backend != ctx_clip.backend_cpu) {
+            std::vector<support_info_op> unsupported_ops;
+            for (const auto & op : info.ops) {
+                if (!op.is_accel) {
+                    unsupported_ops.push_back(op);
+                }
+            }
+            if (!unsupported_ops.empty()) {
+                LOG_WRN("%s: *****************************************************************\n", __func__);
+                LOG_WRN("%s: WARNING: the CLIP graph uses unsupported operators by the backend\n", __func__);
+                LOG_WRN("%s:          the performance will be suboptimal                      \n", __func__);
+                LOG_WRN("%s:          list of unsupported ops (backend=%s):\n", __func__, ggml_backend_name(ctx_clip.backend));
+                for (const auto & op : unsupported_ops) {
+                    LOG_WRN("%s: %16s: type = %s, ne = [%d %d %d %d]\n", __func__,
+                            ggml_op_name(op.op->op),
+                            ggml_type_name(op.op->type),
+                            op.op->ne[0], op.op->ne[1], op.op->ne[2], op.op->ne[3]);
+                }
+                LOG_WRN("%s: flash attention is %s\n", __func__,
+                    (ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) ? "enabled" : "disabled");
+                LOG_WRN("%s: please report this on github as an issue\n", __func__);
+                LOG_WRN("%s: ref: https://github.com/ggml-org/llama.cpp/pull/16837#issuecomment-3461676118\n", __func__);
+                LOG_WRN("%s: *****************************************************************\n", __func__);
+            }
+        }
+    }
+
+    static support_info_graph alloc_compute_meta(clip_ctx & ctx_clip) {
         const auto & hparams = ctx_clip.model.hparams;
         ctx_clip.buf_compute_meta.resize(ctx_clip.max_nodes * ggml_tensor_overhead() + ggml_graph_overhead());
 
@@ -2856,9 +3421,11 @@ struct clip_model_loader {
         if (ctx_clip.model.modality == CLIP_MODALITY_VISION) {
             img->nx = hparams.warmup_image_size;
             img->ny = hparams.warmup_image_size;
+            LOG_INF("%s: warmup with image size = %d x %d\n", __func__, img->nx, img->ny);
         } else {
             img->nx = hparams.warmup_audio_size;
             img->ny = hparams.n_mel_bins;
+            LOG_INF("%s: warmup with audio size = %d\n", __func__, img->nx);
         }
         batch.entries.push_back(std::move(img));
 
@@ -2875,57 +3442,95 @@ struct clip_model_loader {
                         size / 1024.0 / 1024.0);
             }
         }
+
+        const int n_splits = ggml_backend_sched_get_n_splits(ctx_clip.sched.get());
+        const int n_nodes  = ggml_graph_n_nodes(gf);
+
+        LOG_INF("%s: graph splits = %d, nodes = %d\n", __func__,  n_splits, n_nodes);
+
+        support_info_graph res {
+            /*.fattn    = */ true,
+            /*.fattn_op = */ nullptr,
+            /*.ops      = */ {},
+        };
+
+        // check op support
+        for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+            ggml_tensor * node = ggml_graph_node(gf, i);
+            res.ops.push_back({node, true});
+            if (!ggml_backend_supports_op(ctx_clip.backend, node)) {
+                res.ops.back().is_accel = false;
+                if (node->op == GGML_OP_FLASH_ATTN_EXT) {
+                    res.fattn    = false;
+                    res.fattn_op = node;
+                }
+            }
+        }
+
+        return res;
     }
 
-    void get_bool(const std::string & key, bool & output, bool required = true) {
+    void get_bool(const std::string & key, bool & output, bool required = true) const {
         const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
         if (i < 0) {
-            if (required) throw std::runtime_error("Key not found: " + key);
+            if (required) {
+                throw std::runtime_error("Key not found: " + key);
+            }
             return;
         }
         output = gguf_get_val_bool(ctx_gguf.get(), i);
     }
 
-    void get_i32(const std::string & key, int & output, bool required = true) {
+    void get_i32(const std::string & key, int & output, bool required = true) const {
         const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
         if (i < 0) {
-            if (required) throw std::runtime_error("Key not found: " + key);
+            if (required) {
+                throw std::runtime_error("Key not found: " + key);
+            }
             return;
         }
         output = gguf_get_val_i32(ctx_gguf.get(), i);
     }
 
-    void get_u32(const std::string & key, int & output, bool required = true) {
+    void get_u32(const std::string & key, int & output, bool required = true) const {
         const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
         if (i < 0) {
-            if (required) throw std::runtime_error("Key not found: " + key);
+            if (required) {
+                throw std::runtime_error("Key not found: " + key);
+            }
             return;
         }
         output = gguf_get_val_u32(ctx_gguf.get(), i);
     }
 
-    void get_f32(const std::string & key, float & output, bool required = true) {
+    void get_f32(const std::string & key, float & output, bool required = true) const {
         const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
         if (i < 0) {
-            if (required) throw std::runtime_error("Key not found: " + key);
+            if (required) {
+                throw std::runtime_error("Key not found: " + key);
+            }
             return;
         }
         output = gguf_get_val_f32(ctx_gguf.get(), i);
     }
 
-    void get_string(const std::string & key, std::string & output, bool required = true) {
+    void get_string(const std::string & key, std::string & output, bool required = true) const {
         const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
         if (i < 0) {
-            if (required) throw std::runtime_error("Key not found: " + key);
+            if (required) {
+                throw std::runtime_error("Key not found: " + key);
+            }
             return;
         }
         output = std::string(gguf_get_val_str(ctx_gguf.get(), i));
     }
 
-    void get_arr_int(const std::string & key, std::vector<int> & output, bool required = true) {
+    void get_arr_int(const std::string & key, std::vector<int> & output, bool required = true) const {
         const int i = gguf_find_key(ctx_gguf.get(), key.c_str());
         if (i < 0) {
-            if (required) throw std::runtime_error("Key not found: " + key);
+            if (required) {
+                throw std::runtime_error("Key not found: " + key);
+            }
             return;
         }
         int n = gguf_get_arr_n(ctx_gguf.get(), i);
@@ -2936,7 +3541,7 @@ struct clip_model_loader {
         }
     }
 
-    void set_llava_uhd_res_candidates(clip_model & model, const int max_patches_per_side) {
+    static void set_llava_uhd_res_candidates(clip_model & model, const int max_patches_per_side) {
         auto & hparams = model.hparams;
         for (int x = 1; x <= max_patches_per_side; x++) {
             for (int y = 1; y <= max_patches_per_side; y++) {
@@ -2953,7 +3558,6 @@ struct clip_model_loader {
 };
 
 struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params) {
-    g_logger_state.verbosity_thold = ctx_params.verbosity;
     clip_ctx * ctx_vision = nullptr;
     clip_ctx * ctx_audio = nullptr;
 
@@ -2964,24 +3568,22 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
             ctx_vision = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_vision->model, CLIP_MODALITY_VISION);
             loader.load_tensors(*ctx_vision);
-            loader.alloc_compute_meta(*ctx_vision);
+            loader.warmup(*ctx_vision);
         }
 
         if (loader.has_audio) {
             ctx_audio = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_audio->model, CLIP_MODALITY_AUDIO);
             loader.load_tensors(*ctx_audio);
-            loader.alloc_compute_meta(*ctx_audio);
+            loader.warmup(*ctx_audio);
         }
 
     } catch (const std::exception & e) {
         LOG_ERR("%s: failed to load model '%s': %s\n", __func__, fname, e.what());
-        if (ctx_vision) {
-            delete ctx_vision;
-        }
-        if (ctx_audio) {
-            delete ctx_audio;
-        }
+
+        delete ctx_vision;
+        delete ctx_audio;
+
         return {nullptr, nullptr};
     }
 
@@ -3019,10 +3621,10 @@ void clip_image_size_free(struct clip_image_size * load_image_size) {
     }
     delete load_image_size;
 }
-void clip_image_u8_free(struct clip_image_u8  * img) { if (img) delete img; }
-void clip_image_f32_free(struct clip_image_f32 * img) { if (img) delete img; }
-void clip_image_u8_batch_free(struct clip_image_u8_batch * batch) { if (batch) delete batch; }
-void clip_image_f32_batch_free(struct clip_image_f32_batch * batch) { if (batch) delete batch; }
+void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
+void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
+void clip_image_u8_batch_free(struct clip_image_u8_batch * batch) { delete batch; }
+void clip_image_f32_batch_free(struct clip_image_f32_batch * batch) { delete batch; }
 
 size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch) {
     return batch->entries.size();
@@ -3074,9 +3676,169 @@ static void normalize_image_u8_to_f32(const clip_image_u8 & src, clip_image_f32
 
 // set of tools to manupulate images
 // in the future, we can have HW acceleration by allowing this struct to access 3rd party lib like imagick or opencv
-struct image_manipulation {
+struct img_tool {
+    enum resize_algo {
+        RESIZE_ALGO_BILINEAR,
+        RESIZE_ALGO_BICUBIC,
+        // RESIZE_ALGO_LANCZOS, // TODO
+    };
+
+    static void resize(
+            const clip_image_u8 & src,
+            clip_image_u8 & dst,
+            const clip_image_size & target_resolution,
+            resize_algo algo,
+            bool add_padding = true, // TODO: define the behavior for add_padding = false
+            std::array<uint8_t, 3> pad_color = {0, 0, 0}) {
+        dst.nx = target_resolution.width;
+        dst.ny = target_resolution.height;
+        dst.buf.resize(3 * dst.nx * dst.ny);
+
+        if (dst.nx == src.nx && dst.ny == src.ny) {
+            // no resize needed, simple copy
+            dst.buf = src.buf;
+            return;
+        }
+
+        if (!add_padding) {
+            // direct resize
+            switch (algo) {
+                case RESIZE_ALGO_BILINEAR:
+                    resize_bilinear(src, dst, target_resolution.width, target_resolution.height);
+                    break;
+                case RESIZE_ALGO_BICUBIC:
+                    resize_bicubic(src, dst, target_resolution.width, target_resolution.height);
+                    break;
+                default:
+                    throw std::runtime_error("Unsupported resize algorithm");
+            }
+        } else {
+            // resize with padding
+            clip_image_u8 resized_image;
+            float scale_w = static_cast<float>(target_resolution.width) / src.nx;
+            float scale_h = static_cast<float>(target_resolution.height) / src.ny;
+            float scale = std::min(scale_w, scale_h);
+            int new_width  = std::min(static_cast<int>(std::ceil(src.nx * scale)), target_resolution.width);
+            int new_height = std::min(static_cast<int>(std::ceil(src.ny * scale)), target_resolution.height);
+
+            switch (algo) {
+                case RESIZE_ALGO_BILINEAR:
+                    resize_bilinear(src, resized_image, new_width, new_height);
+                    break;
+                case RESIZE_ALGO_BICUBIC:
+                    resize_bicubic(src, resized_image, new_width, new_height);
+                    break;
+                default:
+                    throw std::runtime_error("Unsupported resize algorithm");
+            }
+
+            // fill dst with pad_color
+            fill(dst, pad_color);
+
+            int offset_x = (target_resolution.width  - new_width)  / 2;
+            int offset_y = (target_resolution.height - new_height) / 2;
+
+            composite(dst, resized_image, offset_x, offset_y);
+        }
+    }
+
+    static void crop(const clip_image_u8 & image, clip_image_u8 & dst, int x, int y, int w, int h) {
+        dst.nx = w;
+        dst.ny = h;
+        dst.buf.resize(3 * w * h);
+
+        for (int i = 0; i < h; ++i) {
+            for (int j = 0; j < w; ++j) {
+                int src_idx = 3 * ((y + i)*image.nx + (x + j));
+                int dst_idx = 3 * (i*w + j);
+                dst.buf[dst_idx]     = image.buf[src_idx];
+                dst.buf[dst_idx + 1] = image.buf[src_idx + 1];
+                dst.buf[dst_idx + 2] = image.buf[src_idx + 2];
+            }
+        }
+    }
+
+    // calculate the size of the **resized** image, while preserving the aspect ratio
+    // the calculated size will be aligned to the nearest multiple of align_size
+    // if H or W size is larger than longest_edge, it will be resized to longest_edge
+    static clip_image_size calc_size_preserved_ratio(const clip_image_size & inp_size, const int align_size, const int longest_edge) {
+        GGML_ASSERT(align_size > 0);
+        if (inp_size.width <= 0 || inp_size.height <= 0 || longest_edge <= 0) {
+            return {0, 0};
+        }
+
+        float scale = std::min(static_cast<float>(longest_edge) / inp_size.width,
+                               static_cast<float>(longest_edge) / inp_size.height);
+
+        float target_width_f  = static_cast<float>(inp_size.width)  * scale;
+        float target_height_f = static_cast<float>(inp_size.height) * scale;
+
+        auto ceil_by_factor = [f = align_size](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
+        int aligned_width  = ceil_by_factor(target_width_f);
+        int aligned_height = ceil_by_factor(target_height_f);
+
+        return {aligned_width, aligned_height};
+    }
+
+    // calculate the size of the **resized** image, while preserving the aspect ratio
+    // the calculated size will have min_pixels <= W*H <= max_pixels
+    // this is referred as "smart_resize" in transformers code
+    static clip_image_size calc_size_preserved_ratio(const clip_image_size & inp_size, const int align_size, const int min_pixels, const int max_pixels) {
+        GGML_ASSERT(align_size > 0);
+        const int width  = inp_size.width;
+        const int height = inp_size.height;
+
+        auto ceil_by_factor  = [f = align_size](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
+        auto floor_by_factor = [f = align_size](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
+
+        // always align up first
+        int h_bar = std::max(align_size, ceil_by_factor(height));
+        int w_bar = std::max(align_size, ceil_by_factor(width));
+
+        if (h_bar * w_bar > max_pixels) {
+            const auto beta = std::sqrt(static_cast<float>(height * width) / max_pixels);
+            h_bar = std::max(align_size, floor_by_factor(height / beta));
+            w_bar = std::max(align_size, floor_by_factor(width  / beta));
+        } else if (h_bar * w_bar < min_pixels) {
+            const auto beta = std::sqrt(static_cast<float>(min_pixels) / (height * width));
+            h_bar = ceil_by_factor(height * beta);
+            w_bar = ceil_by_factor(width * beta);
+        }
+
+        return {w_bar, h_bar};
+    }
+
+    // draw src image into dst image at offset (offset_x, offset_y)
+    static void composite(clip_image_u8 & dst, const clip_image_u8 & src, int offset_x, int offset_y) {
+        for (int y = 0; y < src.ny; ++y) {
+            for (int x = 0; x < src.nx; ++x) {
+                int dx = x + offset_x;
+                int dy = y + offset_y;
+                // skip pixels that would be out of bounds in the destination
+                if (dx < 0 || dy < 0 || dx >= dst.nx || dy >= dst.ny) {
+                    continue;
+                }
+                size_t dst_idx = 3 * (static_cast<size_t>(dy) * dst.nx + static_cast<size_t>(dx));
+                size_t src_idx = 3 * (static_cast<size_t>(y) * src.nx + static_cast<size_t>(x));
+                dst.buf[dst_idx + 0] = src.buf[src_idx + 0];
+                dst.buf[dst_idx + 1] = src.buf[src_idx + 1];
+                dst.buf[dst_idx + 2] = src.buf[src_idx + 2];
+            }
+        }
+    }
+
+    // fill the image with a solid color
+    static void fill(clip_image_u8 & img, const std::array<uint8_t, 3> & color) {
+        for (size_t i = 0; i < img.buf.size(); i += 3) {
+            img.buf[i]     = color[0];
+            img.buf[i + 1] = color[1];
+            img.buf[i + 2] = color[2];
+        }
+    }
+
+private:
     // Bilinear resize function
-    static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int target_width, int target_height) {
+    static void resize_bilinear(const clip_image_u8 & src, clip_image_u8 & dst, int target_width, int target_height) {
         dst.nx = target_width;
         dst.ny = target_height;
         dst.buf.resize(3 * target_width * target_height);
@@ -3112,7 +3874,7 @@ struct image_manipulation {
 
     // Bicubic resize function
     // part of image will be cropped if the aspect ratio is different
-    static bool bicubic_resize(const clip_image_u8 & img, clip_image_u8 & dst, int target_width, int target_height) {
+    static bool resize_bicubic(const clip_image_u8 & img, clip_image_u8 & dst, int target_width, int target_height) {
         const int nx = img.nx;
         const int ny = img.ny;
 
@@ -3175,93 +3937,6 @@ struct image_manipulation {
         return true;
     }
 
-    // llava-1.6 type of resize_and_pad
-    // if the ratio is not 1:1, padding with pad_color will be applied
-    // pad_color is single channel, default is 0 (black)
-    static void resize_and_pad_image(const clip_image_u8 & image, clip_image_u8 & dst, const clip_image_size & target_resolution, std::array<uint8_t, 3> pad_color = {0, 0, 0}) {
-        int target_width  = target_resolution.width;
-        int target_height = target_resolution.height;
-
-        float scale_w = static_cast<float>(target_width) / image.nx;
-        float scale_h = static_cast<float>(target_height) / image.ny;
-
-        int new_width, new_height;
-
-        if (scale_w < scale_h) {
-            new_width  = target_width;
-            new_height = std::min(static_cast<int>(std::ceil(image.ny * scale_w)), target_height);
-        } else {
-            new_height = target_height;
-            new_width  = std::min(static_cast<int>(std::ceil(image.nx * scale_h)), target_width);
-        }
-
-        clip_image_u8 resized_image;
-        bicubic_resize(image, resized_image, new_width, new_height);
-
-        clip_image_u8 padded_image;
-        padded_image.nx = target_width;
-        padded_image.ny = target_height;
-        padded_image.buf.resize(3 * target_width * target_height);
-
-        // Fill the padded image with the fill color
-        for (size_t i = 0; i < padded_image.buf.size(); i += 3) {
-            padded_image.buf[i]     = pad_color[0];
-            padded_image.buf[i + 1] = pad_color[1];
-            padded_image.buf[i + 2] = pad_color[2];
-        }
-
-        // Calculate padding offsets
-        int pad_x = (target_width  - new_width)  / 2;
-        int pad_y = (target_height - new_height) / 2;
-
-        // Copy the resized image into the center of the padded buffer
-        for (int y = 0; y < new_height; ++y) {
-            for (int x = 0; x < new_width; ++x) {
-                for (int c = 0; c < 3; ++c) {
-                    padded_image.buf[3 * ((y + pad_y) * target_width + (x + pad_x)) + c] = resized_image.buf[3 * (y * new_width + x) + c];
-                }
-            }
-        }
-        dst = std::move(padded_image);
-    }
-
-    static void crop_image(const clip_image_u8 & image, clip_image_u8 & dst, int x, int y, int w, int h) {
-        dst.nx = w;
-        dst.ny = h;
-        dst.buf.resize(3 * w * h);
-
-        for (int i = 0; i < h; ++i) {
-            for (int j = 0; j < w; ++j) {
-                int src_idx = 3 * ((y + i)*image.nx + (x + j));
-                int dst_idx = 3 * (i*w + j);
-                dst.buf[dst_idx]     = image.buf[src_idx];
-                dst.buf[dst_idx + 1] = image.buf[src_idx + 1];
-                dst.buf[dst_idx + 2] = image.buf[src_idx + 2];
-            }
-        }
-    }
-
-    // calculate the size of the **resized** image, while preserving the aspect ratio
-    // the calculated size will be aligned to the nearest multiple of align_size
-    // if H or W size is larger than max_dimension, it will be resized to max_dimension
-    static clip_image_size calc_size_preserved_ratio(const clip_image_size & inp_size, const int align_size, const int max_dimension) {
-        if (inp_size.width <= 0 || inp_size.height <= 0 || align_size <= 0 || max_dimension <= 0) {
-            return {0, 0};
-        }
-
-        float scale = std::min(1.0f, std::min(static_cast<float>(max_dimension) / inp_size.width,
-                                              static_cast<float>(max_dimension) / inp_size.height));
-
-        float target_width_f  = static_cast<float>(inp_size.width)  * scale;
-        float target_height_f = static_cast<float>(inp_size.height) * scale;
-
-        int aligned_width  = CLIP_ALIGN((int)target_width_f,  align_size);
-        int aligned_height = CLIP_ALIGN((int)target_height_f, align_size);
-
-        return {aligned_width, aligned_height};
-    }
-
-private:
     static inline int clip(int x, int lower, int upper) {
         return std::max(lower, std::min(x, upper));
     }
@@ -3410,10 +4085,11 @@ struct llava_uhd {
 
     static std::vector<clip_image_u8_ptr> slice_image(const clip_image_u8 * img, const slice_instructions & inst) {
         std::vector<clip_image_u8_ptr> output;
+        img_tool::resize_algo interpolation = img_tool::RESIZE_ALGO_BILINEAR; // TODO: make it configurable
 
         // resize to overview size
         clip_image_u8_ptr resized_img(clip_image_u8_init());
-        image_manipulation::bicubic_resize(*img, *resized_img, inst.overview_size.width, inst.overview_size.height);
+        img_tool::resize(*img, *resized_img, inst.overview_size, interpolation);
         output.push_back(std::move(resized_img));
         if (inst.slices.empty()) {
             // no slices, just return the resized image
@@ -3423,9 +4099,11 @@ struct llava_uhd {
         // resize to refined size
         clip_image_u8_ptr refined_img(clip_image_u8_init());
         if (inst.padding_refined) {
-            image_manipulation::resize_and_pad_image(*img, *refined_img, inst.refined_size);
+            img_tool::resize(*img, *refined_img, inst.refined_size, interpolation);
         } else {
-            image_manipulation::bilinear_resize(*img, *refined_img, inst.refined_size.width, inst.refined_size.height);
+            // only algo bicubic preserves the ratio; old models rely on this behavior
+            // TODO: do we need to support other algos here?
+            img_tool::resize(*img, *refined_img, inst.refined_size, img_tool::RESIZE_ALGO_BICUBIC, false);
         }
 
         // create slices
@@ -3436,7 +4114,7 @@ struct llava_uhd {
             int h = slice.size.height;
 
             clip_image_u8_ptr img_slice(clip_image_u8_init());
-            image_manipulation::crop_image(*refined_img, *img_slice, x, y, w, h);
+            img_tool::crop(*refined_img, *img_slice, x, y, w, h);
             output.push_back(std::move(img_slice));
         }
 
@@ -3571,202 +4249,223 @@ private:
 // res_imgs memory is being allocated here, previous allocations will be freed if found
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, struct clip_image_f32_batch * res_imgs) {
     clip_image_size original_size{img->nx, img->ny};
-    bool pad_to_square = true;
     auto & params = ctx->model.hparams;
-    // The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing
-    if (params.mm_patch_merge_type == PATCH_MERGE_SPATIAL_UNPAD) {
-        pad_to_square = false;
-    }
 
-    if (clip_is_minicpmv(ctx)) {
-        auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
-        std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
+    switch (ctx->proj_type()) {
+        case PROJECTOR_TYPE_MINICPMV:
+            {
+                auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
+                std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
 
-        for (size_t i = 0; i < imgs.size(); ++i) {
-            // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
-            clip_image_f32_ptr res(clip_image_f32_init());
-            normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
-            res_imgs->entries.push_back(std::move(res));
-        }
+                for (size_t i = 0; i < imgs.size(); ++i) {
+                    // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
+                    clip_image_f32_ptr res(clip_image_f32_init());
+                    normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
+                    res_imgs->entries.push_back(std::move(res));
+                }
 
-        res_imgs->grid_x = inst.grid_size.width;
-        res_imgs->grid_y = inst.grid_size.height;
-        return true;
+                res_imgs->grid_x = inst.grid_size.width;
+                res_imgs->grid_y = inst.grid_size.height;
+            } break;
 
-    } else if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL) {
-        clip_image_u8 resized;
-        auto patch_size = params.patch_size * 2;
-        auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, patch_size, params.image_size);
-        image_manipulation::bicubic_resize(*img, resized, new_size.width, new_size.height);
+        case PROJECTOR_TYPE_QWEN2VL:
+        case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_QWEN3VL:
+            {
+                GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0);
+                clip_image_u8 resized;
+                const clip_image_size new_size = img_tool::calc_size_preserved_ratio(
+                    original_size,
+                    params.patch_size * 2,
+                    params.image_min_pixels,
+                    params.image_max_pixels);
+                img_tool::resize(*img, resized, new_size, img_tool::RESIZE_ALGO_BILINEAR, false);
+                // clip_image_save_to_bmp(resized, "preproc.bmp");
+                clip_image_f32_ptr img_f32(clip_image_f32_init());
+                // clip_image_f32_ptr res(clip_image_f32_init());
+                normalize_image_u8_to_f32(resized, *img_f32, params.image_mean, params.image_std);
+                // res_imgs->data[0] = *res;
+                res_imgs->entries.push_back(std::move(img_f32));
+            } break;
 
-        clip_image_f32_ptr img_f32(clip_image_f32_init());
-        // clip_image_f32_ptr res(clip_image_f32_init());
-        normalize_image_u8_to_f32(resized, *img_f32, params.image_mean, params.image_std);
-        // res_imgs->data[0] = *res;
-        res_imgs->entries.push_back(std::move(img_f32));
-        return true;
-    } else if (ctx->proj_type() == PROJECTOR_TYPE_IDEFICS3) {
-        // The refined size has two steps:
-        // 1. Resize w/ aspect-ratio preserving such that the longer side is
-        //      the preprocessor longest size
-        // 2. Resize w/out preserving aspect ratio such that both sides are
-        //      multiples of image_size (always rounding up)
-        //
-        // CITE: https://github.com/huggingface/transformers/blob/main/src/transformers/models/idefics3/image_processing_idefics3.py#L737
-        const clip_image_size refined_size = image_manipulation::calc_size_preserved_ratio(
-            original_size, params.image_size, params.preproc_image_size);
+        case PROJECTOR_TYPE_IDEFICS3:
+            {
+                // The refined size has two steps:
+                // 1. Resize w/ aspect-ratio preserving such that the longer side is
+                //      the preprocessor longest size
+                // 2. Resize w/out preserving aspect ratio such that both sides are
+                //      multiples of image_size (always rounding up)
+                //
+                // CITE: https://github.com/huggingface/transformers/blob/main/src/transformers/models/idefics3/image_processing_idefics3.py#L737
+                const clip_image_size refined_size = img_tool::calc_size_preserved_ratio(
+                    original_size, params.image_size, params.image_longest_edge);
+                // LOG_INF("%s: original size: %d x %d, refined size: %d x %d\n",
+                //         __func__, original_size.width, original_size.height,
+                //         refined_size.width, refined_size.height);
 
-        llava_uhd::slice_instructions instructions;
-        instructions.overview_size = clip_image_size{params.image_size, params.image_size};
-        instructions.refined_size = refined_size;
-        instructions.grid_size = clip_image_size{
-            static_cast<int>(std::ceil(static_cast<float>(refined_size.width) / params.image_size)),
-            static_cast<int>(std::ceil(static_cast<float>(refined_size.height) / params.image_size)),
-        };
-        for (int y = 0; y < refined_size.height; y += params.image_size) {
-            for (int x = 0; x < refined_size.width; x += params.image_size) {
-                instructions.slices.push_back(llava_uhd::slice_coordinates{
-                    /* x    */x,
-                    /* y    */y,
-                    /* size */clip_image_size{
-                        std::min(params.image_size, refined_size.width - x),
-                        std::min(params.image_size, refined_size.height - y)
+                llava_uhd::slice_instructions instructions;
+                instructions.overview_size = clip_image_size{params.image_size, params.image_size};
+                instructions.refined_size = refined_size;
+                instructions.grid_size = clip_image_size{
+                    static_cast<int>(std::ceil(static_cast<float>(refined_size.width) / params.image_size)),
+                    static_cast<int>(std::ceil(static_cast<float>(refined_size.height) / params.image_size)),
+                };
+                for (int y = 0; y < refined_size.height; y += params.image_size) {
+                    for (int x = 0; x < refined_size.width; x += params.image_size) {
+                        // LOG_INF("%s: adding slice at x=%d, y=%d\n", __func__, x, y);
+                        instructions.slices.push_back(llava_uhd::slice_coordinates{
+                            /* x    */x,
+                            /* y    */y,
+                            /* size */clip_image_size{
+                                std::min(params.image_size, refined_size.width - x),
+                                std::min(params.image_size, refined_size.height - y)
+                            }
+                        });
                     }
-                });
-            }
-        }
-        auto imgs = llava_uhd::slice_image(img, instructions);
+                }
+                auto imgs = llava_uhd::slice_image(img, instructions);
 
-        // cast and normalize to f32
-        for (size_t i = 0; i < imgs.size(); ++i) {
-            // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
-            clip_image_f32_ptr res(clip_image_f32_init());
-            normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
-            res_imgs->entries.push_back(std::move(res));
-        }
+                // cast and normalize to f32
+                for (size_t i = 0; i < imgs.size(); ++i) {
+                    // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
+                    clip_image_f32_ptr res(clip_image_f32_init());
+                    normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
+                    res_imgs->entries.push_back(std::move(res));
+                }
 
-        res_imgs->grid_x = instructions.grid_size.width;
-        res_imgs->grid_y = instructions.grid_size.height;
-        return true;
-    } else if (ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE
-            || ctx->proj_type() == PROJECTOR_TYPE_GEMMA3
-            || ctx->proj_type() == PROJECTOR_TYPE_INTERNVL // TODO @ngxson : support dynamic resolution
-    ) {
-        clip_image_u8 resized_image;
-        int sz = params.image_size;
-        image_manipulation::resize_and_pad_image(*img, resized_image, {sz, sz});
-        clip_image_f32_ptr img_f32(clip_image_f32_init());
-        //clip_image_save_to_bmp(resized_image, "resized.bmp");
-        normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
-        res_imgs->entries.push_back(std::move(img_f32));
-        return true;
+                res_imgs->grid_x = instructions.grid_size.width;
+                res_imgs->grid_y = instructions.grid_size.height;
+            } break;
 
-    } else if (ctx->proj_type() == PROJECTOR_TYPE_PIXTRAL) {
-        clip_image_u8 resized_image;
-        auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, params.patch_size, params.image_size);
-        image_manipulation::bilinear_resize(*img, resized_image, new_size.width, new_size.height);
-        clip_image_f32_ptr img_f32(clip_image_f32_init());
-        normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
-        res_imgs->entries.push_back(std::move(img_f32));
-        return true;
+        case PROJECTOR_TYPE_GLM_EDGE:
+        case PROJECTOR_TYPE_GEMMA3:
+        case PROJECTOR_TYPE_INTERNVL: // TODO @ngxson : support dynamic resolution
+            {
+                clip_image_u8 resized_image;
+                int sz = params.image_size;
+                img_tool::resize(*img, resized_image, {sz, sz}, img_tool::RESIZE_ALGO_BILINEAR);
+                clip_image_f32_ptr img_f32(clip_image_f32_init());
+                //clip_image_save_to_bmp(resized_image, "resized.bmp");
+                normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
+                res_imgs->entries.push_back(std::move(img_f32));
+            } break;
 
-    } else if (ctx->proj_type() == PROJECTOR_TYPE_LLAMA4) {
-        GGML_ASSERT(!params.image_res_candidates.empty());
-        auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
-        std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
+        case PROJECTOR_TYPE_JANUS_PRO:
+            {
+                // Janus Pro preprocessing: pad to square with gray(127), resize to 384x384
+                const std::array<uint8_t, 3> pad_color = {127, 127, 127};
+                clip_image_u8 resized_image;
+                int sz = params.image_size;
+                img_tool::resize(*img, resized_image, {sz, sz}, img_tool::RESIZE_ALGO_BILINEAR, true, pad_color);
+                clip_image_f32_ptr img_f32(clip_image_f32_init());
+                normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
+                res_imgs->entries.push_back(std::move(img_f32));
+            } break;
 
-        for (size_t i = 0; i < imgs.size(); ++i) {
-            clip_image_f32_ptr res(clip_image_f32_init());
-            normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
-            res_imgs->entries.push_back(std::move(res));
-        }
+        case PROJECTOR_TYPE_PIXTRAL:
+        case PROJECTOR_TYPE_LIGHTONOCR:
+            {
+                GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0);
+                clip_image_u8 resized_image;
+                // the original pixtral model doesn't have n_merge
+                const int cur_merge = params.n_merge == 0 ? 1 : params.n_merge;
+                const clip_image_size target_size = img_tool::calc_size_preserved_ratio(
+                    original_size,
+                    params.patch_size * cur_merge,
+                    params.image_min_pixels,
+                    params.image_max_pixels);
+                img_tool::resize(*img, resized_image, target_size, img_tool::RESIZE_ALGO_BILINEAR);
+                clip_image_f32_ptr img_f32(clip_image_f32_init());
+                normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
+                res_imgs->entries.push_back(std::move(img_f32));
+            } break;
 
-        res_imgs->grid_x = inst.grid_size.width;
-        res_imgs->grid_y = inst.grid_size.height;
-        return true;
+        case PROJECTOR_TYPE_LLAMA4:
+            {
+                GGML_ASSERT(!params.image_res_candidates.empty());
+                auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
+                std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
 
-    } else if ( ctx->proj_type() == PROJECTOR_TYPE_LFM2
-             || ctx->proj_type() == PROJECTOR_TYPE_KIMIVL
-    ) {
-        GGML_ASSERT(params.proj_scale_factor);
+                for (size_t i = 0; i < imgs.size(); ++i) {
+                    clip_image_f32_ptr res(clip_image_f32_init());
+                    normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
+                    res_imgs->entries.push_back(std::move(res));
+                }
 
-        // smart resize
-        const int width = img->nx;
-        const int height = img->ny;
-        const int total_factor = params.patch_size * params.proj_scale_factor;
-        constexpr int min_image_tokens = 64;
-        constexpr int max_image_tokens = 1024;
-        const float min_pixels = min_image_tokens * total_factor * total_factor;
-        const float max_pixels = max_image_tokens * total_factor * total_factor;
+                res_imgs->grid_x = inst.grid_size.width;
+                res_imgs->grid_y = inst.grid_size.height;
+            } break;
 
-        auto round_by_factor = [f = total_factor](float x) { return static_cast<int>(std::nearbyintf(x / static_cast<float>(f))) * f; };
-        auto ceil_by_factor  = [f = total_factor](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
-        auto floor_by_factor = [f = total_factor](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
+        case PROJECTOR_TYPE_LFM2:
+        case PROJECTOR_TYPE_KIMIVL:
+            {
+                GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0);
+                const clip_image_size target_size = img_tool::calc_size_preserved_ratio(
+                    original_size,
+                    params.patch_size * params.n_merge,
+                    params.image_min_pixels,
+                    params.image_max_pixels);
+                const std::array<uint8_t, 3> pad_color = {122, 116, 104};
 
-        int h_bar = std::max(total_factor, round_by_factor(height));
-        int w_bar = std::max(total_factor, round_by_factor(width));
+                clip_image_u8 resized_img;
+                img_tool::resize(*img, resized_img, target_size, img_tool::RESIZE_ALGO_BILINEAR, true, pad_color);
+                clip_image_f32_ptr res(clip_image_f32_init());
+                normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
+                res_imgs->entries.push_back(std::move(res));
+            } break;
 
-        if (h_bar * w_bar > max_pixels) {
-            const auto beta = std::sqrt((height * width) / max_pixels);
-            h_bar = std::max(total_factor, floor_by_factor(height / beta));
-            w_bar = std::max(total_factor, floor_by_factor(width / beta));
-        } else if (h_bar * w_bar < min_pixels) {
-            const auto beta = std::sqrt(min_pixels / (height * width));
-            h_bar = ceil_by_factor(height * beta);
-            w_bar = ceil_by_factor(width * beta);
-        }
+        case PROJECTOR_TYPE_MLP:
+        case PROJECTOR_TYPE_MLP_NORM:
+        case PROJECTOR_TYPE_LDP:
+        case PROJECTOR_TYPE_LDPV2:
+        case PROJECTOR_TYPE_COGVLM: // TODO @ngxson : is this correct for cogvlm?
+            {
+                // TODO @ngxson : refactor the code below to avoid duplicated logic
 
-        const std::array<uint8_t, 3> pad_color = {122, 116, 104};
+                // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
+                // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
 
-        clip_image_u8 resized_img;
-        image_manipulation::resize_and_pad_image(*img, resized_img, clip_image_size{w_bar, h_bar}, pad_color);
-        clip_image_f32_ptr res(clip_image_f32_init());
-        normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
-        res_imgs->entries.push_back(std::move(res));
-        return true;
-    }
-
-    // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
-    // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
-
-    clip_image_u8_ptr temp(clip_image_u8_init()); // we will keep the input image data here temporarily
-
-    if (pad_to_square) {
-        // for llava-1.5, we resize image to a square, and pad the shorter side with a background color
-        // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
-        const int longer_side = std::max(img->nx, img->ny);
-        temp->nx = longer_side;
-        temp->ny = longer_side;
-        temp->buf.resize(3 * longer_side * longer_side);
-
-        // background color in RGB from LLaVA (this is the mean rgb color * 255)
-        const std::array<uint8_t, 3> pad_color = {122, 116, 104};
-
-        // resize the image to the target_size
-        image_manipulation::resize_and_pad_image(*img, *temp, clip_image_size{params.image_size, params.image_size}, pad_color);
-
-        clip_image_f32_ptr res(clip_image_f32_init());
-        normalize_image_u8_to_f32(*temp, *res, params.image_mean, params.image_std);
-        res_imgs->entries.push_back(std::move(res));
-        return true;
-
-    } else if (!params.image_res_candidates.empty()) {
-        // "spatial_unpad" with "anyres" processing for llava-1.6
-        auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
-        std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
-
-        for (size_t i = 0; i < imgs.size(); ++i) {
-            // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
-            clip_image_f32_ptr res(clip_image_f32_init());
-            normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
-            res_imgs->entries.push_back(std::move(res));
-        }
-
-        return true;
-    } else {
-        GGML_ABORT("Unknown image preprocessing type");
+                clip_image_u8_ptr temp(clip_image_u8_init()); // we will keep the input image data here temporarily
+
+                // The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing
+                if (params.image_res_candidates.empty()) { // pad_to_square
+                    // for llava-1.5, we resize image to a square, and pad the shorter side with a background color
+                    // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
+                    const int longer_side = std::max(img->nx, img->ny);
+                    temp->nx = longer_side;
+                    temp->ny = longer_side;
+                    temp->buf.resize(3 * longer_side * longer_side);
+
+                    // background color in RGB from LLaVA (this is the mean rgb color * 255)
+                    const std::array<uint8_t, 3> pad_color = {122, 116, 104};
+
+                    // resize the image to the target_size
+                    img_tool::resize(*img, *temp, clip_image_size{params.image_size, params.image_size}, img_tool::RESIZE_ALGO_BILINEAR, true, pad_color);
+
+                    clip_image_f32_ptr res(clip_image_f32_init());
+                    normalize_image_u8_to_f32(*temp, *res, params.image_mean, params.image_std);
+                    res_imgs->entries.push_back(std::move(res));
+
+                } else {
+                    // "spatial_unpad" with "anyres" processing for llava-1.6
+                    auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
+                    std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
+
+                    for (size_t i = 0; i < imgs.size(); ++i) {
+                        // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
+                        clip_image_f32_ptr res(clip_image_f32_init());
+                        normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
+                        res_imgs->entries.push_back(std::move(res));
+                    }
+                }
+            } break;
+
+        default:
+            LOG_ERR("%s: unsupported projector type %d\n", __func__, ctx->proj_type());
+            return false;
     }
 
+    return true;
 }
 
 ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
@@ -3813,16 +4512,16 @@ const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
 int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
     const int n_total = clip_n_output_tokens(ctx, img);
-    if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL) {
-        return img->nx / (params.patch_size * 2) + (int)(img->nx % params.patch_size > 0);
+    if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN3VL) {
+        return img->nx / (params.patch_size * 2);
     }
     return n_total;
 }
 
 int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
-    if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL) {
-        return img->ny / (params.patch_size * 2) + (int)(img->ny % params.patch_size > 0);
+    if (ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN3VL) {
+        return img->ny / (params.patch_size * 2);
     }
     return 1;
 }
@@ -3839,6 +4538,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
     switch (proj) {
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_MLP_NORM:
+        case PROJECTOR_TYPE_JANUS_PRO:
             {
                 // do nothing
             } break;
@@ -3847,7 +4547,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_GLM_EDGE:
             {
                 n_patches /= 4;
-                if (ctx->model.mm_glm_tok_boi) {
+                if (ctx->model.mm_boi) {
                     n_patches += 2; // for BOI and EOI token embeddings
                 }
             } break;
@@ -3877,11 +4577,11 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
             } break;
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_QWEN3VL:
             {
                 // dynamic size (2 conv, so double patch size)
-                int patch_size = params.patch_size * 2;
-                int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
-                int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0);
+                int x_patch = img->nx / (params.patch_size * 2);
+                int y_patch = img->ny / (params.patch_size * 2);
                 n_patches = x_patch * y_patch;
             } break;
         case PROJECTOR_TYPE_GEMMA3:
@@ -3890,26 +4590,30 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_LLAMA4:
             {
                 // both X and Y are downscaled by the scale factor
-                int scale_factor = ctx->model.hparams.proj_scale_factor;
+                int scale_factor = ctx->model.hparams.n_merge;
                 n_patches /= (scale_factor * scale_factor);
             } break;
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL:
             {
                 // dynamic size
-                int scale_factor = ctx->model.hparams.proj_scale_factor;
-                int out_patch_size = params.patch_size * scale_factor;
+                int out_patch_size = params.patch_size * ctx->model.hparams.n_merge;
                 int x_patch = CLIP_ALIGN(img->nx, out_patch_size) / out_patch_size;
                 int y_patch = CLIP_ALIGN(img->ny, out_patch_size) / out_patch_size;
                 n_patches = x_patch * y_patch;
             } break;
         case PROJECTOR_TYPE_PIXTRAL:
+        case PROJECTOR_TYPE_LIGHTONOCR:
             {
                 // dynamic size
-                int n_merge = params.spatial_merge_size;
+                int n_merge = ctx->model.hparams.n_merge;
                 int n_patches_x = img->nx / patch_size / (n_merge > 0 ? n_merge : 1);
                 int n_patches_y = img->ny / patch_size / (n_merge > 0 ? n_merge : 1);
-                n_patches = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
+                if (ctx->model.token_embd_img_break) {
+                    n_patches = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
+                } else {
+                    n_patches = n_patches_y * n_patches_x;
+                }
             } break;
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_ULTRAVOX:
@@ -3932,6 +4636,10 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
                     n_patches /= 2;
                 }
             } break;
+        case PROJECTOR_TYPE_COGVLM:
+            {
+                n_patches += 2; // for BOI and EOI token embeddings
+            } break;
         default:
             GGML_ABORT("unsupported projector type");
     }
@@ -3939,92 +4647,6 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
     return n_patches;
 }
 
-static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
-    assert(embed_dim % 2 == 0);
-    int H = pos.size();
-    int W = pos[0].size();
-
-    std::vector<float> omega(embed_dim / 2);
-    for (int i = 0; i < embed_dim / 2; ++i) {
-        omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
-    }
-
-    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
-    for (int h = 0; h < H; ++h) {
-        for (int w = 0; w < W; ++w) {
-            for (int d = 0; d < embed_dim / 2; ++d) {
-                float out_value = pos[h][w] * omega[d];
-                emb[h][w][d] = sin(out_value);
-                emb[h][w][d + embed_dim / 2] = cos(out_value);
-            }
-        }
-    }
-
-    return emb;
-}
-
-static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
-    assert(embed_dim % 2 == 0);
-    std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
-    std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)
-
-    int H = emb_h.size();
-    int W = emb_h[0].size();
-    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
-
-    for (int h = 0; h < H; ++h) {
-        for (int w = 0; w < W; ++w) {
-            for (int d = 0; d < embed_dim / 2; ++d) {
-                emb[h][w][d] = emb_h[h][w][d];
-                emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
-            }
-        }
-    }
-    return emb;
-}
-
-static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
-    int grid_h_size = image_size.first;
-    int grid_w_size = image_size.second;
-
-    std::vector<float> grid_h(grid_h_size);
-    std::vector<float> grid_w(grid_w_size);
-
-    for (int i = 0; i < grid_h_size; ++i) {
-        grid_h[i] = static_cast<float>(i);
-    }
-    for (int i = 0; i < grid_w_size; ++i) {
-        grid_w[i] = static_cast<float>(i);
-    }
-
-    std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
-    for (int h = 0; h < grid_h_size; ++h) {
-        for (int w = 0; w < grid_w_size; ++w) {
-            grid[h][w] = grid_w[w];
-        }
-    }
-    std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
-    for (int h = 0; h < grid_h_size; ++h) {
-        for (int w = 0; w < grid_w_size; ++w) {
-            grid_2d[0][h][w] = grid_h[h];
-            grid_2d[1][h][w] = grid_w[w];
-        }
-    }
-
-    std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);
-
-    int H = image_size.first;
-    int W = image_size.second;
-    std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
-    for (int h = 0; h < H; ++h) {
-        for (int w = 0; w < W; ++w) {
-            pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
-        }
-    }
-
-    return pos_embed_2d;
-}
-
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     clip_image_f32_batch imgs;
     clip_image_f32_ptr img_copy(clip_image_f32_init());
@@ -4163,26 +4785,33 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 }
                 set_input_i32("positions", positions);
 
-                // inspired from resampler of Qwen-VL:
-                //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
-                //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
-                int embed_dim = clip_n_mmproj_embd(ctx);
-
-                // TODO @ngxson : this is very inefficient, can we do this using ggml_sin and ggml_cos?
-                auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
-
-                std::vector<float> pos_embed(embed_dim * pos_w * pos_h);
-                for(int i = 0; i < pos_w * pos_h; ++i){
-                    for(int j = 0; j < embed_dim; ++j){
-                        pos_embed[i * embed_dim + j] = pos_embed_t[i][j];
-                    }
+                // inputs for resampler projector
+                // set the 2D positions (using float for sinusoidal embedding)
+                int n_patches_per_col = image_size_width / patch_size;
+                std::vector<float> pos_data(n_pos);
+                // dimension H
+                for (int i = 0; i < n_pos; i++) {
+                    pos_data[i] = static_cast<float>(i / n_patches_per_col);
                 }
-
-                set_input_f32("pos_embed", pos_embed);
+                set_input_f32("pos_h", pos_data);
+                // dimension W
+                for (int i = 0; i < n_pos; i++) {
+                    pos_data[i] = static_cast<float>(i % n_patches_per_col);
+                }
+                set_input_f32("pos_w", pos_data);
+                // base frequency omega
+                const float base_freq   = 10000.0f;
+                const int   n_embd_proj = clip_n_mmproj_embd(ctx);
+                std::vector<float> omega(n_embd_proj / 4);
+                for (int i = 0; i < n_embd_proj / 4; ++i) {
+                    omega[i] = 1.0f / std::pow(base_freq, static_cast<float>(i) / (n_embd_proj / 4));
+                }
+                set_input_f32("omega", omega);
             } break;
         case PROJECTOR_TYPE_QWEN2VL:
+        case PROJECTOR_TYPE_QWEN3VL:
             {
-                const int merge_ratio = 2;
+                const int merge_ratio = hparams.n_merge;
                 const int pw = image_size_width  / patch_size;
                 const int ph = image_size_height / patch_size;
                 std::vector<int> positions(n_pos * 4);
@@ -4286,6 +4915,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             } break;
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_KIMIVL:
+        case PROJECTOR_TYPE_LIGHTONOCR:
             {
                 // set the 2D positions
                 int n_patches_per_col = image_size_width / patch_size;
@@ -4339,6 +4969,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_VOXTRAL:
+        case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_COGVLM:
             {
                 // do nothing
             } break;
@@ -4416,6 +5048,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_model_peg_0_b->ne[0];
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_PIXTRAL:
+        case PROJECTOR_TYPE_LIGHTONOCR:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_MLP_NORM:
             return ctx->model.mm_3_b->ne[0];
@@ -4425,7 +5058,11 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_model_mlp_3_w->ne[1];
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_JANUS_PRO:
             return ctx->model.mm_1_b->ne[0];
+        case PROJECTOR_TYPE_QWEN3VL:
+            // main path + deepstack paths
+            return ctx->model.mm_1_b->ne[0] * (1 + ctx->model.n_deepstack_layers);
         case PROJECTOR_TYPE_GEMMA3:
             return ctx->model.mm_input_proj_w->ne[0];
         case PROJECTOR_TYPE_IDEFICS3:
@@ -4442,6 +5079,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL:
             return ctx->model.mm_2_w->ne[1];
+        case PROJECTOR_TYPE_COGVLM:
+            return ctx->model.mm_4h_to_h_w->ne[1];
         default:
             GGML_ABORT("Unknown projector type");
     }
@@ -4460,7 +5099,8 @@ bool clip_is_glm(const struct clip_ctx * ctx) {
 
 bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
     return ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL
-        || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL;
+        || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL
+        || ctx->proj_type() == PROJECTOR_TYPE_QWEN3VL;
 }
 
 bool clip_is_llava(const struct clip_ctx * ctx) {
diff --git a/llama/llama.cpp/tools/mtmd/clip.h b/llama/llama.cpp/tools/mtmd/clip.h
index 3387cdbd..c1442afe 100644
--- a/llama/llama.cpp/tools/mtmd/clip.h
+++ b/llama/llama.cpp/tools/mtmd/clip.h
@@ -1,6 +1,7 @@
 #pragma once
 
 #include "ggml.h"
+
 #include <stddef.h>
 #include <stdint.h>
 
@@ -22,9 +23,17 @@ enum clip_modality {
     CLIP_MODALITY_AUDIO,
 };
 
+enum clip_flash_attn_type {
+    CLIP_FLASH_ATTN_TYPE_AUTO     = -1,
+    CLIP_FLASH_ATTN_TYPE_DISABLED = 0,
+    CLIP_FLASH_ATTN_TYPE_ENABLED  = 1,
+};
+
 struct clip_context_params {
     bool use_gpu;
-    enum ggml_log_level verbosity;
+    enum clip_flash_attn_type flash_attn_type;
+    int image_min_tokens;
+    int image_max_tokens;
 };
 
 struct clip_init_result {
diff --git a/llama/llama.cpp/tools/mtmd/mtmd-helper.cpp b/llama/llama.cpp/tools/mtmd/mtmd-helper.cpp
index 686f42f3..f0891bba 100644
--- a/llama/llama.cpp/tools/mtmd/mtmd-helper.cpp
+++ b/llama/llama.cpp/tools/mtmd/mtmd-helper.cpp
@@ -32,8 +32,65 @@
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb/stb_image.h"
 
-#define LOG_INF(...) fprintf(stdout, __VA_ARGS__)
-#define LOG_ERR(...) fprintf(stderr, __VA_ARGS__)
+//
+// internal logging functions
+//
+
+struct mtmd_helper_logger {
+    ggml_log_callback default_callback = [](ggml_log_level level, const char * text, void * user_data) {
+        (void) level;
+        (void) user_data;
+        fputs(text, stderr);
+        fflush(stderr);
+    };
+
+    ggml_log_callback log_callback = default_callback;
+    void * log_callback_user_data;
+
+    void log_v(enum ggml_log_level level, const char * format, va_list args) {
+        if (format == NULL) {
+            return;
+        }
+        va_list args_copy;
+        va_copy(args_copy, args);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            log_callback(level, buffer, log_callback_user_data);
+        } else {
+            char * buffer2 = (char *) calloc(len + 1, sizeof(char));
+            vsnprintf(buffer2, len + 1, format, args_copy);
+            buffer2[len] = 0;
+            log_callback(level, buffer2, log_callback_user_data);
+            free(buffer2);
+        }
+        va_end(args_copy);
+    }
+
+    void log(enum ggml_log_level level, const char * format, ...) {
+        va_list args;
+        va_start(args, format);
+        log_v(level, format, args);
+        va_end(args);
+    }
+} g_logger;
+
+#define LOG_INF(...) g_logger.log(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define LOG_WRN(...) g_logger.log(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
+#define LOG_ERR(...) g_logger.log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+void mtmd_helper_log_set(ggml_log_callback log_callback, void * user_data) {
+    if (log_callback == nullptr) {
+        log_callback = g_logger.default_callback;
+    }
+    g_logger.log_callback = log_callback;
+    g_logger.log_callback_user_data = user_data;
+    mtmd_log_set(log_callback, user_data);
+}
+
+//
+// helper functions
+//
 
 size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks) {
     size_t n_tokens = 0;
@@ -182,7 +239,7 @@ int32_t mtmd_helper_decode_image_chunk(
     }
 
     const llama_model * model = llama_get_model(lctx);
-    int n_mmproj_embd = llama_model_n_embd(model);
+    int n_mmproj_embd = llama_model_n_embd_inp(model);
     int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
 
     int32_t n_tokens = mtmd_input_chunk_get_n_tokens(chunk);
@@ -325,7 +382,7 @@ int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
                                 llama_pos * new_n_past) {
     size_t n_chunks = mtmd_input_chunks_size(chunks);
     if (n_chunks == 0) {
-        LOG_ERR("no chunks to eval\n");
+        LOG_WRN("no chunks to eval\n");
         return 0;
     }
 
diff --git a/llama/llama.cpp/tools/mtmd/mtmd-helper.h b/llama/llama.cpp/tools/mtmd/mtmd-helper.h
index 5c0edc69..5036b924 100644
--- a/llama/llama.cpp/tools/mtmd/mtmd-helper.h
+++ b/llama/llama.cpp/tools/mtmd/mtmd-helper.h
@@ -20,6 +20,11 @@ extern "C" {
 // BREAKING CHANGES are expected.
 //
 
+// Set callback for all future logging events.
+// If this is not called, or NULL is supplied, everything is output on stderr.
+// Note: this also call mtmd_log_set() internally
+MTMD_API void mtmd_helper_log_set(ggml_log_callback log_callback, void * user_data);
+
 // helper function to construct a mtmd_bitmap from a file
 // it calls mtmd_helper_bitmap_init_from_buf() internally
 // returns nullptr on failure
diff --git a/llama/llama.cpp/tools/mtmd/mtmd.cpp b/llama/llama.cpp/tools/mtmd/mtmd.cpp
index 35a0d25e..9858de63 100644
--- a/llama/llama.cpp/tools/mtmd/mtmd.cpp
+++ b/llama/llama.cpp/tools/mtmd/mtmd.cpp
@@ -5,12 +5,20 @@
 
 #include "llama.h"
 
+// fix problem with std::min and std::max
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 #include <algorithm>
 #include <cerrno>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
-#include <limits>
 #include <vector>
 
 // represents raw image data, layout is RGBRGBRGB...
@@ -93,14 +101,26 @@ const char * mtmd_default_marker() {
     return "<__media__>";
 }
 
+static clip_flash_attn_type mtmd_get_clip_flash_attn_type(enum llama_flash_attn_type flash_attn_type) {
+    switch (flash_attn_type) {
+        case LLAMA_FLASH_ATTN_TYPE_AUTO:     return CLIP_FLASH_ATTN_TYPE_AUTO;
+        case LLAMA_FLASH_ATTN_TYPE_DISABLED: return CLIP_FLASH_ATTN_TYPE_DISABLED;
+        case LLAMA_FLASH_ATTN_TYPE_ENABLED:  return CLIP_FLASH_ATTN_TYPE_ENABLED;
+    }
+    return CLIP_FLASH_ATTN_TYPE_AUTO;
+}
+
 mtmd_context_params mtmd_context_params_default() {
-    mtmd_context_params params;
-    params.use_gpu = true;
-    params.print_timings = true;
-    params.n_threads = 4;
-    params.verbosity = GGML_LOG_LEVEL_INFO;
-    params.image_marker = MTMD_DEFAULT_IMAGE_MARKER;
-    params.media_marker = mtmd_default_marker();
+    mtmd_context_params params {
+        /* use_gpu           */ true,
+        /* print_timings     */ true,
+        /* n_threads         */ 4,
+        /* image_marker      */ MTMD_DEFAULT_IMAGE_MARKER,
+        /* media_marker      */ mtmd_default_marker(),
+        /* flash_attn_type   */ LLAMA_FLASH_ATTN_TYPE_AUTO,
+        /* image_min_tokens  */ -1,
+        /* image_max_tokens  */ -1,
+    };
     return params;
 }
 
@@ -152,7 +172,7 @@ struct mtmd_context {
         print_timings(ctx_params.print_timings),
         n_threads    (ctx_params.n_threads),
         media_marker (ctx_params.media_marker),
-        n_embd_text  (llama_model_n_embd(text_model))
+        n_embd_text  (llama_model_n_embd_inp(text_model))
     {
         if (std::string(ctx_params.image_marker) != MTMD_DEFAULT_IMAGE_MARKER) {
             throw std::runtime_error("custom image_marker is not supported anymore, use media_marker instead");
@@ -162,9 +182,13 @@ struct mtmd_context {
             throw std::runtime_error("media_marker must not be empty");
         }
 
-        clip_context_params ctx_clip_params;
-        ctx_clip_params.use_gpu   = ctx_params.use_gpu;
-        ctx_clip_params.verbosity = ctx_params.verbosity;
+        clip_context_params ctx_clip_params {
+            /* use_gpu           */ ctx_params.use_gpu,
+            /* flash_attn_type   */ CLIP_FLASH_ATTN_TYPE_AUTO,
+            /* image_min_tokens  */ ctx_params.image_min_tokens,
+            /* image_max_tokens  */ ctx_params.image_max_tokens,
+        };
+
         auto res = clip_init(mmproj_fname, ctx_clip_params);
         ctx_v = res.ctx_v;
         ctx_a = res.ctx_a;
@@ -268,7 +292,7 @@ struct mtmd_context {
             // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
             img_end = "[IMG_END]";
 
-        } else if (proj == PROJECTOR_TYPE_QWEN2VL || proj == PROJECTOR_TYPE_QWEN25VL) {
+        } else if (proj == PROJECTOR_TYPE_QWEN2VL || proj == PROJECTOR_TYPE_QWEN25VL || proj == PROJECTOR_TYPE_QWEN3VL) {
             // <|vision_start|> ... (image embeddings) ... <|vision_end|>
             img_beg = "<|vision_start|>";
             img_end = "<|vision_end|>";
@@ -285,6 +309,11 @@ struct mtmd_context {
             img_beg = "<img>";
             img_end = "</img>";
 
+        } else if (proj == PROJECTOR_TYPE_LIGHTONOCR) {
+            // <|im_start|> ... (image embeddings) ... <|im_end|>
+            img_beg = "<|im_start|>";
+            img_end = "<|im_end|>";
+
         }
     }
 
@@ -374,9 +403,7 @@ mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
 }
 
 void mtmd_free(mtmd_context * ctx) {
-    if (ctx) {
-        delete ctx;
-    }
+    delete ctx;
 }
 
 struct mtmd_tokenizer {
@@ -1036,7 +1063,9 @@ const char * mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
 
 llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
     if (image_tokens->use_mrope_pos) {
-        return 1; // for M-RoPE, the whole image is 1 in temporal dimension
+        // for M-RoPE, temporal dimension = max(t,h,w)
+        // t is omitted as we don't support video input
+        return std::max(image_tokens->nx, image_tokens->ny);
     }
     return image_tokens->n_tokens();
 }
@@ -1075,3 +1104,8 @@ mtmd_input_chunks * mtmd_test_create_input_chunks() {
 
     return chunks;
 }
+
+void mtmd_log_set(ggml_log_callback log_callback, void * user_data) {
+    g_logger_state.log_callback = log_callback ? log_callback : clip_log_callback_default;
+    g_logger_state.log_callback_user_data = user_data;
+}
diff --git a/llama/llama.cpp/tools/mtmd/mtmd.h b/llama/llama.cpp/tools/mtmd/mtmd.h
index cf287224..8d3fa5d3 100644
--- a/llama/llama.cpp/tools/mtmd/mtmd.h
+++ b/llama/llama.cpp/tools/mtmd/mtmd.h
@@ -82,9 +82,13 @@ struct mtmd_context_params {
     bool use_gpu;
     bool print_timings;
     int n_threads;
-    enum ggml_log_level verbosity;
     const char * image_marker; // deprecated, use media_marker instead
     const char * media_marker;
+    enum llama_flash_attn_type flash_attn_type;
+
+    // limit number of image tokens, only for vision models with dynamic resolution
+    int image_min_tokens; // minimum number of tokens for image input (default: read from metadata)
+    int image_max_tokens; // maximum number of tokens for image input (default: read from metadata)
 };
 
 MTMD_API const char * mtmd_default_marker(void);
@@ -156,7 +160,7 @@ MTMD_API const mtmd_image_tokens *  mtmd_input_chunk_get_tokens_image(const mtmd
 MTMD_API size_t                     mtmd_input_chunk_get_n_tokens    (const mtmd_input_chunk * chunk);
 // returns nullptr for ID on text chunk
 MTMD_API const char *               mtmd_input_chunk_get_id          (const mtmd_input_chunk * chunk);
-// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
+// number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)
 MTMD_API llama_pos                  mtmd_input_chunk_get_n_pos       (const mtmd_input_chunk * chunk);
 
 // in case you want to use custom logic to handle the chunk (i.e. KV cache management)
@@ -174,7 +178,7 @@ MTMD_API size_t       mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * i
 MTMD_API size_t       mtmd_image_tokens_get_nx      (const mtmd_image_tokens * image_tokens);
 MTMD_API size_t       mtmd_image_tokens_get_ny      (const mtmd_image_tokens * image_tokens);
 MTMD_API const char * mtmd_image_tokens_get_id      (const mtmd_image_tokens * image_tokens); // TODO: deprecate
-// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
+// number of temporal positions (equals to max(t,h,w) for M-RoPE; equals to n_tokens otherwise)
 MTMD_API llama_pos    mtmd_image_tokens_get_n_pos   (const mtmd_image_tokens * image_tokens); // TODO: deprecate
 
 // tokenize an input text prompt and a list of bitmaps (images/audio)
@@ -213,6 +217,10 @@ MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
 // llama_model_n_embd(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
 MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 
+// Set callback for all future logging events.
+// If this is not called, or NULL is supplied, everything is output on stderr.
+MTMD_API void mtmd_log_set(ggml_log_callback log_callback, void * user_data);
+
 /////////////////////////////////////////
 
 // test function, to be used in test-mtmd-c-api.c
diff --git a/llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch b/llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch
index e201f83b..4a2ee02f 100644
--- a/llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch
+++ b/llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch
@@ -23,7 +23,7 @@ problem.
  8 files changed, 21 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index ff9135fe2..8ba86f824 100644
+index 4cf377e7f..4882541c8 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -113,7 +113,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@@ -42,7 +42,7 @@ index ff9135fe2..8ba86f824 100644
  }
  
  static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-@@ -2075,6 +2075,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+@@ -2079,6 +2079,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
  static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      GGML_ASSERT(buffer);
      ggml_aligned_free(buffer->context, buffer->size);
@@ -54,7 +54,7 @@ index ff9135fe2..8ba86f824 100644
  }
  
  static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-@@ -2127,7 +2132,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+@@ -2131,7 +2136,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
  };
  
  static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
@@ -64,10 +64,10 @@ index ff9135fe2..8ba86f824 100644
      /* .init_tensor     = */ NULL, // no initialization required
      /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
 diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
-index 8bd5449f1..01e2df61a 100644
+index df28d67fb..1f6a56ba2 100644
 --- a/ggml/src/ggml-cann/ggml-cann.cpp
 +++ b/ggml/src/ggml-cann/ggml-cann.cpp
-@@ -820,6 +820,7 @@ static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
+@@ -831,6 +831,7 @@ static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
  static void ggml_backend_cann_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cann_buffer_context * ctx = (ggml_backend_cann_buffer_context *) buffer->context;
      delete ctx;
@@ -75,7 +75,7 @@ index 8bd5449f1..01e2df61a 100644
  }
  
  /**
-@@ -1560,6 +1561,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
+@@ -1570,6 +1571,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
   */
  static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
      ACL_CHECK(aclrtFreeHost(buffer->context));
@@ -84,10 +84,10 @@ index 8bd5449f1..01e2df61a 100644
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index bc396b521..aefc6935e 100644
+index fa7e1e13a..8f3b1c173 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -576,6 +576,7 @@ struct ggml_backend_cuda_buffer_context {
+@@ -579,6 +579,7 @@ struct ggml_backend_cuda_buffer_context {
  static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
      delete ctx;
@@ -95,7 +95,7 @@ index bc396b521..aefc6935e 100644
  }
  
  static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
-@@ -831,6 +832,7 @@ struct ggml_backend_cuda_split_buffer_context {
+@@ -834,6 +835,7 @@ struct ggml_backend_cuda_split_buffer_context {
  static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
      delete ctx;
@@ -103,7 +103,7 @@ index bc396b521..aefc6935e 100644
  }
  
  static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1112,6 +1114,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
+@@ -1115,6 +1117,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
  
  static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      CUDA_CHECK(cudaFreeHost(buffer->context));
@@ -112,7 +112,7 @@ index bc396b521..aefc6935e 100644
  
  static void * ggml_cuda_host_malloc(size_t size) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
-index 7afc881fa..bf0962274 100644
+index 70bf6f3d9..f2b7fe692 100644
 --- a/ggml/src/ggml-metal/ggml-metal.cpp
 +++ b/ggml/src/ggml-metal/ggml-metal.cpp
 @@ -25,6 +25,7 @@ static void ggml_backend_metal_buffer_shared_free_buffer(ggml_backend_buffer_t b
@@ -132,10 +132,10 @@ index 7afc881fa..bf0962274 100644
  
  static void * ggml_backend_metal_buffer_private_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
-index db33a4ab6..c42ee26e1 100644
+index e5302f455..43fa83e8f 100644
 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp
 +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
-@@ -3266,6 +3266,7 @@ struct ggml_backend_opencl_buffer_context {
+@@ -3412,6 +3412,7 @@ struct ggml_backend_opencl_buffer_context {
  static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
      delete ctx;
@@ -144,10 +144,10 @@ index db33a4ab6..c42ee26e1 100644
  
  static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
-index a38df5a97..fd07e4a21 100644
+index 48fd99a76..da2aab3df 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -528,6 +528,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -555,6 +555,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
      RPC_STATUS_ASSERT(status);
      delete ctx;
@@ -156,10 +156,10 @@ index a38df5a97..fd07e4a21 100644
  
  static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
-index b695ba051..37e853120 100644
+index 3f1bdfb9f..a95c2f305 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
-@@ -352,6 +352,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
+@@ -355,6 +355,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
      ggml_sycl_set_device(ctx->device);
  
      delete ctx;
@@ -167,7 +167,7 @@ index b695ba051..37e853120 100644
  }
  catch (sycl::exception const &exc) {
    std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-@@ -813,6 +814,7 @@ struct ggml_backend_sycl_split_buffer_context {
+@@ -816,6 +817,7 @@ struct ggml_backend_sycl_split_buffer_context {
  static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
      delete ctx;
@@ -175,7 +175,7 @@ index b695ba051..37e853120 100644
  }
  
  static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1155,6 +1157,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
+@@ -1158,6 +1160,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
  
  static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_sycl_host_free(buffer->context);
@@ -184,10 +184,10 @@ index b695ba051..37e853120 100644
  
  static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index b783f7805..216dc167c 100644
+index 66dd0bfab..83cdec29e 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -11828,6 +11828,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -12368,6 +12368,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
      ggml_vk_destroy_buffer(ctx->dev_buffer);
      delete ctx;
@@ -195,7 +195,7 @@ index b783f7805..216dc167c 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -11971,6 +11972,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -12511,6 +12512,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
  static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
      ggml_vk_host_free(vk_instance.devices[0], buffer->context);
diff --git a/llama/patches/0002-pretokenizer.patch b/llama/patches/0002-pretokenizer.patch
index 1a90f06d..096d5f4e 100644
--- a/llama/patches/0002-pretokenizer.patch
+++ b/llama/patches/0002-pretokenizer.patch
@@ -10,10 +10,10 @@ logs instead of throwing an error
  1 file changed, 3 insertions(+), 11 deletions(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 639fecbd3..a7ce6f8e1 100644
+index a73c4c448..b9f0631f4 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -1812,16 +1812,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1825,16 +1825,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
          if (type == LLAMA_VOCAB_TYPE_BPE) {
              add_space_prefix = false;
              clean_spaces = true;
@@ -31,8 +31,8 @@ index 639fecbd3..a7ce6f8e1 100644
                  pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -1993,7 +1984,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
-                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
+@@ -2014,7 +2005,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+                 pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
                  clean_spaces = false;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
diff --git a/llama/patches/0003-clip-unicode.patch b/llama/patches/0003-clip-unicode.patch
index c80a09d3..1f83a77e 100644
--- a/llama/patches/0003-clip-unicode.patch
+++ b/llama/patches/0003-clip-unicode.patch
@@ -10,11 +10,11 @@ filesystems for paths that include wide characters
  1 file changed, 39 insertions(+)
 
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index f2abf8852..c984e6282 100644
+index 05777d2d9..f4c4d2c48 100644
 --- a/tools/mtmd/clip.cpp
 +++ b/tools/mtmd/clip.cpp
-@@ -28,6 +28,19 @@
- #include <numeric>
+@@ -24,6 +24,19 @@
+ #include <array>
  #include <functional>
  
 +#if defined(_WIN32)
@@ -30,10 +30,10 @@ index f2abf8852..c984e6282 100644
 +#endif
 +#endif
 +
- struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
+ struct clip_logger_state g_logger_state = {clip_log_callback_default, NULL};
  
  enum ffn_op_type {
-@@ -2774,7 +2787,29 @@ struct clip_model_loader {
+@@ -3255,7 +3268,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index f2abf8852..c984e6282 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -2801,7 +2836,11 @@ struct clip_model_loader {
+@@ -3282,7 +3317,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }
diff --git a/llama/patches/0004-solar-pro.patch b/llama/patches/0004-solar-pro.patch
index 74bff354..82241b87 100644
--- a/llama/patches/0004-solar-pro.patch
+++ b/llama/patches/0004-solar-pro.patch
@@ -5,20 +5,36 @@ Subject: [PATCH] solar-pro
 
 adds support for the Solar Pro architecture
 ---
- src/llama-arch.cpp         |  21 ++++
+ src/CMakeLists.txt         |   1 +
+ src/llama-arch.cpp         |  21 +++++
  src/llama-arch.h           |   3 +
  src/llama-hparams.cpp      |   8 ++
- src/llama-hparams.h        |   5 +
+ src/llama-hparams.h        |   5 ++
  src/llama-model-loader.cpp |   2 +-
- src/llama-model.cpp        | 207 +++++++++++++++++++++++++++++++++++++
+ src/llama-model.cpp        |  48 +++++++++++
  src/llama-model.h          |   3 +
- 7 files changed, 248 insertions(+), 1 deletion(-)
+ src/models/models.h        |   5 ++
+ src/models/solar.cpp       | 158 +++++++++++++++++++++++++++++++++++++
+ 10 files changed, 253 insertions(+), 1 deletion(-)
+ create mode 100644 src/models/solar.cpp
 
+diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
+index 67c7807e0..fda881640 100644
+--- a/src/CMakeLists.txt
++++ b/src/CMakeLists.txt
+@@ -125,6 +125,7 @@ add_library(llama
+             models/seed-oss.cpp
+             models/smallthinker.cpp
+             models/smollm3.cpp
++            models/solar.cpp
+             models/stablelm.cpp
+             models/starcoder.cpp
+             models/starcoder2.cpp
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index 8ca769c5f..ab262ec0c 100644
+index 8571a2e02..b6bde25d5 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
-@@ -82,6 +82,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+@@ -85,6 +85,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
      { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
      { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
      { LLM_ARCH_CHAMELEON,        "chameleon"        },
@@ -26,7 +42,7 @@ index 8ca769c5f..ab262ec0c 100644
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-@@ -183,6 +184,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+@@ -204,6 +205,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
      { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
      { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
      { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
@@ -34,7 +50,7 @@ index 8ca769c5f..ab262ec0c 100644
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -1901,6 +1903,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -2023,6 +2025,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
          },
      },
@@ -59,7 +75,7 @@ index 8ca769c5f..ab262ec0c 100644
      {
          LLM_ARCH_WAVTOKENIZER_DEC,
          {
-@@ -2469,6 +2489,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -2681,6 +2701,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
@@ -68,10 +84,10 @@ index 8ca769c5f..ab262ec0c 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index dea725c1a..ea2b4ffb9 100644
+index 150646478..3936a4687 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -86,6 +86,7 @@ enum llm_arch {
+@@ -89,6 +89,7 @@ enum llm_arch {
      LLM_ARCH_GRANITE_MOE,
      LLM_ARCH_GRANITE_HYBRID,
      LLM_ARCH_CHAMELEON,
@@ -79,7 +95,7 @@ index dea725c1a..ea2b4ffb9 100644
      LLM_ARCH_WAVTOKENIZER_DEC,
      LLM_ARCH_PLM,
      LLM_ARCH_BAILINGMOE,
-@@ -187,6 +188,7 @@ enum llm_kv {
+@@ -208,6 +209,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_SCALE,
      LLM_KV_ATTENTION_OUTPUT_SCALE,
      LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
@@ -87,7 +103,7 @@ index dea725c1a..ea2b4ffb9 100644
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -436,6 +438,7 @@ enum llm_tensor {
+@@ -459,6 +461,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -96,11 +112,11 @@ index dea725c1a..ea2b4ffb9 100644
      LLM_TENSOR_CONVNEXT_DW,
      LLM_TENSOR_CONVNEXT_NORM,
 diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
-index db65d69ea..b6bf6bbf2 100644
+index 8cdbaf69f..41127bf91 100644
 --- a/src/llama-hparams.cpp
 +++ b/src/llama-hparams.cpp
-@@ -151,6 +151,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
-     return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+@@ -161,6 +161,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
+     return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
  }
  
 +bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
@@ -115,7 +131,7 @@ index db65d69ea..b6bf6bbf2 100644
      if (il < n_layer) {
          return swa_layers[il];
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index 6fcf91b7d..24569a258 100644
+index c3a53be79..2ffe7dd30 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
 @@ -64,6 +64,8 @@ struct llama_hparams {
@@ -127,7 +143,7 @@ index 6fcf91b7d..24569a258 100644
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
      uint32_t n_lora_kv          = 0;
-@@ -250,6 +252,9 @@ struct llama_hparams {
+@@ -256,6 +258,9 @@ struct llama_hparams {
  
      uint32_t n_pos_per_embd() const;
  
@@ -151,10 +167,10 @@ index aa3a65f87..ee303bd58 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 2a83d6627..54621ea39 100644
+index c2a545531..4468de2f9 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1890,6 +1890,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1961,6 +1961,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -176,7 +192,7 @@ index 2a83d6627..54621ea39 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -5224,6 +5239,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -5350,6 +5365,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
  
                          layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
  
@@ -211,12 +227,71 @@ index 2a83d6627..54621ea39 100644
                          layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                          layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
                          layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
-@@ -16515,6 +16558,165 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
-     }
+@@ -7425,6 +7468,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
+             {
+                 llm = std::make_unique<llm_build_chameleon>(*this, params);
+             } break;
++        case LLM_ARCH_SOLAR:
++            {
++                llm = std::make_unique<llm_build_solar>(*this, params);
++            } break;
+         case LLM_ARCH_WAVTOKENIZER_DEC:
+             {
+                 llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params);
+@@ -7684,6 +7731,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+         case LLM_ARCH_GRANITE_MOE:
+         case LLM_ARCH_GRANITE_HYBRID:
+         case LLM_ARCH_CHAMELEON:
++        case LLM_ARCH_SOLAR:
+         case LLM_ARCH_BAILINGMOE:
+         case LLM_ARCH_NEO_BERT:
+         case LLM_ARCH_SMOLLM3:
+diff --git a/src/llama-model.h b/src/llama-model.h
+index f8342cf2c..cbf4e1bfa 100644
+--- a/src/llama-model.h
++++ b/src/llama-model.h
+@@ -76,6 +76,7 @@ enum llm_type {
+     LLM_TYPE_15B,
+     LLM_TYPE_16B,
+     LLM_TYPE_20B,
++    LLM_TYPE_22B,
+     LLM_TYPE_26B,
+     LLM_TYPE_27B,
+     LLM_TYPE_30B,
+@@ -404,6 +405,8 @@ struct llama_layer {
+     struct ggml_tensor * ffn_act_beta    = nullptr;
+     struct ggml_tensor * ffn_act_eps     = nullptr;
+ 
++    struct ggml_tensor * bskcn_tv = nullptr;
++
+     struct llama_layer_posnet posnet;
+ 
+     struct llama_layer_convnext convnext;
+diff --git a/src/models/models.h b/src/models/models.h
+index 7ba225b47..71fea796d 100644
+--- a/src/models/models.h
++++ b/src/models/models.h
+@@ -510,6 +510,11 @@ struct llm_build_smollm3 : public llm_graph_context {
+     llm_build_smollm3(const llama_model & model, const llm_graph_params & params);
  };
  
 +struct llm_build_solar : public llm_graph_context {
-+    llm_build_solar(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
++    llm_build_solar(const llama_model & model, const llm_graph_params & params);
++};
++
++
+ struct llm_build_stablelm : public llm_graph_context {
+     llm_build_stablelm(const llama_model & model, const llm_graph_params & params);
+ };
+diff --git a/src/models/solar.cpp b/src/models/solar.cpp
+new file mode 100644
+index 000000000..97383928c
+--- /dev/null
++++ b/src/models/solar.cpp
+@@ -0,0 +1,158 @@
++#include "models.h"
++
++llm_build_solar::llm_build_solar(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
 +        const int64_t n_embd_head = hparams.n_embd_head_v;
 +        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 +        GGML_ASSERT(n_embd_head == hparams.n_rot);
@@ -285,7 +360,7 @@ index 2a83d6627..54621ea39 100644
 +                cb(Kcur, "Kcur", il);
 +                if (model.layers[il].bk) {
 +                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-+                    cb(Kcur, "Kcur", il);
++                   cb(Kcur, "Kcur", il);
 +                }
 +
 +                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
@@ -371,49 +446,4 @@ index 2a83d6627..54621ea39 100644
 +        res->t_logits = cur;
 +
 +        ggml_build_forward_expand(gf, cur);
-+    }
-+};
-+
- // ref: https://github.com/facebookresearch/chameleon
- // based on the original build_llama() function, changes:
- //   * qk-norm
-@@ -20096,6 +20298,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
-             {
-                 llm = std::make_unique<llm_build_chameleon>(*this, params);
-             } break;
-+        case LLM_ARCH_SOLAR:
-+            {
-+                llm = std::make_unique<llm_build_solar>(*this, params);
-+            } break;
-         case LLM_ARCH_WAVTOKENIZER_DEC:
-             {
-                 llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params);
-@@ -20331,6 +20537,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
-         case LLM_ARCH_GRANITE_MOE:
-         case LLM_ARCH_GRANITE_HYBRID:
-         case LLM_ARCH_CHAMELEON:
-+        case LLM_ARCH_SOLAR:
-         case LLM_ARCH_BAILINGMOE:
-         case LLM_ARCH_NEO_BERT:
-         case LLM_ARCH_SMOLLM3:
-diff --git a/src/llama-model.h b/src/llama-model.h
-index 248f85410..4a7924aaa 100644
---- a/src/llama-model.h
-+++ b/src/llama-model.h
-@@ -76,6 +76,7 @@ enum llm_type {
-     LLM_TYPE_15B,
-     LLM_TYPE_16B,
-     LLM_TYPE_20B,
-+    LLM_TYPE_22B,
-     LLM_TYPE_27B,
-     LLM_TYPE_30B,
-     LLM_TYPE_32B,
-@@ -390,6 +391,8 @@ struct llama_layer {
-     struct ggml_tensor * ffn_act_beta    = nullptr;
-     struct ggml_tensor * ffn_act_eps     = nullptr;
- 
-+    struct ggml_tensor * bskcn_tv = nullptr;
-+
-     struct llama_layer_posnet posnet;
- 
-     struct llama_layer_convnext convnext;
++}
diff --git a/llama/patches/0005-fix-deepseek-deseret-regex.patch b/llama/patches/0005-fix-deepseek-deseret-regex.patch
index 79debec5..0cebdb58 100644
--- a/llama/patches/0005-fix-deepseek-deseret-regex.patch
+++ b/llama/patches/0005-fix-deepseek-deseret-regex.patch
@@ -12,7 +12,7 @@ regex
  2 files changed, 22 insertions(+), 1 deletion(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index a7ce6f8e1..8064dc197 100644
+index b9f0631f4..1525283d7 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -299,7 +299,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
@@ -25,7 +25,7 @@ index a7ce6f8e1..8064dc197 100644
                      "\\s+$",
                      "[一-龥ࠀ-一가-퟿]+",
 diff --git a/src/unicode.cpp b/src/unicode.cpp
-index 65f366517..ce336a228 100644
+index 77ba4fc46..040518e1e 100644
 --- a/src/unicode.cpp
 +++ b/src/unicode.cpp
 @@ -2,6 +2,11 @@
diff --git a/llama/patches/0006-maintain-ordering-for-rules-for-grammar.patch b/llama/patches/0006-maintain-ordering-for-rules-for-grammar.patch
index 5ffe836d..66ac01c1 100644
--- a/llama/patches/0006-maintain-ordering-for-rules-for-grammar.patch
+++ b/llama/patches/0006-maintain-ordering-for-rules-for-grammar.patch
@@ -8,10 +8,10 @@ Subject: [PATCH] maintain ordering for rules for grammar
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
-index dd9b51a9e..d88f43209 100644
+index c8421e1e8..cb659915d 100644
 --- a/common/json-schema-to-grammar.cpp
 +++ b/common/json-schema-to-grammar.cpp
-@@ -308,7 +308,7 @@ private:
+@@ -310,7 +310,7 @@ private:
      friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
      std::function<json(const std::string &)> _fetch_json;
      bool _dotall;
diff --git a/llama/patches/0008-add-phony-target-ggml-cpu-for-all-cpu-variants.patch b/llama/patches/0008-add-phony-target-ggml-cpu-for-all-cpu-variants.patch
index 8fa52c1c..d3ab6500 100644
--- a/llama/patches/0008-add-phony-target-ggml-cpu-for-all-cpu-variants.patch
+++ b/llama/patches/0008-add-phony-target-ggml-cpu-for-all-cpu-variants.patch
@@ -8,10 +8,10 @@ Subject: [PATCH] add phony target ggml-cpu for all cpu variants
  1 file changed, 2 insertions(+)
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index ba281b8e6..ead235878 100644
+index d93664b8b..800f98b65 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -314,6 +314,7 @@ function(ggml_add_cpu_backend_variant tag_name)
+@@ -349,6 +349,7 @@ function(ggml_add_cpu_backend_variant tag_name)
      endif()
  
      ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -19,7 +19,7 @@ index ba281b8e6..ead235878 100644
  endfunction()
  
  ggml_add_backend(CPU)
-@@ -324,6 +325,7 @@ if (GGML_CPU_ALL_VARIANTS)
+@@ -359,6 +360,7 @@ if (GGML_CPU_ALL_VARIANTS)
      elseif (GGML_CPU_ARM_ARCH)
          message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
      endif()
diff --git a/llama/patches/0009-remove-amx.patch b/llama/patches/0009-remove-amx.patch
index 51a34bbc..bfb3727a 100644
--- a/llama/patches/0009-remove-amx.patch
+++ b/llama/patches/0009-remove-amx.patch
@@ -9,10 +9,10 @@ disable amx as it reduces performance on some systems
  1 file changed, 4 deletions(-)
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index ead235878..f9a6587f1 100644
+index 800f98b65..6d493a4ff 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -334,10 +334,6 @@ if (GGML_CPU_ALL_VARIANTS)
+@@ -369,10 +369,6 @@ if (GGML_CPU_ALL_VARIANTS)
          ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
          ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
          ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
diff --git a/llama/patches/0010-fix-string-arr-kv-loading.patch b/llama/patches/0010-fix-string-arr-kv-loading.patch
index c0cab979..ce151948 100644
--- a/llama/patches/0010-fix-string-arr-kv-loading.patch
+++ b/llama/patches/0010-fix-string-arr-kv-loading.patch
@@ -53,10 +53,10 @@ index 8cc4ef1cf..d950dbdf5 100644
  }
  
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 8064dc197..31f49801c 100644
+index 1525283d7..ea450c361 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -1768,9 +1768,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1781,9 +1781,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
              const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str());
              if (precompiled_charsmap_keyidx != -1) {
                  const gguf_type pc_type = gguf_get_arr_type(ctx, precompiled_charsmap_keyidx);
diff --git a/llama/patches/0011-ollama-debug-tensor.patch b/llama/patches/0011-ollama-debug-tensor.patch
index 6706c4ed..76db920f 100644
--- a/llama/patches/0011-ollama-debug-tensor.patch
+++ b/llama/patches/0011-ollama-debug-tensor.patch
@@ -8,7 +8,7 @@ Subject: [PATCH] ollama debug tensor
  1 file changed, 6 insertions(+)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index 9ec485cfa..4b2f8b7bd 100644
+index 3247af8bb..5be08d6f4 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
 @@ -15,6 +15,8 @@
@@ -20,7 +20,7 @@ index 9ec485cfa..4b2f8b7bd 100644
  #if defined(_MSC_VER) || defined(__MINGW32__)
  #include <malloc.h> // using malloc.h with MSC/MINGW
  #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
-@@ -2891,6 +2893,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+@@ -2922,6 +2924,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
  
          ggml_compute_forward(&params, node);
  
diff --git a/llama/patches/0012-add-ollama-vocab-for-grammar-support.patch b/llama/patches/0012-add-ollama-vocab-for-grammar-support.patch
index 82217739..e84bc875 100644
--- a/llama/patches/0012-add-ollama-vocab-for-grammar-support.patch
+++ b/llama/patches/0012-add-ollama-vocab-for-grammar-support.patch
@@ -6,14 +6,14 @@ Subject: [PATCH] add ollama vocab for grammar support
 ---
  src/llama-grammar.cpp  | 49 ++++++++++++++++++++++++++++++++++++------
  src/llama-grammar.h    | 14 ++++++++++++
- src/llama-sampling.cpp |  4 ++--
- 3 files changed, 58 insertions(+), 9 deletions(-)
+ src/llama-sampling.cpp |  6 +++---
+ 3 files changed, 59 insertions(+), 10 deletions(-)
 
 diff --git a/src/llama-grammar.cpp b/src/llama-grammar.cpp
-index bed706bb2..b51cee090 100644
+index b3c5eb571..a7307c47f 100644
 --- a/src/llama-grammar.cpp
 +++ b/src/llama-grammar.cpp
-@@ -907,6 +907,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
+@@ -915,6 +915,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
  
  struct llama_grammar * llama_grammar_init_impl(
          const struct llama_vocab * vocab,
@@ -21,7 +21,7 @@ index bed706bb2..b51cee090 100644
          const llama_grammar_element ** rules,
          size_t n_rules,
          size_t start_rule_index) {
-@@ -962,6 +963,7 @@ struct llama_grammar * llama_grammar_init_impl(
+@@ -970,6 +971,7 @@ struct llama_grammar * llama_grammar_init_impl(
      // then the pointers would be invalidated when the local vec_rules goes out of scope.
      return new llama_grammar {
          vocab,
@@ -29,7 +29,7 @@ index bed706bb2..b51cee090 100644
          std::move(vec_rules),
          std::move(stacks),
          /* .partial_utf8 = */     {},
-@@ -975,6 +977,7 @@ struct llama_grammar * llama_grammar_init_impl(
+@@ -983,6 +985,7 @@ struct llama_grammar * llama_grammar_init_impl(
  
  struct llama_grammar * llama_grammar_init_impl(
          const struct llama_vocab * vocab,
@@ -37,7 +37,7 @@ index bed706bb2..b51cee090 100644
                        const char * grammar_str,
                        const char * grammar_root,
                                bool lazy,
-@@ -1067,6 +1070,7 @@ struct llama_grammar * llama_grammar_init_impl(
+@@ -1075,6 +1078,7 @@ struct llama_grammar * llama_grammar_init_impl(
      // then the pointers would be invalidated when the local vec_rules goes out of scope.
      return new llama_grammar {
          vocab,
@@ -45,7 +45,7 @@ index bed706bb2..b51cee090 100644
          std::move(vec_rules),
          std::move(stacks),
          /* .partial_utf8 = */     {},
-@@ -1089,6 +1093,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
+@@ -1097,6 +1101,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
  struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
      auto * result = new llama_grammar {
          grammar.vocab,
@@ -53,7 +53,7 @@ index bed706bb2..b51cee090 100644
          grammar.rules,
          grammar.stacks,
          grammar.partial_utf8,
-@@ -1116,7 +1121,6 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
+@@ -1124,7 +1129,6 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
  }
  
  void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_data_array * cur_p) {
@@ -61,7 +61,7 @@ index bed706bb2..b51cee090 100644
  
      if (grammar.awaiting_trigger) {
          return;
-@@ -1138,9 +1142,13 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
+@@ -1146,9 +1150,13 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
  
      for (size_t i = 0; i < cur_p->size; ++i) {
          const llama_token id      = cur_p->data[i].id;
@@ -77,7 +77,7 @@ index bed706bb2..b51cee090 100644
              if (!allow_eog) {
                  cur_p->data[i].logit = -INFINITY;
              }
-@@ -1159,9 +1167,10 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
+@@ -1167,9 +1175,10 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
  }
  
  void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token) {
@@ -90,7 +90,7 @@ index bed706bb2..b51cee090 100644
  
      if (grammar.awaiting_trigger) {
          if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
-@@ -1201,13 +1210,14 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
+@@ -1209,13 +1218,14 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
          }
      }
  
@@ -107,7 +107,7 @@ index bed706bb2..b51cee090 100644
      }
  
      llama_grammar_accept_str(grammar, piece);
-@@ -1227,3 +1237,28 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
+@@ -1235,3 +1245,28 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
          throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece);
      }
  }
@@ -184,10 +184,10 @@ index f8c291de9..2a3a62db3 100644
                        const char * grammar_root,
                                bool lazy,
 diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp
-index 55d2e355f..da34526b1 100644
+index 3f4a729bc..38a30ea05 100644
 --- a/src/llama-sampling.cpp
 +++ b/src/llama-sampling.cpp
-@@ -1563,7 +1563,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
+@@ -1561,7 +1561,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
          trigger_patterns_c.push_back(trigger_pattern.pattern.c_str());
      }
  
@@ -196,12 +196,15 @@ index 55d2e355f..da34526b1 100644
                                                   ctx->grammar->lazy, trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                   ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
  
-@@ -1645,7 +1645,7 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
+@@ -1639,9 +1639,9 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
+             trigger_pattern += ")[\\s\\S]*";
+ 
+             std::array<const char *, 1> tmp_trigger_patterns = { trigger_pattern.c_str() };
+-            grammar = llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, tmp_trigger_patterns.data(), tmp_trigger_patterns.size(), trigger_tokens, num_trigger_tokens);
++            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, tmp_trigger_patterns.data(), tmp_trigger_patterns.size(), trigger_tokens, num_trigger_tokens);
+         } else {
+-            grammar = llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens);
++            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens);
+         }
+         *ctx = {
              /* .vocab        = */ vocab,
-             /* .grammar_str  = */ grammar_str,
-             /* .grammar_root = */ grammar_root,
--            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
-+            /* .grammar      = */ llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
-         };
-         if (!ctx->grammar) {
-             delete ctx;
diff --git a/llama/patches/0013-add-argsort-and-cuda-copy-for-i32.patch b/llama/patches/0013-add-argsort-and-cuda-copy-for-i32.patch
index ef4b359e..5e5bc110 100644
--- a/llama/patches/0013-add-argsort-and-cuda-copy-for-i32.patch
+++ b/llama/patches/0013-add-argsort-and-cuda-copy-for-i32.patch
@@ -8,14 +8,14 @@ Subject: [PATCH] add argsort and cuda copy for i32
  ggml/src/ggml-cuda/argsort.cu        | 122 ++++++++++++++++++++++++---
  ggml/src/ggml-cuda/cpy-utils.cuh     |   6 ++
  ggml/src/ggml-cuda/cpy.cu            |  40 +++++++++
- ggml/src/ggml-metal/ggml-metal.metal |  64 ++++++++++++++
- 5 files changed, 263 insertions(+), 12 deletions(-)
+ ggml/src/ggml-metal/ggml-metal.metal |  69 +++++++++++++++
+ 5 files changed, 268 insertions(+), 12 deletions(-)
 
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index b52f0f847..902fdad69 100644
+index 2745fc54e..40666bab6 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -7889,6 +7889,45 @@ static void ggml_compute_forward_argsort_f32(
+@@ -7846,6 +7846,45 @@ static void ggml_compute_forward_argsort_f32(
      }
  }
  
@@ -61,7 +61,7 @@ index b52f0f847..902fdad69 100644
  void ggml_compute_forward_argsort(
      const ggml_compute_params * params,
      ggml_tensor * dst) {
-@@ -7900,6 +7939,10 @@ void ggml_compute_forward_argsort(
+@@ -7857,6 +7896,10 @@ void ggml_compute_forward_argsort(
              {
                  ggml_compute_forward_argsort_f32(params, dst);
              } break;
@@ -73,7 +73,7 @@ index b52f0f847..902fdad69 100644
              {
                  GGML_ABORT("fatal error");
 diff --git a/ggml/src/ggml-cuda/argsort.cu b/ggml/src/ggml-cuda/argsort.cu
-index 6e7b90d42..08dd30525 100644
+index da9652c3b..b82be371c 100644
 --- a/ggml/src/ggml-cuda/argsort.cu
 +++ b/ggml/src/ggml-cuda/argsort.cu
 @@ -168,13 +168,107 @@ static void argsort_f32_i32_cuda_bitonic(const float *   x,
@@ -220,11 +220,11 @@ index 6e7b90d42..08dd30525 100644
 +    }
  }
 diff --git a/ggml/src/ggml-cuda/cpy-utils.cuh b/ggml/src/ggml-cuda/cpy-utils.cuh
-index e621cb981..597c0c8b3 100644
+index 7697c292d..00d773dd3 100644
 --- a/ggml/src/ggml-cuda/cpy-utils.cuh
 +++ b/ggml/src/ggml-cuda/cpy-utils.cuh
 @@ -215,3 +215,9 @@ template<typename src_t, typename dst_t>
- static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
+ static __device__ void cpy_1_scalar(const char * cxi, char * cdsti) {
      *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
  }
 +
@@ -234,10 +234,10 @@ index e621cb981..597c0c8b3 100644
 +    *dst = *src;
 +}
 diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
-index 12d5bf776..a0e34030e 100644
+index c4ceb4fc5..0e53ecc39 100644
 --- a/ggml/src/ggml-cuda/cpy.cu
 +++ b/ggml/src/ggml-cuda/cpy.cu
-@@ -251,6 +251,43 @@ static void ggml_cpy_f32_iq4_nl_cuda(
+@@ -352,6 +352,43 @@ static void ggml_cpy_f32_iq4_nl_cuda(
          (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
  }
  
@@ -281,73 +281,76 @@ index 12d5bf776..a0e34030e 100644
  void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
      const int64_t ne = ggml_nelements(src0);
      GGML_ASSERT(ne == ggml_nelements(src1));
-@@ -332,6 +369,9 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
-         ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-         ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+@@ -481,6 +518,9 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
+             ggml_cpy_scalar_cuda<half, float>
+                 (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+         }
 +    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
 +        // TODO consider converting to template
 +        ggml_cpy_i32_i32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
      } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-         ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+         if (can_be_transposed) {
+             ggml_cpy_scalar_cuda<nv_bfloat16, nv_bfloat16, true>
 diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
-index 2c2f01415..50b8071de 100644
+index 73b45c762..aed013a9d 100644
 --- a/ggml/src/ggml-metal/ggml-metal.metal
 +++ b/ggml/src/ggml-metal/ggml-metal.metal
-@@ -4467,8 +4467,72 @@ kernel void kernel_argsort_f32_i32(
+@@ -4721,8 +4721,77 @@ kernel void kernel_argsort_f32_i32(
      }
  }
  
 +typedef void (i32_argsort_t)(
 +        constant   ggml_metal_kargs_argsort & args,
-+        device  const int32_t * x,
++        device  const int32_t * src0,
 +        device        int32_t * dst,
-+        threadgroup   int32_t * shared_values [[threadgroup(0)]],
-+        uint3 tgpig[[threadgroup_position_in_grid]],
-+        uint3 tpitg[[thread_position_in_threadgroup]]);
++        threadgroup   int32_t * shmem_i32 [[threadgroup(0)]],
++        uint3   tgpig[[threadgroup_position_in_grid]],
++        ushort3 tpitg[[thread_position_in_threadgroup]],
++        ushort3   ntg[[threads_per_threadgroup]]);
 +
 +template<ggml_sort_order order>
 +kernel void kernel_argsort_i32_i32(
 +        constant   ggml_metal_kargs_argsort & args,
-+        device const int32_t * x,
++        device const int32_t * src0,
 +        device       int32_t * dst,
-+        threadgroup int32_t  * shared_values [[threadgroup(0)]],
-+        uint3 tgpig[[threadgroup_position_in_grid]],
-+        uint3 tpitg[[thread_position_in_threadgroup]]) {
++        threadgroup int32_t  * shmem_i32 [[threadgroup(0)]],
++        uint3   tgpig[[threadgroup_position_in_grid]],
++        ushort3 tpitg[[thread_position_in_threadgroup]],
++        ushort3   ntg[[threads_per_threadgroup]]) {
 +    // bitonic sort
-+    int col = tpitg[0];
-+    int row = tgpig[1];
++    const int col = tpitg[0];
 +
-+    if (col >= args.ncols_pad) return;
++    const int i00 = (tgpig[0]/args.ne01)*ntg.x;
++    const int i01 =  tgpig[0]%args.ne01;
++    const int i02 =  tgpig[1];
++    const int i03 =  tgpig[2];
 +
-+    device const int32_t * x_row   = x + row * args.ncols;
-+    threadgroup int32_t  * dst_row = shared_values;
++    device const int32_t * src0_row = (device const int32_t *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 +
 +    // initialize indices
-+    dst_row[col] = col;
++    shmem_i32[col] = i00 + col;
 +
 +    threadgroup_barrier(mem_flags::mem_threadgroup);
 +
-+    for (int k = 2; k <= args.ncols_pad; k *= 2) {
++    for (int k = 2; k <= ntg.x; k *= 2) {
 +        for (int j = k / 2; j > 0; j /= 2) {
 +            int ixj = col ^ j;
 +            if (ixj > col) {
 +                if ((col & k) == 0) {
-+                    if (dst_row[col] >= args.ncols ||
-+                        (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
++                    if (shmem_i32[col] >= args.ne00 ||
++                        (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
++                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
++                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
 +                    ) {
-+                        SWAP(dst_row[col], dst_row[ixj]);
++                        SWAP(shmem_i32[col], shmem_i32[ixj]);
 +                    }
 +                } else {
-+                    if (dst_row[ixj] >= args.ncols ||
-+                        (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
++                    if (shmem_i32[ixj] >= args.ne00 ||
++                        (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
++                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
++                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
 +                    ) {
-+                        SWAP(dst_row[col], dst_row[ixj]);
++                        SWAP(shmem_i32[col], shmem_i32[ixj]);
 +                    }
 +                }
 +            }
@@ -356,8 +359,10 @@ index 2c2f01415..50b8071de 100644
 +    }
 +
 +    // copy the result to dst without the padding
-+    if (col < args.ncols) {
-+        dst[row * args.ncols + col] = dst_row[col];
++    if (i00 + col < args.ne00) {
++        dst += i00 + args.ne00*i01 + args.ne00*args.ne01*i02 + args.ne00*args.ne01*args.ne02*i03;
++
++        dst[col] = shmem_i32[col];
 +    }
 +}
 +
@@ -366,5 +371,5 @@ index 2c2f01415..50b8071de 100644
 +template [[host_name("kernel_argsort_i32_i32_asc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_ASC>;
 +template [[host_name("kernel_argsort_i32_i32_desc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_DESC>;
  
- kernel void kernel_leaky_relu_f32(
-         constant     ggml_metal_kargs_leaky_relu & args,
+ typedef void (argsort_merge_t)(
+         constant   ggml_metal_kargs_argsort_merge & args,
diff --git a/llama/patches/0014-graph-memory-reporting-on-failure.patch b/llama/patches/0014-graph-memory-reporting-on-failure.patch
index b657a398..fdb462c9 100644
--- a/llama/patches/0014-graph-memory-reporting-on-failure.patch
+++ b/llama/patches/0014-graph-memory-reporting-on-failure.patch
@@ -35,10 +35,10 @@ index f1b740785..c54ff98bf 100644
      GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
      GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
-index c830c0965..363853873 100644
+index 218222ece..06ee502ab 100644
 --- a/ggml/src/ggml-alloc.c
 +++ b/ggml/src/ggml-alloc.c
-@@ -486,6 +486,7 @@ struct node_alloc {
+@@ -493,6 +493,7 @@ struct node_alloc {
  struct ggml_gallocr {
      ggml_backend_buffer_type_t * bufts; // [n_buffers]
      struct vbuffer ** buffers; // [n_buffers]
@@ -46,7 +46,7 @@ index c830c0965..363853873 100644
      struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
      int n_buffers;
  
-@@ -509,6 +510,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
+@@ -516,6 +517,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
      galloc->buffers = calloc(n_bufs, sizeof(struct vbuffer *));
      GGML_ASSERT(galloc->buffers != NULL);
  
@@ -56,7 +56,7 @@ index c830c0965..363853873 100644
      galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
      GGML_ASSERT(galloc->buf_tallocs != NULL);
  
-@@ -576,6 +580,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
+@@ -583,6 +587,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
      ggml_hash_set_free(&galloc->hash_set);
      free(galloc->hash_values);
      free(galloc->bufts);
@@ -64,7 +64,7 @@ index c830c0965..363853873 100644
      free(galloc->buffers);
      free(galloc->buf_tallocs);
      free(galloc->node_allocs);
-@@ -891,6 +896,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+@@ -898,6 +903,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
          }
      }
  
@@ -73,8 +73,8 @@ index c830c0965..363853873 100644
      // reallocate buffers if needed
      for (int i = 0; i < galloc->n_buffers; i++) {
          // if the buffer type is used multiple times, we reuse the same buffer
-@@ -920,14 +927,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
- 
+@@ -932,14 +939,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+ #endif
              ggml_vbuffer_free(galloc->buffers[i]);
              galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
 -            if (galloc->buffers[i] == NULL) {
@@ -96,7 +96,7 @@ index c830c0965..363853873 100644
  }
  
  bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-@@ -1082,6 +1094,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+@@ -1094,6 +1106,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
      return ggml_vbuffer_size(galloc->buffers[buffer_id]);
  }
  
@@ -120,10 +120,10 @@ index c830c0965..363853873 100644
  
  static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 8ba86f824..cb2b99562 100644
+index 4882541c8..ff41c7712 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -1809,6 +1809,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
+@@ -1813,6 +1813,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
      return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
  }
  
diff --git a/llama/patches/0015-ggml-Export-GPU-UUIDs.patch b/llama/patches/0015-ggml-Export-GPU-UUIDs.patch
index 28c11241..c3836536 100644
--- a/llama/patches/0015-ggml-Export-GPU-UUIDs.patch
+++ b/llama/patches/0015-ggml-Export-GPU-UUIDs.patch
@@ -1,10 +1,8 @@
 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Jesse Gross <jesse@ollama.com>
-Date: Thu, 24 Apr 2025 14:48:51 -0700
+From: Daniel Hiltgen <daniel@ollama.com>
+Date: Sun, 30 Nov 2025 11:05:56 -0800
 Subject: [PATCH] ggml: Export GPU UUIDs
 
-This enables matching up devices and information reported by the backend
-with tools (e.g. nvidia-smi) and system management libraries (e.g. nvml).
 ---
  ggml/include/ggml-backend.h        |  1 +
  ggml/src/ggml-cuda/ggml-cuda.cu    | 67 +++++++++++++++++++++++++++---
@@ -24,10 +22,10 @@ index c54ff98bf..229bf387b 100644
          size_t memory_total;
          // device type
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index aefc6935e..cc201afff 100644
+index 8f3b1c173..e803f4af6 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -183,6 +183,51 @@ static int ggml_cuda_parse_id(char devName[]) {
+@@ -185,6 +185,51 @@ static int ggml_cuda_parse_id(char devName[]) {
  }
  #endif // defined(GGML_USE_HIP)
  
@@ -79,7 +77,7 @@ index aefc6935e..cc201afff 100644
  static ggml_cuda_device_info ggml_cuda_init() {
      ggml_cuda_device_info info = {};
  
-@@ -249,22 +294,24 @@ static ggml_cuda_device_info ggml_cuda_init() {
+@@ -251,22 +296,24 @@ static ggml_cuda_device_info ggml_cuda_init() {
                  info.devices[id].cc += prop.minor * 0x10;
              }
          }
@@ -110,7 +108,7 @@ index aefc6935e..cc201afff 100644
          std::string device_name(prop.name);
          if (device_name == "NVIDIA GeForce MX450") {
              turing_devices_without_mma.push_back({ id, device_name });
-@@ -3268,6 +3315,7 @@ struct ggml_backend_cuda_device_context {
+@@ -4048,6 +4095,7 @@ struct ggml_backend_cuda_device_context {
      std::string name;
      std::string description;
      std::string pci_bus_id;
@@ -118,9 +116,9 @@ index aefc6935e..cc201afff 100644
  };
  
  static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
-@@ -3280,6 +3328,11 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
-     return ctx->description.c_str();
+@@ -4136,6 +4184,11 @@ static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_k
  }
+ #endif // defined(__linux__)
  
 +static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
 +    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
@@ -130,7 +128,7 @@ index aefc6935e..cc201afff 100644
  static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
      ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
      ggml_cuda_set_device(ctx->device);
-@@ -3296,6 +3349,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+@@ -4176,6 +4229,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
  
      props->name        = ggml_backend_cuda_device_get_name(dev);
      props->description = ggml_backend_cuda_device_get_description(dev);
@@ -138,7 +136,7 @@ index aefc6935e..cc201afff 100644
      props->type        = ggml_backend_cuda_device_get_type(dev);
      props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
      ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
-@@ -3869,6 +3923,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -4767,6 +4821,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  cudaDeviceProp prop;
                  CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                  dev_ctx->description = prop.name;
@@ -147,10 +145,10 @@ index aefc6935e..cc201afff 100644
                  char pci_bus_id[16] = {};
                  snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
 diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
-index bf0962274..f2ff9f322 100644
+index f2b7fe692..8fc1c2fb5 100644
 --- a/ggml/src/ggml-metal/ggml-metal.cpp
 +++ b/ggml/src/ggml-metal/ggml-metal.cpp
-@@ -538,6 +538,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
+@@ -547,6 +547,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
  static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
      props->name        = ggml_backend_metal_device_get_name(dev);
      props->description = ggml_backend_metal_device_get_description(dev);
diff --git a/llama/patches/0016-add-C-API-for-mtmd_input_text.patch b/llama/patches/0016-add-C-API-for-mtmd_input_text.patch
index a2efcbab..fa371e8e 100644
--- a/llama/patches/0016-add-C-API-for-mtmd_input_text.patch
+++ b/llama/patches/0016-add-C-API-for-mtmd_input_text.patch
@@ -10,10 +10,10 @@ Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
  2 files changed, 13 insertions(+)
 
 diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
-index 4d487581a..35a0d25ed 100644
+index dfad9cd79..9858de630 100644
 --- a/tools/mtmd/mtmd.cpp
 +++ b/tools/mtmd/mtmd.cpp
-@@ -79,6 +79,16 @@ enum mtmd_slice_tmpl {
+@@ -87,6 +87,16 @@ enum mtmd_slice_tmpl {
      MTMD_SLICE_TMPL_IDEFICS3,
  };
  
@@ -31,7 +31,7 @@ index 4d487581a..35a0d25ed 100644
      return "<__media__>";
  }
 diff --git a/tools/mtmd/mtmd.h b/tools/mtmd/mtmd.h
-index f4ea07d3a..cf287224b 100644
+index 015119be8..8d3fa5d34 100644
 --- a/tools/mtmd/mtmd.h
 +++ b/tools/mtmd/mtmd.h
 @@ -75,6 +75,9 @@ typedef struct mtmd_input_chunk  mtmd_input_chunk;
diff --git a/llama/patches/0017-no-power-throttling-win32-with-gnuc.patch b/llama/patches/0017-no-power-throttling-win32-with-gnuc.patch
index fa98defd..549e48fa 100644
--- a/llama/patches/0017-no-power-throttling-win32-with-gnuc.patch
+++ b/llama/patches/0017-no-power-throttling-win32-with-gnuc.patch
@@ -8,10 +8,10 @@ Subject: [PATCH] no power throttling win32 with gnuc
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index 4b2f8b7bd..046646282 100644
+index 5be08d6f4..7a0df30c3 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
-@@ -2441,7 +2441,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
+@@ -2463,7 +2463,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
          // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
          // all our threads onto the first 4 cores which results in terrible performance with
          // n_threads > 4
diff --git a/llama/patches/0018-ggml-Add-batch-size-hint.patch b/llama/patches/0018-ggml-Add-batch-size-hint.patch
index e9629a7d..f917f397 100644
--- a/llama/patches/0018-ggml-Add-batch-size-hint.patch
+++ b/llama/patches/0018-ggml-Add-batch-size-hint.patch
@@ -58,7 +58,7 @@ index 6792ba986..0f5b03cef 100644
          // (optional) event synchronization
          // record an event on this stream
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index cb2b99562..41eef3b5f 100644
+index ff41c7712..f511e8d76 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -348,14 +348,14 @@ enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_ba
@@ -97,7 +97,7 @@ index cb2b99562..41eef3b5f 100644
                  for (int b = 0; b < src_backend_id; b++) {
                      if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                          SET_CAUSE(tensor, "1.off");
-@@ -1550,7 +1552,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
+@@ -1556,7 +1558,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
          }
  
          if (!sched->callback_eval) {
@@ -106,7 +106,7 @@ index cb2b99562..41eef3b5f 100644
              if (ec != GGML_STATUS_SUCCESS) {
                  return ec;
              }
-@@ -1572,7 +1574,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
+@@ -1578,7 +1580,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
  
                  struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
  
@@ -115,7 +115,7 @@ index cb2b99562..41eef3b5f 100644
                  if (ec != GGML_STATUS_SUCCESS) {
                      return ec;
                  }
-@@ -1651,6 +1653,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1657,6 +1659,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
  
      sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
      sched->op_offload = op_offload;
@@ -123,7 +123,7 @@ index cb2b99562..41eef3b5f 100644
  
      ggml_backend_sched_reset(sched);
  
-@@ -1682,6 +1685,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+@@ -1688,6 +1691,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
      free(sched);
  }
  
@@ -178,10 +178,10 @@ index 3191faaa4..32f14c811 100644
  
  static const struct ggml_backend_i ggml_backend_cpu_i = {
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index cc201afff..02d413467 100644
+index e803f4af6..78fb2d8b3 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -2693,7 +2693,7 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
+@@ -2885,7 +2885,7 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
  
  #ifdef USE_CUDA_GRAPH
  static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
@@ -190,7 +190,7 @@ index cc201afff..02d413467 100644
  
      // Loop over nodes in GGML graph to obtain info needed for CUDA graph
  
-@@ -2726,24 +2726,34 @@ static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
+@@ -2918,24 +2918,34 @@ static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
  #endif
          }
  
@@ -241,7 +241,7 @@ index cc201afff..02d413467 100644
          }
  
          if (!use_cuda_graph) {
-@@ -3128,7 +3138,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
+@@ -3679,7 +3689,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
      }
  }
  
@@ -250,7 +250,7 @@ index cc201afff..02d413467 100644
      ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
  
      ggml_cuda_set_device(cuda_ctx->device);
-@@ -3166,7 +3176,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
+@@ -3717,7 +3727,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
      if (use_cuda_graph) {
          cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
  
@@ -260,10 +260,10 @@ index cc201afff..02d413467 100644
          // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
          if (use_cuda_graph && cuda_graph_update_required) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
-index f2ff9f322..05ff6a5a6 100644
+index 8fc1c2fb5..ba95b4acc 100644
 --- a/ggml/src/ggml-metal/ggml-metal.cpp
 +++ b/ggml/src/ggml-metal/ggml-metal.cpp
-@@ -410,10 +410,12 @@ static bool ggml_backend_metal_cpy_tensor_async(ggml_backend_t backend_src, ggml
+@@ -419,10 +419,12 @@ static bool ggml_backend_metal_cpy_tensor_async(ggml_backend_t backend_src, ggml
      GGML_UNUSED(dst);
  }
  
@@ -278,10 +278,10 @@ index f2ff9f322..05ff6a5a6 100644
  
  static void ggml_backend_metal_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 216dc167c..3a6bbe564 100644
+index 83cdec29e..a36c6560c 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -12357,7 +12357,7 @@ static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const stru
+@@ -13103,7 +13103,7 @@ static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const stru
      return num_adds;
  }
  
@@ -290,7 +290,7 @@ index 216dc167c..3a6bbe564 100644
      VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
      ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
  
-@@ -12561,6 +12561,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
+@@ -13320,6 +13320,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
      return GGML_STATUS_SUCCESS;
  
      UNUSED(backend);
diff --git a/llama/patches/0020-ggml-No-alloc-mode.patch b/llama/patches/0020-ggml-No-alloc-mode.patch
index 48dda776..0dff5573 100644
--- a/llama/patches/0020-ggml-No-alloc-mode.patch
+++ b/llama/patches/0020-ggml-No-alloc-mode.patch
@@ -12,8 +12,8 @@ must be recreated with no-alloc set to false before loading data.
  ggml/src/ggml-backend-impl.h    |  16 +++
  ggml/src/ggml-backend.cpp       |  72 ++++++++++-
  ggml/src/ggml-cuda/common.cuh   |  58 ++++++++-
- ggml/src/ggml-cuda/ggml-cuda.cu | 217 ++++++++++++++++++++++++++------
- 5 files changed, 320 insertions(+), 44 deletions(-)
+ ggml/src/ggml-cuda/ggml-cuda.cu | 218 ++++++++++++++++++++++++++------
+ 5 files changed, 321 insertions(+), 44 deletions(-)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
 index 2763f2bd6..b3b5b356a 100644
@@ -75,7 +75,7 @@ index 0f5b03cef..7bdf9d81f 100644
  
      struct ggml_backend {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 41eef3b5f..c81a2e48a 100644
+index f511e8d76..74b7f070c 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -41,6 +41,19 @@ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t
@@ -134,7 +134,7 @@ index 41eef3b5f..c81a2e48a 100644
  };
  
  #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
-@@ -1608,6 +1634,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1614,6 +1640,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
          size_t graph_size,
          bool parallel,
          bool op_offload) {
@@ -152,7 +152,7 @@ index 41eef3b5f..c81a2e48a 100644
      GGML_ASSERT(n_backends > 0);
      GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
      GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
-@@ -1649,11 +1686,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1655,11 +1692,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
                  sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
              }
          }
@@ -167,7 +167,7 @@ index 41eef3b5f..c81a2e48a 100644
  
      ggml_backend_sched_reset(sched);
  
-@@ -1668,6 +1708,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+@@ -1674,6 +1714,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
          for (int c = 0; c < sched->n_copies; c++) {
              ggml_backend_event_free(sched->events[b][c]);
          }
@@ -178,7 +178,7 @@ index 41eef3b5f..c81a2e48a 100644
      }
      ggml_gallocr_free(sched->galloc);
      ggml_free(sched->ctx);
-@@ -1715,6 +1759,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
+@@ -1719,6 +1763,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
          return false;
      }
  
@@ -203,7 +203,7 @@ index 41eef3b5f..c81a2e48a 100644
      ggml_backend_sched_reset(sched);
  
      return true;
-@@ -1820,7 +1882,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
+@@ -1824,7 +1886,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
      int backend_index = ggml_backend_sched_backend_id(sched, backend);
      GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
  
@@ -219,10 +219,10 @@ index 41eef3b5f..c81a2e48a 100644
  
  void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
 diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
-index 41ff89c4d..2931c15ca 100644
+index 611341deb..c3f8ca914 100644
 --- a/ggml/src/ggml-cuda/common.cuh
 +++ b/ggml/src/ggml-cuda/common.cuh
-@@ -35,6 +35,41 @@
+@@ -37,6 +37,41 @@
  #include "vendors/cuda.h"
  #endif // defined(GGML_USE_HIP)
  
@@ -264,7 +264,7 @@ index 41ff89c4d..2931c15ca 100644
  #define STRINGIZE_IMPL(...) #__VA_ARGS__
  #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
  
-@@ -856,6 +891,9 @@ struct ggml_cuda_pool {
+@@ -891,6 +926,9 @@ struct ggml_cuda_pool {
  
      virtual void * alloc(size_t size, size_t * actual_size) = 0;
      virtual void free(void * ptr, size_t size) = 0;
@@ -274,46 +274,48 @@ index 41ff89c4d..2931c15ca 100644
  };
  
  template<typename T>
-@@ -992,11 +1030,11 @@ struct ggml_backend_cuda_context {
+@@ -1179,11 +1217,11 @@ struct ggml_backend_cuda_context {
      // pool
-     std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
+     std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];
  
--    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
-+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, bool alloc);
+-    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, int stream_no);
++    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, int stream_no, bool alloc);
  
      ggml_cuda_pool & pool(int device) {
-         if (pools[device] == nullptr) {
--            pools[device] = new_pool_for_device(device);
-+            pools[device] = new_pool_for_device(device, true);
+         if (pools[device][curr_stream_no] == nullptr) {
+-            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no);
++            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no, true);
          }
-         return *pools[device];
+         return *pools[device][curr_stream_no];
      }
-@@ -1004,4 +1042,20 @@ struct ggml_backend_cuda_context {
+@@ -1191,6 +1229,22 @@ struct ggml_backend_cuda_context {
      ggml_cuda_pool & pool() {
          return pool(device);
      }
 +
 +    void pool_set_alloc(bool alloc) {
-+        GGML_ASSERT(pools[device] == nullptr || pools[device]->alloc_memory() == alloc);
++        GGML_ASSERT(pools[device][curr_stream_no] == nullptr || pools[device][curr_stream_no]->alloc_memory() == alloc);
 +
-+        if (pools[device] == nullptr) {
-+            pools[device] = new_pool_for_device(device, alloc);
++        if (pools[device][curr_stream_no] == nullptr) {
++            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no, alloc);
 +        }
 +    }
 +
 +    size_t pool_get_alloc_size() {
-+        if (pools[device] == nullptr) {
++        if (pools[device][curr_stream_no] == nullptr) {
 +            return 0;
 +        }
 +
-+        return pools[device]->alloc_size();
++        return pools[device][curr_stream_no]->alloc_size();
 +    }
  };
+ 
+ struct ggml_cuda_mm_fusion_args_host {
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 02d413467..f79e5d65c 100644
+index 78fb2d8b3..fe0da71ca 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -359,6 +359,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {
+@@ -361,6 +361,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {
  
  // #define DEBUG_CUDA_MALLOC
  
@@ -322,7 +324,7 @@ index 02d413467..f79e5d65c 100644
  // buffer pool for cuda (legacy)
  struct ggml_cuda_pool_leg : public ggml_cuda_pool {
      static const int MAX_BUFFERS = 256;
-@@ -371,9 +373,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -373,9 +375,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
  
      ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
      size_t pool_size = 0;
@@ -337,7 +339,7 @@ index 02d413467..f79e5d65c 100644
      }
  
      ~ggml_cuda_pool_leg() {
-@@ -381,7 +386,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -383,7 +388,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
          for (int i = 0; i < MAX_BUFFERS; ++i) {
              ggml_cuda_buffer & b = buffer_pool[i];
              if (b.ptr != nullptr) {
@@ -348,7 +350,7 @@ index 02d413467..f79e5d65c 100644
                  pool_size -= b.size;
              }
          }
-@@ -429,8 +436,15 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -431,8 +438,15 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
          void * ptr;
          size_t look_ahead_size = (size_t) (1.05 * size);
          look_ahead_size = 256 * ((look_ahead_size + 255)/256);
@@ -366,7 +368,7 @@ index 02d413467..f79e5d65c 100644
          *actual_size = look_ahead_size;
          pool_size += look_ahead_size;
  #ifdef DEBUG_CUDA_MALLOC
-@@ -450,10 +464,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -452,10 +466,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
              }
          }
          GGML_LOG_DEBUG(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
@@ -389,7 +391,7 @@ index 02d413467..f79e5d65c 100644
  };
  
  // pool with virtual memory
-@@ -465,18 +489,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
+@@ -467,18 +491,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
      CUdeviceptr pool_addr = 0;
      size_t pool_used = 0;
      size_t pool_size = 0;
@@ -417,7 +419,7 @@ index 02d413467..f79e5d65c 100644
  #if defined(GGML_USE_HIP)
              // Workaround for https://github.com/ROCm/ROCR-Runtime/issues/285
              for (std::pair<CUdeviceptr, size_t> & mapping : mappings) {
-@@ -503,35 +533,49 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
+@@ -505,35 +535,49 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
  
              GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
  
@@ -493,7 +495,7 @@ index 02d413467..f79e5d65c 100644
  
              // add to the pool
              pool_size += reserve_size;
-@@ -564,16 +608,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
+@@ -566,17 +610,27 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
          // all deallocations must be in reverse order of the allocations
          GGML_ASSERT(ptr == (void *) ((char *)(pool_addr) + pool_used));
      }
@@ -505,11 +507,14 @@ index 02d413467..f79e5d65c 100644
 +    size_t alloc_size() override {
 +        return pool_size + last_alloc;
 +    }
++
  };
  #endif // defined(GGML_USE_VMM)
  
--std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device, bool alloc) {
+ std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int                  device,
+-                                                                               [[maybe_unused]] int stream_no) {
++                                                                               [[maybe_unused]] int stream_no,
++                                                                               bool alloc) {
  #if defined(GGML_USE_VMM)
      if (ggml_cuda_info().devices[device].vmm) {
 -        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
@@ -521,7 +526,7 @@ index 02d413467..f79e5d65c 100644
  }
  
  // destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
-@@ -757,11 +809,20 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
+@@ -760,11 +814,20 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
  }
  
  static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -543,7 +548,7 @@ index 02d413467..f79e5d65c 100644
  static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
      size_t size = ggml_nbytes(tensor);
      int64_t ne0 = tensor->ne[0];
-@@ -785,6 +846,7 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
+@@ -788,6 +851,7 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
      /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
      /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
      /* .is_host          = */ NULL,
@@ -551,7 +556,7 @@ index 02d413467..f79e5d65c 100644
  };
  
  ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
-@@ -2986,6 +3048,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
+@@ -3258,6 +3322,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
  
  static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
      bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
@@ -559,7 +564,7 @@ index 02d413467..f79e5d65c 100644
      // flag used to determine whether it is an integrated_gpu
      const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
  
-@@ -3001,6 +3064,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
+@@ -3347,6 +3412,10 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                      continue;
                  }
  
@@ -567,11 +572,10 @@ index 02d413467..f79e5d65c 100644
 +                if (reserving_graph && node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
 +                    continue;
 +                }
-+
-                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
-                 if (!disable_fusion) {
  
-@@ -3140,6 +3208,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
+                 // start of fusion operations
+                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
+@@ -3691,6 +3760,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
  
  static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph, int batch_size) {
      ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
@@ -579,7 +583,7 @@ index 02d413467..f79e5d65c 100644
  
      ggml_cuda_set_device(cuda_ctx->device);
  
-@@ -3215,6 +3284,71 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
+@@ -3766,6 +3836,71 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
      return GGML_STATUS_SUCCESS;
  }
  
@@ -645,16 +649,16 @@ index 02d413467..f79e5d65c 100644
 +
 +static void ggml_backend_cuda_reset(ggml_backend_t backend) {
 +    ggml_backend_cuda_context * ctx = (ggml_backend_cuda_context *)backend->context;
-+    ctx->pools[ctx->device] = NULL;
++    ctx->pools[ctx->device][ctx->curr_stream_no] = NULL;
 +}
 +
  static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
      ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
  
-@@ -3255,6 +3389,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
+@@ -4035,6 +4170,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
      /* .event_record            = */ ggml_backend_cuda_event_record,
      /* .event_wait              = */ ggml_backend_cuda_event_wait,
-     /* .graph_optimize          = */ NULL,
+     /* .graph_optimize          = */ ggml_backend_cuda_graph_optimize,
 +    /* .graph_reserve           = */ ggml_backend_cuda_graph_reserve,
 +    /* .buffer_size             = */ ggml_backend_cuda_buffer_size,
 +    /* .reset                   = */ ggml_backend_cuda_reset,
diff --git a/llama/patches/0021-decode-disable-output_all.patch b/llama/patches/0021-decode-disable-output_all.patch
index 6b32e0da..c92e3910 100644
--- a/llama/patches/0021-decode-disable-output_all.patch
+++ b/llama/patches/0021-decode-disable-output_all.patch
@@ -8,12 +8,12 @@ Subject: [PATCH] decode: disable output_all
  1 file changed, 1 insertion(+), 2 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index bd348bcad..8b4a89d38 100644
+index e04f0fc4f..1359c614b 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -974,8 +974,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
+@@ -999,8 +999,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
      const int64_t n_vocab = vocab.n_tokens();
-     const int64_t n_embd  = hparams.n_embd;
+     const int64_t n_embd  = hparams.n_embd_inp();
  
 -    // when computing embeddings, all tokens are output
 -    const bool output_all = cparams.embeddings;
diff --git a/llama/patches/0022-ggml-Enable-resetting-backend-devices.patch b/llama/patches/0022-ggml-Enable-resetting-backend-devices.patch
index 54b2754b..c65d84f7 100644
--- a/llama/patches/0022-ggml-Enable-resetting-backend-devices.patch
+++ b/llama/patches/0022-ggml-Enable-resetting-backend-devices.patch
@@ -43,7 +43,7 @@ index 7bdf9d81f..21b35ac5c 100644
  
      struct ggml_backend_device {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index c81a2e48a..9b0a9b91f 100644
+index 74b7f070c..8d2cc167f 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -526,6 +526,14 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
@@ -62,10 +62,10 @@ index c81a2e48a..9b0a9b91f 100644
      GGML_ASSERT(device);
      return device->iface.get_buffer_type(device);
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index f79e5d65c..c9333689f 100644
+index fe0da71ca..0787e443c 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -107,6 +107,11 @@ int ggml_cuda_get_device() {
+@@ -109,6 +109,11 @@ int ggml_cuda_get_device() {
      return id;
  }
  
@@ -77,7 +77,7 @@ index f79e5d65c..c9333689f 100644
  static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
      ggml_cuda_set_device(device);
      cudaError_t err;
-@@ -3499,7 +3504,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+@@ -4380,7 +4385,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
      props->id          = ggml_backend_cuda_device_get_id(dev);
      props->type        = ggml_backend_cuda_device_get_type(dev);
      props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
@@ -89,7 +89,7 @@ index f79e5d65c..c9333689f 100644
  
      bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
  #ifdef GGML_CUDA_NO_PEER_COPY
-@@ -3936,6 +3944,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
+@@ -4835,6 +4843,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
      CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
  }
  
@@ -101,7 +101,7 @@ index f79e5d65c..c9333689f 100644
  static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
      /* .get_name                = */ ggml_backend_cuda_device_get_name,
      /* .get_description         = */ ggml_backend_cuda_device_get_description,
-@@ -3952,6 +3965,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
+@@ -4851,6 +4864,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
      /* .event_new               = */ ggml_backend_cuda_device_event_new,
      /* .event_free              = */ ggml_backend_cuda_device_event_free,
      /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
@@ -110,7 +110,7 @@ index f79e5d65c..c9333689f 100644
  
  // backend reg
 diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
-index 890c10364..1f06be80e 100644
+index b7d6edf7f..b987d7aeb 100644
 --- a/ggml/src/ggml-cuda/vendors/hip.h
 +++ b/ggml/src/ggml-cuda/vendors/hip.h
 @@ -45,6 +45,7 @@
diff --git a/llama/patches/0024-GPU-discovery-enhancements.patch b/llama/patches/0024-GPU-discovery-enhancements.patch
index 5a2adf8d..c372f0bc 100644
--- a/llama/patches/0024-GPU-discovery-enhancements.patch
+++ b/llama/patches/0024-GPU-discovery-enhancements.patch
@@ -20,10 +20,10 @@ fix vulkan PCI ID and ID handling
  ggml/src/ggml-cuda/vendors/hip.h     |   3 +
  ggml/src/ggml-impl.h                 |   8 +
  ggml/src/ggml-metal/ggml-metal.cpp   |   2 +
- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 209 +++++++++--
+ ggml/src/ggml-vulkan/ggml-vulkan.cpp | 169 ++++++++-
  ggml/src/mem_hip.cpp                 | 529 +++++++++++++++++++++++++++
  ggml/src/mem_nvml.cpp                | 209 +++++++++++
- 9 files changed, 1003 insertions(+), 30 deletions(-)
+ 9 files changed, 976 insertions(+), 17 deletions(-)
  create mode 100644 ggml/src/mem_hip.cpp
  create mode 100644 ggml/src/mem_nvml.cpp
 
@@ -45,7 +45,7 @@ index 69223c488..6510e0cba 100644
  
      GGML_API const char *                  ggml_backend_dev_name(ggml_backend_dev_t device);
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index f9a6587f1..03f359ae9 100644
+index 6d493a4ff..ac8f38464 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
 @@ -209,6 +209,8 @@ add_library(ggml-base
@@ -56,12 +56,12 @@ index f9a6587f1..03f359ae9 100644
 +            mem_nvml.cpp
              gguf.cpp)
  
- target_include_directories(ggml-base PRIVATE .)
+ set_target_properties(ggml-base PROPERTIES
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index c9333689f..f1a20e7fe 100644
+index 0787e443c..736d47c07 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -261,6 +261,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
+@@ -263,6 +263,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
      for (int id = 0; id < info.device_count; ++id) {
          int device_vmm = 0;
  
@@ -78,7 +78,7 @@ index c9333689f..f1a20e7fe 100644
  #if defined(GGML_USE_VMM)
          CUdevice device;
          CU_CHECK(cuDeviceGet(&device, id));
-@@ -314,6 +324,11 @@ static ggml_cuda_device_info ggml_cuda_init() {
+@@ -316,6 +326,11 @@ static ggml_cuda_device_info ggml_cuda_init() {
  #else
          info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
          info.devices[id].cc = 100*prop.major + 10*prop.minor;
@@ -90,7 +90,7 @@ index c9333689f..f1a20e7fe 100644
          GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, ID: %s\n",
                          id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
                          ggml_cuda_parse_uuid(prop, id).c_str());
-@@ -3468,6 +3483,11 @@ struct ggml_backend_cuda_device_context {
+@@ -4249,6 +4264,11 @@ struct ggml_backend_cuda_device_context {
      std::string description;
      std::string pci_bus_id;
      std::string id;
@@ -102,7 +102,7 @@ index c9333689f..f1a20e7fe 100644
  };
  
  static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
-@@ -3488,6 +3508,28 @@ static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
+@@ -4345,6 +4365,28 @@ static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
  static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
      ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
      ggml_cuda_set_device(ctx->device);
@@ -129,9 +129,9 @@ index c9333689f..f1a20e7fe 100644
 +    }
 +#endif
      CUDA_CHECK(cudaMemGetInfo(free, total));
- }
  
-@@ -3496,6 +3538,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
+ // ref: https://github.com/ggml-org/llama.cpp/pull/17368
+@@ -4377,6 +4419,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
      return GGML_BACKEND_DEVICE_TYPE_GPU;
  }
  
@@ -139,7 +139,7 @@ index c9333689f..f1a20e7fe 100644
  static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
      ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
  
-@@ -3509,6 +3552,19 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+@@ -4390,6 +4433,19 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
      // If you need the memory data, call ggml_backend_dev_memory() explicitly.
      props->memory_total = props->memory_free = 0;
  
@@ -159,7 +159,7 @@ index c9333689f..f1a20e7fe 100644
      bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
  #ifdef GGML_CUDA_NO_PEER_COPY
      bool events = false;
-@@ -4075,6 +4131,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -4974,6 +5030,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
          std::lock_guard<std::mutex> lock(mutex);
          if (!initialized) {
              ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
@@ -167,7 +167,7 @@ index c9333689f..f1a20e7fe 100644
  
              for (int i = 0; i < ggml_cuda_info().device_count; i++) {
                  ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
-@@ -4090,6 +4147,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -4989,6 +5046,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
                  dev_ctx->pci_bus_id = pci_bus_id;
  
@@ -183,7 +183,7 @@ index c9333689f..f1a20e7fe 100644
                      /* .iface   = */ ggml_backend_cuda_device_interface,
                      /* .reg     = */ &reg,
 diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
-index 1f06be80e..2f9ef2dc0 100644
+index b987d7aeb..5ad5623ae 100644
 --- a/ggml/src/ggml-cuda/vendors/hip.h
 +++ b/ggml/src/ggml-cuda/vendors/hip.h
 @@ -5,6 +5,8 @@
@@ -204,7 +204,7 @@ index 1f06be80e..2f9ef2dc0 100644
  #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
  #define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
 diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
-index e9201cdc6..44ae76d66 100644
+index fe57d4c58..1c07e767a 100644
 --- a/ggml/src/ggml-impl.h
 +++ b/ggml/src/ggml-impl.h
 @@ -677,6 +677,14 @@ static inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
@@ -223,10 +223,10 @@ index e9201cdc6..44ae76d66 100644
  }
  #endif
 diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
-index 05ff6a5a6..032dee76d 100644
+index ba95b4acc..f6f8f7a10 100644
 --- a/ggml/src/ggml-metal/ggml-metal.cpp
 +++ b/ggml/src/ggml-metal/ggml-metal.cpp
-@@ -537,6 +537,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
+@@ -546,6 +546,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
      GGML_UNUSED(dev);
  }
  
@@ -234,7 +234,7 @@ index 05ff6a5a6..032dee76d 100644
  static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
      props->name        = ggml_backend_metal_device_get_name(dev);
      props->description = ggml_backend_metal_device_get_description(dev);
-@@ -545,6 +546,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
+@@ -554,6 +555,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
  
      ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
  
@@ -243,18 +243,18 @@ index 05ff6a5a6..032dee76d 100644
          /* .async                 = */ true,
          /* .host_buffer           = */ false,
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 3a6bbe564..ca02ea079 100644
+index a36c6560c..a234eda2e 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -229,6 +229,7 @@ class vk_memory_logger;
- #endif
+@@ -236,6 +236,7 @@ class vk_memory_logger;
  class vk_perf_logger;
  static void ggml_vk_destroy_buffer(vk_buffer& buf);
+ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx);
 +static std::string ggml_vk_get_device_id(int device);
  
  static constexpr uint32_t mul_mat_vec_max_cols = 8;
  static constexpr uint32_t p021_max_gqa_ratio = 8;
-@@ -11813,6 +11814,29 @@ static void ggml_vk_get_device_description(int device, char * description, size_
+@@ -12353,6 +12354,29 @@ static void ggml_vk_get_device_description(int device, char * description, size_
      snprintf(description, description_size, "%s", props.deviceName.data());
  }
  
@@ -284,7 +284,7 @@ index 3a6bbe564..ca02ea079 100644
  // backend interface
  
  #define UNUSED GGML_UNUSED
-@@ -12761,31 +12785,102 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
+@@ -13614,15 +13638,72 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
      ggml_vk_get_device_description(dev_idx, description, description_size);
  }
  
@@ -312,24 +312,23 @@ index 3a6bbe564..ca02ea079 100644
 +    int driver_major;
 +    int driver_minor;
 +};
-+
+ 
+-    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
 +void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size_t * free, size_t * total) {
 +    GGML_ASSERT(ctx->device < (int) vk_instance.device_indices.size());
 +    GGML_ASSERT(ctx->device < (int) vk_instance.device_supports_membudget.size());
 +
 +    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[ctx->device]];
- 
--    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
--    vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops;
--    vk::PhysicalDeviceMemoryProperties2 memprops = {};
--    bool membudget_supported = vk_instance.device_supports_membudget[device];
-+    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+     vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops;
+     vk::PhysicalDeviceMemoryProperties2 memprops = {};
+-    const bool membudget_supported = vk_instance.device_supports_membudget[device];
++    const bool membudget_supported = vk_instance.device_supports_membudget[ctx->device];
+     const bool is_integrated_gpu = vkdev.getProperties().deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
++    
 +    vk::PhysicalDeviceProperties2 props2;
 +    vkdev.getProperties2(&props2);
- 
--    if (membudget_supported) {
--        memprops.pNext = &budgetprops;
-+    if (!ctx->is_integrated_gpu)
++
++    if (!is_integrated_gpu)
 +    {
 +        // Use vendor specific management libraries for best VRAM reporting if available
 +        switch (props2.properties.vendorID) {
@@ -356,55 +355,13 @@ index 3a6bbe564..ca02ea079 100644
 +            }
 +            break;
 +        }
-     }
--    vkdev.getMemoryProperties2(&memprops);
++    }
 +    // else fallback to memory budget if supported
++
  
--    for (uint32_t i = 0; i < memprops.memoryProperties.memoryHeapCount; ++i) {
--        const vk::MemoryHeap & heap = memprops.memoryProperties.memoryHeaps[i];
-+    *total = 0;
-+    *free = 0;
-+    vk::PhysicalDeviceMemoryBudgetPropertiesEXT mem_budget_props;
-+    vk::PhysicalDeviceMemoryProperties2 memprops2;
-+    memprops2.pNext = &mem_budget_props;
-+    vkdev.getMemoryProperties2(&memprops2);
-+    for (int i = 0; i < memprops2.memoryProperties.memoryHeapCount; i++) {
-+        if (memprops2.memoryProperties.memoryHeaps[i].flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-+            *total += memprops2.memoryProperties.memoryHeaps[i].size;
-+        } else if (ctx->is_integrated_gpu) {
-+            // Include shared memory on iGPUs
-+            *total += memprops2.memoryProperties.memoryHeaps[i].size;
-+        }
-+    }
-+    for (int i = 0; i < memprops2.memoryProperties.memoryHeapCount; i++) {
-+        if (memprops2.memoryProperties.memoryHeaps[i].flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-+            *free += mem_budget_props.heapBudget[i];
-+        } else if (ctx->is_integrated_gpu) {
-+            *free += mem_budget_props.heapBudget[i];
-+        }
-+    }
-+    if (*total > 0 && *free > 0) {
-+        return;
-+    } else if (*total > 0) {
-+        *free = *total;
-+        return;
-+    }
- 
-+    // else just report the physical memory
-+    for (const vk::MemoryHeap& heap : memprops2.memoryProperties.memoryHeaps) {
-         if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-             *total = heap.size;
--
--            if (membudget_supported && i < budgetprops.heapUsage.size()) {
--                *free = budgetprops.heapBudget[i] - budgetprops.heapUsage[i];
--            } else {
--                *free = heap.size;
--            }
-+            *free = heap.size;
-             break;
-         }
-     }
-@@ -12818,8 +12913,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
+     if (membudget_supported) {
+         memprops.pNext = &budgetprops;
+@@ -13674,8 +13755,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
          }
      }
  
@@ -419,7 +376,7 @@ index 3a6bbe564..ca02ea079 100644
      }
  
      vk::PhysicalDeviceProperties2 props = {};
-@@ -12836,19 +12936,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
+@@ -13692,19 +13778,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
  
      char pci_bus_id[16] = {};
      snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.%x", pci_domain, pci_bus, pci_device, pci_function);
@@ -453,7 +410,7 @@ index 3a6bbe564..ca02ea079 100644
  
  static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
      ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-@@ -12860,9 +12965,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
+@@ -13716,9 +13807,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
      return ctx->description.c_str();
  }
  
@@ -469,7 +426,7 @@ index 3a6bbe564..ca02ea079 100644
  }
  
  static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
-@@ -12886,8 +12996,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
+@@ -13742,8 +13838,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
  
      props->name        = ggml_backend_vk_device_get_name(dev);
      props->description = ggml_backend_vk_device_get_description(dev);
@@ -480,7 +437,7 @@ index 3a6bbe564..ca02ea079 100644
      ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
      props->caps = {
          /* .async                 = */ false,
-@@ -12895,6 +13006,13 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
+@@ -13751,6 +13848,13 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
          /* .buffer_from_host_ptr  = */ false,
          /* .events                = */ false,
      };
@@ -494,7 +451,7 @@ index 3a6bbe564..ca02ea079 100644
  }
  
  static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
-@@ -13365,6 +13483,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14319,6 +14423,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
          static std::mutex mutex;
          std::lock_guard<std::mutex> lock(mutex);
          if (!initialized) {
@@ -503,7 +460,7 @@ index 3a6bbe564..ca02ea079 100644
              for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
                  ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
                  char desc[256];
-@@ -13373,12 +13493,41 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14327,12 +14433,41 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                  ctx->name = GGML_VK_NAME + std::to_string(i);
                  ctx->description = desc;
                  ctx->is_integrated_gpu = ggml_backend_vk_get_device_type(i) == vk::PhysicalDeviceType::eIntegratedGpu;
diff --git a/llama/patches/0027-interleave-multi-rope.patch b/llama/patches/0027-interleave-multi-rope.patch
index 07873fed..1fee6b75 100644
--- a/llama/patches/0027-interleave-multi-rope.patch
+++ b/llama/patches/0027-interleave-multi-rope.patch
@@ -6,108 +6,101 @@ Subject: [PATCH] interleave multi rope
 since ollama doesn't use mrope for anything else, change it to mean the
 interleaved version used for qwen3vl
 ---
- ggml/src/ggml-cpu/ops.cpp                           |  7 ++-----
- ggml/src/ggml-cuda/rope.cu                          | 12 +++---------
- ggml/src/ggml-metal/ggml-metal.metal                | 10 +++-------
- ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp | 12 +++---------
- 4 files changed, 11 insertions(+), 30 deletions(-)
+ ggml/src/ggml-cpu/ops.cpp                           | 8 ++++----
+ ggml/src/ggml-cuda/rope.cu                          | 8 ++++----
+ ggml/src/ggml-metal/ggml-metal.metal                | 8 ++++----
+ ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl | 8 ++++----
+ 4 files changed, 16 insertions(+), 16 deletions(-)
 
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index 902fdad69..70955347d 100644
+index 40666bab6..3155cb4bb 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -5509,15 +5509,12 @@ static void ggml_mrope_cache_init(
-         }
+@@ -5599,14 +5599,14 @@ static void ggml_mrope_cache_init(
  
          float theta = theta_t;
--        if (sector >= sections[0] && sector < sec_w) {
-+        if (sector % 3 == 1 && sector < 1 + 3 * sections[1]) {
-             theta = theta_h;
-         }
--        else if (sector >= sec_w && sector < sec_w + sections[2]) {
-+        else if (sector % 3 == 2 && sector < 2 + 3 * sections[2]) {
-             theta = theta_w;
-         }
--        else if (sector >= sec_w + sections[2]) {
--            theta = theta_e;
--        }
- 
-         rope_yarn(
-             theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1]
+         if (is_imrope) { // qwen3vl apply interleaved mrope
+-            if (sector % 3 == 1 && sector < 3 * sections[1]) {
++            if (sector % 3 == 1 && sector < 1 + 3 * sections[1]) {
+                 theta = theta_h;
+-            } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
++            } else if (sector % 3 == 2 && sector < 2 + 3 * sections[2]) {
+                 theta = theta_w;
+             } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                 theta = theta_t;
+-            } else {
+-                theta = theta_e;
++            // } else {
++            //     theta = theta_e;
+             }
+         } else {
+             if (sector >= sections[0] && sector < sec_w) {
 diff --git a/ggml/src/ggml-cuda/rope.cu b/ggml/src/ggml-cuda/rope.cu
-index d058504cd..287fe9d2c 100644
+index 88ed79111..71ca60214 100644
 --- a/ggml/src/ggml-cuda/rope.cu
 +++ b/ggml/src/ggml-cuda/rope.cu
-@@ -151,19 +151,13 @@ static __global__ void rope_multi(
-     const int sec_w = sections.v[1] + sections.v[0];
-     const int sector = (i0 / 2) % sect_dims;
- 
--    float theta_base = 0.0;
--    if (sector < sections.v[0]) {
--        theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
--    }
--    else if (sector >= sections.v[0] && sector < sec_w) {
-+    float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
-+    if (sector % 3 == 1 && sector < 1 + 3 * sections.v[1]) {
-         theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
-     }
--    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-+    else if (sector % 3 == 2 && sector < 2 + 3 * sections.v[2]) {
-         theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
-     }
--    else if (sector >= sec_w + sections.v[2]) {
--        theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
--    }
- 
-     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+@@ -200,14 +200,14 @@ static __global__ void rope_multi(
  
+     float theta_base = 0.0;
+     if (is_imrope) {
+-        if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
++        if (sector % 3 == 1 && sector < 1 + 3 * sections.v[1]) { // h
+             theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
+-        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
++        } else if (sector % 3 == 2 && sector < 2 + 3 * sections.v[2]) { // w
+             theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+         } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
+             theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+-        } else {
+-            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
++        // } else {
++        //     theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+         }
+     } else {
+         if (sector < sections.v[0]) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
-index 50b8071de..65a3183c8 100644
+index aed013a9d..a489de435 100644
 --- a/ggml/src/ggml-metal/ggml-metal.metal
 +++ b/ggml/src/ggml-metal/ggml-metal.metal
-@@ -3888,15 +3888,11 @@ kernel void kernel_rope_multi(
-             const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
-             const int sector    = ic % sect_dims;
+@@ -4009,14 +4009,14 @@ kernel void kernel_rope_multi(
  
--            float theta_base;
--            if (sector < args.sect_0) {
--                theta_base = (float) pos[i2];
--            } else if (sector < sec_w01) {
-+            float theta_base = (float) pos[i2];
-+            if (sector % 3 == 1 && sector < 1 + 3 * args.sect_1) {
-                 theta_base = (float) pos[i2 + args.ne02];
--            } else if (sector < sec_w012) {
-+            } else if (sector % 3 == 2 && sector < 2 + 3 * args.sect_2) {
-                 theta_base = (float) pos[i2 + args.ne02 * 2];
--            } else {
--                theta_base = (float) pos[i2 + args.ne02 * 3];
-             }
-             // end of mrope
- 
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
-index 111286b49..633dc20ff 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
-@@ -31,19 +31,13 @@ void main() {
-     const int sec_w = p.sections[1] + p.sections[0];
-     const uint sector = (i0 / 2) % sect_dims;
- 
--    float theta_base = 0.0;
--    if (sector < p.sections[0]) {
--        theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
--    }
--    else if (sector >= p.sections[0] && sector < sec_w) {
-+    float theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
-+    if (sector % 3 == 1 && sector < 1 + 3 * p.sections[1]) {
-         theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
-     }
--    else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
-+    else if (sector % 3 == 2 && sector < 2 + 3 * p.sections[2]) {
-         theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
-     }
--    else if (sector >= sec_w + p.sections[2]) {
--        theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
--    }
- 
-     const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
+             float theta_base;
+             if (FC_rope_is_imrope) {
+-                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
++                if (sector % 3 == 1 && sector < 1 + 3 * args.sect_1) { // h
+                     theta_base = (float) pos[i2 + args.ne02 * 1];
+-                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
++                } else if (sector % 3 == 2 && sector < 2 + 3 * args.sect_2) { // w
+                     theta_base = (float) pos[i2 + args.ne02 * 2];
+                 } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                     theta_base = (float) pos[i2 + args.ne02 * 0];
+-                } else { // e
+-                    theta_base = (float) pos[i2 + args.ne02 * 3];
++                // } else { // e
++                //     theta_base = (float) pos[i2 + args.ne02 * 3];
+                 }
+             } else {
+                 if (sector < args.sect_0) {
+diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl
+index 9726b722d..1c8c69422 100644
+--- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl
++++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl
+@@ -148,14 +148,14 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
  
+     float theta_base = 0.0;
+     if (p.is_imrope != 0) {
+-        if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
++        if (sector % 3 == 1 && sector < 1 + 3 * p.sections[1]) {
+             theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+-        } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
++        } else if (sector % 3 == 2 && sector < 2 + 3 * p.sections[2]) {
+             theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+         } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
+             theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+-        } else {
+-            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
++        //} else {
++        //    theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+         }
+     } else {
+         if (sector < p.sections[0]) {
diff --git a/llama/patches/0028-Add-memory-detection-using-DXGI-PDH.patch b/llama/patches/0028-Add-memory-detection-using-DXGI-PDH.patch
index 1d7018b8..17656838 100644
--- a/llama/patches/0028-Add-memory-detection-using-DXGI-PDH.patch
+++ b/llama/patches/0028-Add-memory-detection-using-DXGI-PDH.patch
@@ -6,13 +6,13 @@ Subject: [PATCH] Add memory detection using DXGI + PDH
 ---
  ggml/src/CMakeLists.txt              |   1 +
  ggml/src/ggml-impl.h                 |   3 +
- ggml/src/ggml-vulkan/ggml-vulkan.cpp |  29 ++-
+ ggml/src/ggml-vulkan/ggml-vulkan.cpp |  26 ++-
  ggml/src/mem_dxgi_pdh.cpp            | 297 +++++++++++++++++++++++++++
- 4 files changed, 327 insertions(+), 3 deletions(-)
+ 4 files changed, 325 insertions(+), 2 deletions(-)
  create mode 100644 ggml/src/mem_dxgi_pdh.cpp
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index 03f359ae9..4b3e5efb5 100644
+index ac8f38464..faa1beed2 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
 @@ -211,6 +211,7 @@ add_library(ggml-base
@@ -22,9 +22,9 @@ index 03f359ae9..4b3e5efb5 100644
 +            mem_dxgi_pdh.cpp
              gguf.cpp)
  
- target_include_directories(ggml-base PRIVATE .)
+ set_target_properties(ggml-base PROPERTIES
 diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
-index 44ae76d66..639d551a2 100644
+index 1c07e767a..0da3e065b 100644
 --- a/ggml/src/ggml-impl.h
 +++ b/ggml/src/ggml-impl.h
 @@ -684,6 +684,9 @@ GGML_API void ggml_nvml_release();
@@ -38,10 +38,10 @@ index 44ae76d66..639d551a2 100644
  #ifdef __cplusplus
  }
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index ca02ea079..c12b069e5 100644
+index a234eda2e..c98f98c73 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -73,6 +73,7 @@ DispatchLoaderDynamic & ggml_vk_default_dispatcher();
+@@ -74,6 +74,7 @@ DispatchLoaderDynamic & ggml_vk_default_dispatcher();
  #define VK_KHR_SHADER_BFLOAT16_EXTENSION_NAME                        "VK_KHR_shader_bfloat16"
  #define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR ((VkStructureType)1000141000)
  #define VK_COMPONENT_TYPE_BFLOAT16_KHR                               ((VkComponentTypeKHR)1000141000)
@@ -49,7 +49,7 @@ index ca02ea079..c12b069e5 100644
  
  typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR {
      VkStructureType                       sType;
-@@ -12802,6 +12803,7 @@ struct ggml_backend_vk_device_context {
+@@ -13655,6 +13656,7 @@ struct ggml_backend_vk_device_context {
      std::string pci_id;
      std::string id;
      std::string uuid;
@@ -57,8 +57,8 @@ index ca02ea079..c12b069e5 100644
      int major;
      int minor;
      int driver_major;
-@@ -12817,8 +12819,22 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
-     vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+@@ -13673,6 +13675,20 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
+     
      vk::PhysicalDeviceProperties2 props2;
      vkdev.getProperties2(&props2);
 +    GGML_LOG_DEBUG("ggml_backend_vk_get_device_memory called: uuid %s\n", ctx->uuid.c_str());
@@ -76,22 +76,17 @@ index ca02ea079..c12b069e5 100644
 +        ggml_dxgi_pdh_release();
 +    }
  
--    if (!ctx->is_integrated_gpu)
-+    if (!ctx->is_integrated_gpu) 
+     if (!is_integrated_gpu)
      {
-         // Use vendor specific management libraries for best VRAM reporting if available
-         switch (props2.properties.vendorID) {
-@@ -12846,8 +12862,8 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
-             break;
-         }
+@@ -13704,7 +13720,6 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
      }
--    // else fallback to memory budget if supported
+     // else fallback to memory budget if supported
  
-+    // else fallback to memory budget if supported
-     *total = 0;
-     *free = 0;
-     vk::PhysicalDeviceMemoryBudgetPropertiesEXT mem_budget_props;
-@@ -13500,7 +13516,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+-
+     if (membudget_supported) {
+         memprops.pNext = &budgetprops;
+     }
+@@ -14440,7 +14455,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                      /* .reg     = */ reg,
                      /* .context = */ ctx,
                  });
@@ -99,7 +94,7 @@ index ca02ea079..c12b069e5 100644
                  // Gather additional information about the device
                  int dev_idx = vk_instance.device_indices[i];
                  vk::PhysicalDeviceProperties props1;
-@@ -13523,6 +13538,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14463,6 +14477,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                      }
                  }
                  ctx->uuid = oss.str();
diff --git a/llama/patches/0036-ggml-cuda-skip-large-batches.patch b/llama/patches/0029-ggml-cuda-skip-large-batches.patch
similarity index 91%
rename from llama/patches/0036-ggml-cuda-skip-large-batches.patch
rename to llama/patches/0029-ggml-cuda-skip-large-batches.patch
index 1c9ee45f..d1d1addd 100644
--- a/llama/patches/0036-ggml-cuda-skip-large-batches.patch
+++ b/llama/patches/0029-ggml-cuda-skip-large-batches.patch
@@ -10,10 +10,10 @@ fallback to cpu
  1 file changed, 3 insertions(+)
 
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index f1a20e7fe..1a71e07c9 100644
+index 736d47c07..7350f6758 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -3677,6 +3677,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
+@@ -4564,6 +4564,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                  if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
                      return false;
                  }
diff --git a/llama/patches/0029-vulkan-Call-ggml_vk_buffer_write_2d-from-ggml_vk_buf.patch b/llama/patches/0029-vulkan-Call-ggml_vk_buffer_write_2d-from-ggml_vk_buf.patch
deleted file mode 100644
index a99b6120..00000000
--- a/llama/patches/0029-vulkan-Call-ggml_vk_buffer_write_2d-from-ggml_vk_buf.patch
+++ /dev/null
@@ -1,32 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Jeff Bolz <jbolz@nvidia.com>
-Date: Wed, 29 Oct 2025 03:53:04 -0500
-Subject: [PATCH] vulkan: Call ggml_vk_buffer_write_2d from ggml_vk_buffer_copy
- (#16793)
-
-This lets the copy to the destination device use the host-visible
-vidmem optimization.
----
- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 5 +----
- 1 file changed, 1 insertion(+), 4 deletions(-)
-
-diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index c12b069e5..76c78c2ea 100644
---- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -5654,14 +5654,11 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
-         VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")");
-         // Copy device to device
-         ggml_vk_ensure_sync_staging_buffer(src->device, size);
--        ggml_vk_ensure_sync_staging_buffer(dst->device, size);
- 
-         // Copy to src staging buffer
-         ggml_vk_buffer_copy(src->device->sync_staging, 0, src, src_offset, size);
--        // memcpy to dst staging buffer
--        memcpy(dst->device->sync_staging->ptr, src->device->sync_staging->ptr, size);
-         // Copy to dst buffer
--        ggml_vk_buffer_copy(dst, dst_offset, dst->device->sync_staging, 0, size);
-+        ggml_vk_buffer_write_2d(dst, dst_offset, src->device->sync_staging->ptr, 0, size, 1);
-     }
- }
- 
diff --git a/llama/patches/0030-Vulkan-MMQ-Integer-Dot-Refactor-and-K-Quant-support-.patch b/llama/patches/0030-Vulkan-MMQ-Integer-Dot-Refactor-and-K-Quant-support-.patch
deleted file mode 100644
index 24286766..00000000
--- a/llama/patches/0030-Vulkan-MMQ-Integer-Dot-Refactor-and-K-Quant-support-.patch
+++ /dev/null
@@ -1,2140 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Ruben Ortlam <picard12@live.de>
-Date: Wed, 29 Oct 2025 14:39:03 +0100
-Subject: [PATCH] Vulkan MMQ Integer Dot Refactor and K-Quant support (#16536)
-
-* vulkan: add mmq q2_k integer dot support
-
-* Refactor mmq caching
-
-* Reduce mmq register use
-
-* Load 4 quant blocks into shared memory in one step
-
-* Pack q2_k blocks into caches of 32
-
-* Use 32-bit accumulators for integer dot matmul
-
-* Add q4_k mmq
-
-* Add q3_k mmq
-
-* Add q5_k mmq
-
-* Add q6_k mmq
-
-* Add mxfp4 mmq, enable MMQ MUL_MAT_ID
-
-* Fix mmv dm loads
----
- ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 165 +++++-
- .../vulkan-shaders/dequant_funcs.glsl         |  10 +-
- .../vulkan-shaders/dequant_funcs_cm2.glsl     |   6 +-
- .../vulkan-shaders/dequant_mxfp4.comp         |   4 +-
- .../vulkan-shaders/dequant_q2_k.comp          |   4 +-
- .../vulkan-shaders/dequant_q4_k.comp          |   4 +-
- .../vulkan-shaders/dequant_q5_k.comp          |   4 +-
- .../vulkan-shaders/mul_mat_vec_q2_k.comp      |   6 +-
- .../vulkan-shaders/mul_mat_vec_q4_k.comp      |   6 +-
- .../vulkan-shaders/mul_mat_vec_q5_k.comp      |   6 +-
- .../ggml-vulkan/vulkan-shaders/mul_mm.comp    |  72 +--
- .../vulkan-shaders/mul_mm_funcs.glsl          |  14 +-
- .../vulkan-shaders/mul_mm_id_funcs.glsl       |  70 +++
- .../ggml-vulkan/vulkan-shaders/mul_mmq.comp   | 288 +++-------
- .../vulkan-shaders/mul_mmq_funcs.glsl         | 538 ++++++++++++++++--
- .../vulkan-shaders/mul_mmq_shmem_types.glsl   |  78 +++
- .../src/ggml-vulkan/vulkan-shaders/types.glsl |  53 +-
- .../vulkan-shaders/vulkan-shaders-gen.cpp     |   5 +-
- 18 files changed, 928 insertions(+), 405 deletions(-)
- create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl
- create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl
-
-diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 76c78c2ea..7669ed206 100644
---- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -488,6 +488,7 @@ struct vk_device_struct {
-     vk_matmul_pipeline2 pipeline_matmul_id_f16_f32;
- 
-     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
-+    vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_COUNT];
- 
-     vk_pipeline pipeline_matmul_split_k_reduce;
-     vk_pipeline pipeline_quantize_q8_1;
-@@ -2449,8 +2450,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
-                           l_warptile_id, m_warptile_id, s_warptile_id,
-                           l_warptile_mmq, m_warptile_mmq, s_warptile_mmq,
-                           l_warptile_mmq_int, m_warptile_mmq_int, s_warptile_mmq_int,
-+                          l_warptile_mmq_int_k, m_warptile_mmq_int_k, s_warptile_mmq_int_k,
-                           l_warptile_mmq_k, m_warptile_mmq_k, s_warptile_mmq_k,
--                          l_warptile_mmqid, m_warptile_mmqid, s_warptile_mmqid;
-+                          l_warptile_mmqid, m_warptile_mmqid, s_warptile_mmqid,
-+                          l_warptile_mmqid_int, m_warptile_mmqid_int, s_warptile_mmqid_int,
-+                          l_warptile_mmqid_int_k, m_warptile_mmqid_int_k, s_warptile_mmqid_int_k;
-     std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms,
-                             l_mmq_wg_denoms, m_mmq_wg_denoms, s_mmq_wg_denoms,
-                             l_mmq_wg_denoms_k, m_mmq_wg_denoms_k, s_mmq_wg_denoms_k,
-@@ -2513,10 +2517,16 @@ static void ggml_vk_load_shaders(vk_device& device) {
-         m_warptile_mmq = { 128,  64,  64, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
-         s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };
- 
-+        // Integer MMQ has a smaller shared memory profile, but heavier register use
-         l_warptile_mmq_int = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
-         m_warptile_mmq_int = { 128,  64,  64, 32, subgroup_size_8,     32, 2, 2, 2, 1, subgroup_size_8 };
-         s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32,       32, 2, 2, 1, 1, subgroup_size_8 };
- 
-+        // K-quants use even more registers, mitigate by setting WMITER to 1
-+        l_warptile_mmq_int_k = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 1, 4, 4, 1, subgroup_size_8 };
-+        m_warptile_mmq_int_k = { 128,  64,  64, 32, subgroup_size_8,     32, 1, 2, 2, 1, subgroup_size_8 };
-+        s_warptile_mmq_int_k = { subgroup_size_32, 32, 32, 32, 32,       32, 1, 2, 1, 1, subgroup_size_8 };
-+
-         l_warptile_id = { 128, 128, 128, 16, mul_mat_subgroup_size_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_16 };
-         m_warptile_id = { 128,  64,  64, 16, mul_mat_subgroup_size_16, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_16 };
-         s_warptile_id = { mul_mat_subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 };
-@@ -2525,10 +2535,18 @@ static void ggml_vk_load_shaders(vk_device& device) {
-         m_warptile_mmqid = { 128,  64,  64, 32, mul_mat_subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_8 };
-         s_warptile_mmqid = { mul_mat_subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 };
- 
-+        l_warptile_mmqid_int = { 128, 128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, 4, 4, 1, mul_mat_subgroup_size_8 };
-+        m_warptile_mmqid_int = { 128,  64,  64, 32, mul_mat_subgroup_size_8,     32, 2, 2, 2, 1, mul_mat_subgroup_size_8 };
-+        s_warptile_mmqid_int = { mul_mat_subgroup_size_32, 32, 32, 32, 32,       32, 2, 2, 1, 1, mul_mat_subgroup_size_8 };
-+
-+        l_warptile_mmqid_int_k = { 128, 128, 128, 32, mul_mat_subgroup_size_16 * 2, 64, 1, 4, 4, 1, mul_mat_subgroup_size_16 };
-+        m_warptile_mmqid_int_k = { 128,  64,  64, 32, mul_mat_subgroup_size_16,     32, 1, 2, 2, 1, mul_mat_subgroup_size_16 };
-+        s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32, 32, 32, 32,       32, 1, 2, 1, 1, mul_mat_subgroup_size_16 };
-+
-         // chip specific tuning
-         if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
-             m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
--            m_warptile_mmqid = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
-+            m_warptile_mmqid = m_warptile_mmqid_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
-         }
- 
-         l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
-@@ -2913,18 +2931,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
-         if (device->mul_mat ## ID ## _s[TYPE]) \
-             ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
- 
--#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
-+#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
-         if (device->mul_mat ## ID ## _l[TYPE]) { \
--            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->l, #NAMELC "_f16acc_l", NAMELC ## _f16acc_len, NAMELC ##  _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
--            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->l, #NAMELC        "_l", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
-+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->l, #NAMELC        "_l", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
-         } \
-         if (device->mul_mat ## ID ## _m[TYPE]) { \
--            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->m, #NAMELC "_f16acc_m", NAMELC ## _f16acc_len, NAMELC ##  _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
--            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->m, #NAMELC        "_m", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
-+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->m, #NAMELC        "_m", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
-         } \
-         if (device->mul_mat ## ID ## _s[TYPE]) { \
--            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f16acc->s, #NAMELC "_f16acc_s", NAMELC ## _f16acc_len, NAMELC ##  _f16acc_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
--            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->s, #NAMELC        "_s", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
-+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME .f32acc->s, #NAMELC        "_s", NAMELC ## _len,        NAMELC ##  _data,        "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
-         } \
- 
-         // Create 2 variants, {f16,f32} accumulator
-@@ -2963,11 +2978,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
- 
- #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-         if (device->integer_dot_product) {
--            CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0], matmul_q4_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
--            CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1], matmul_q4_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
--            CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0], matmul_q5_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
--            CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1], matmul_q5_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
--            CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0], matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-+            CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0], matmul_q4_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1], matmul_q4_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0], matmul_q5_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1], matmul_q5_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0], matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, , 0);
-+
-+            CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_MXFP4], matmul_mxfp4_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, , 0);
-+
-+            CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q2_K], matmul_q2_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q3_K], matmul_q3_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_K], matmul_q4_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_K], matmul_q5_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, , 0);
-+            CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q6_K], matmul_q6_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, , 0);
-         }
- #endif
- 
-@@ -2997,6 +3020,24 @@ static void ggml_vk_load_shaders(vk_device& device) {
-             CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_subgroup_iq4_xs_f32,  mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-             CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_subgroup_iq4_nl_f32,  mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-             CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_subgroup_mxfp4_f32,   mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-+
-+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-+            if (device->integer_dot_product) {
-+                CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-+                CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-+                CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-+                CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-+                CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-+
-+                CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
-+
-+                CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
-+                CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
-+                CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
-+                CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
-+                CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
-+            }
-+#endif
-         } else {
-             CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
-             CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
-@@ -3023,6 +3064,24 @@ static void ggml_vk_load_shaders(vk_device& device) {
-             CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f32,  mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
-             CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f32,  mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
-             CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f32,   mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
-+
-+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-+            if (device->integer_dot_product) {
-+                CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, 0);
-+
-+                CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int,   vk_mat_mat_id_push_constants, 4, _id, 0);
-+
-+                CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
-+                CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
-+            }
-+#endif
-         }
- #undef CREATE_MM2
- #undef CREATE_MMQ
-@@ -3087,6 +3146,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
-             CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-             CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-             CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, );
-+
-+            CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, );
-+            CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, );
-+            CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, );
-+            CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, );
-+            CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_q8_1, mmq_wg_denoms, warptile_mmq_int_k, vk_mat_mat_push_constants, 3, );
-         }
- #endif
- 
-@@ -3146,7 +3211,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
-     }
-     // reusing CREATE_MM from the fp32 path
-     if ((device->coopmat2 || device->coopmat_support)
--#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
-         && !device->coopmat_bf16_support
- #endif
-         ) {
-@@ -4930,7 +4995,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
- 
-     // MMQ
-     if (src1_type == GGML_TYPE_Q8_1) {
--        vk_matmul_pipeline pipelines = (ctx->device->fp16 && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f32acc;
-+        vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f32acc;
- 
-         if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
-             return nullptr;
-@@ -5077,6 +5142,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
-         }
-     }
- 
-+    // MMQ
-+    if (src1_type == GGML_TYPE_Q8_1) {
-+        vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_id_q8_1[src0_type].f32acc;
-+
-+        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
-+            return nullptr;
-+        }
-+
-+        return pipelines;
-+    }
-+
-     GGML_ASSERT(src1_type == GGML_TYPE_F32 || (ctx->device->coopmat2 && src1_type == GGML_TYPE_F16));
- 
-     switch (src0_type) {
-@@ -6879,10 +6955,19 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
- 
-     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
- 
--    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
-+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;
-+
-+    // Check for mmq first
-+    vk_matmul_pipeline mmp = quantize_y ? ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, GGML_TYPE_Q8_1, (ggml_prec)dst->op_params[0]) : nullptr;
-+
-+    if (mmp == nullptr) {
-+        // Fall back to f16 dequant mul mat
-+        mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
-+        quantize_y = false;
-+    }
- 
-     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
--    const bool qy_needs_dequant = (src1->type != f16_type && !y_f32_kernel) || y_non_contig;
-+    const bool qy_needs_dequant = !quantize_y && ((src1->type != f16_type && !y_f32_kernel) || y_non_contig);
- 
-     if (qx_needs_dequant) {
-         // Fall back to dequant + f16 mulmat
-@@ -6892,8 +6977,8 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-     // Not implemented
-     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
- 
--    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type));
--    const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
-+    const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type));
-+    const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && nei1 > 8;
- 
-     vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type);
- 
-@@ -6906,12 +6991,13 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
-     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
-     const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
--    const uint64_t y_sz = y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
-+    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
-     const uint64_t ids_sz = nbi2;
-     const uint64_t d_sz = sizeof(float) * d_ne;
- 
-     vk_pipeline to_fp16_vk_0 = nullptr;
-     vk_pipeline to_fp16_vk_1 = nullptr;
-+    vk_pipeline to_q8_1 = nullptr;
- 
-     if (x_non_contig) {
-         to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
-@@ -6926,9 +7012,16 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
-     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
- 
-+    if (quantize_y) {
-+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
-+    }
-+
-     if (dryrun) {
-         const uint64_t x_sz_upd = x_sz * ne02 * ne03;
--        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
-+        uint64_t y_sz_upd = y_sz * ne12 * ne13;
-+        if (quantize_y) {
-+            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
-+        }
-         if (
-                 (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                 (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange)) {
-@@ -6937,7 +7030,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-         if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-             ctx->prealloc_size_x = x_sz_upd;
-         }
--        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
-+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
-             ctx->prealloc_size_y = y_sz_upd;
-         }
- 
-@@ -6949,6 +7042,9 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-         if (qy_needs_dequant) {
-             ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
-         }
-+        if (quantize_y) {
-+            ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
-+        }
-         return;
-     }
- 
-@@ -6985,6 +7081,9 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-     if (qy_needs_dequant) {
-         d_Y = ctx->prealloc_y;
-         GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
-+    } else if (quantize_y) {
-+        d_Y = ctx->prealloc_y;
-+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
-     } else {
-         d_Y = d_Qy;
-         y_buf_offset = qy_buf_offset;
-@@ -7016,6 +7115,17 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-             ctx->prealloc_y_last_tensor_used = src1;
-         }
-     }
-+    if (quantize_y) {
-+        if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
-+            ctx->prealloc_y_last_tensor_used != src1) {
-+            if (ctx->prealloc_y_need_sync) {
-+                ggml_vk_sync_buffers(ctx, subctx);
-+            }
-+            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne * ne12 * ne13, true);
-+            ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
-+            ctx->prealloc_y_last_tensor_used = src1;
-+        }
-+    }
- 
-     uint32_t stride_batch_x = ne00*ne01;
-     uint32_t stride_batch_y = ne10*ne11;
-@@ -7024,14 +7134,19 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
-         stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
-     }
- 
--    if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
-+    if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant && !quantize_y) {
-         stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
-     }
- 
-+    uint32_t y_sz_total = y_sz * ne12 * ne13;
-+    if (quantize_y) {
-+        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
-+    }
-+
-     // compute
-     ggml_vk_matmul_id(
-         ctx, subctx, pipeline,
--        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 },
-+        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz_total },
-         { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz },
-         ne01, ne21, ne10, ne10, ne10, ne01,
-         stride_batch_x, stride_batch_y, ne20*ne21,
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
-index 0d98f5a9d..09676a623 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
-@@ -437,7 +437,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
- #if defined(DATA_A_MXFP4)
- vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
--    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]);
-+    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]) * 0.5;
- }
- vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-     vec2 v0 = dequantize(ib, iqs, a_offset);
-@@ -488,9 +488,9 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
- 
-     const uvec2 qs = uvec2(data_a[a_offset + ib].qs[qsi], data_a[a_offset + ib].qs[qsi + 1]);
-     const uint scales = data_a[a_offset + ib].scales[scalesi];
--    const vec2 d = vec2(data_a[a_offset + ib].d);
-+    const vec2 dm = vec2(data_a[a_offset + ib].dm);
- 
--    return d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
-+    return dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
- }
- vec2 get_dm(uint ib, uint a_offset) {
-     return vec2(1, 0);
-@@ -529,7 +529,7 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-     const uint is = 2 * n + b;                 // 0..7
-     const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
- 
--    const vec2 loadd = vec2(data_a[a_offset + ib].d);
-+    const vec2 loadd = vec2(data_a[a_offset + ib].dm);
- 
-     const uint scidx0 = (is < 4) ? is : (is + 4);
-     const uint scidx1 = (is < 4) ? is : (is - 4);
-@@ -567,7 +567,7 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
- 
-     const uint8_t hm = uint8_t(1 << (iqs / 16));
- 
--    const vec2 loadd = vec2(data_a[a_offset + ib].d);
-+    const vec2 loadd = vec2(data_a[a_offset + ib].dm);
- 
-     const uint scidx0 = (is < 4) ? is : (is + 4);
-     const uint scidx1 = (is < 4) ? is : (is - 4);
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl
-index 67baedf7c..8ac6482dc 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl
-@@ -120,7 +120,7 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2
- float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
- {
-     decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
--    const f16vec2 d = bl.block.d;
-+    const f16vec2 dm = bl.block.dm;
-     const uint idx = coordInBlock[1];
- 
-     const uint scalesi = (idx & 0xF0) >> 4;             // 0..15
-@@ -131,7 +131,7 @@ float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2
-     qs = unpack8(qs)[idx & 1];
- 
-     const uint scales = bl.block.scales[scalesi];
--    float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
-+    float16_t ret = dm.x * float16_t(scales & 0xF) * float16_t(qs) - dm.y * float16_t(scales >> 4);
-     return ret;
- }
- 
-@@ -680,7 +680,7 @@ float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords
-     uint32_t qs = bl.block.qs[iqs];
-     qs >>= shift;
-     qs &= 0xF;
--    float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
-+    float16_t ret = float16_t(kvalues_mxfp4[qs] * d * 0.5);
-     return ret;
- }
- #endif
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp
-index ffba5a77d..3194ba291 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp
-@@ -26,7 +26,7 @@ void main() {
-     const float d = e8m0_to_fp32(data_a[ib].e);
- 
-     [[unroll]] for (uint l = 0; l < 8; ++l) {
--        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]);
--        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]);
-+        data_b[b_idx + l +  0] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]));
-+        data_b[b_idx + l + 16] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]));
-     }
- }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp
-index 58dc2e5df..dc05a7834 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp
-@@ -24,8 +24,8 @@ void main() {
-         const uint ql_idx = 32 * ip + il;
-         const uint8_t qs = data_a[i].qs[32 * ip + il];
- 
--        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
--        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
-+        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].dm.x);
-+        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].dm.y);
-         data_b[y_idx +  0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
-         data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
-         data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp
-index 8b7be557e..0f23dc0a3 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp
-@@ -20,8 +20,8 @@ void main() {
-         const uint is = 2 * il;
-         const uint n = 4;
- 
--        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
--        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
-+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].dm.x);
-+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].dm.y);
- 
-         const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
-         const uint qs_idx = 32*il + n * ir;
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp
-index 6bc04670f..970469a60 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp
-@@ -19,8 +19,8 @@ void main() {
-         const uint ir = tid % 16;
-         const uint is = 2 * il;
- 
--        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
--        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
-+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].dm.x);
-+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].dm.y);
- 
-         const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir;
-         const uint qs_idx = 32*il + 2 * ir;
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
-index 03ed25d3b..14093c0de 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
-@@ -41,9 +41,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
-         const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
-         const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
- 
--        vec2 d = vec2(data_a[ib0 + i].d);
--        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
--        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-+        const FLOAT_TYPE_VEC2 dm = vec2(data_a[ib0 + i].dm);
- 
-         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-             vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
-@@ -75,7 +73,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
-                        fma(FLOAT_TYPE(b96[l]),  sccache2[csel][ix][6 + 8*v_im],
-                        fma(FLOAT_TYPE(b112[l]), sccache2[csel][ix][7 + 8*v_im], sum2))))))));
-             }
--            temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
-+            temp[j][n] = fma(dm.x, sum1, fma(-dm.y, sum2, temp[j][n]));
-         }
-     }
- }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
-index 21d07d2e5..49d91ad59 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
-@@ -14,9 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
- 
-     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-         const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
--        vec2 d = vec2(data_a[ib0 + i].d);
--        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
--        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-+        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
- 
-         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-@@ -81,7 +79,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
-                 fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
-                 fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
-                 fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
--            temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
-+            temp[j][n] = fma(dm.x, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dm.y, smin, temp[j][n]));
-         }
-     }
- }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
-index 9e46c89a1..0d61b4966 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
-@@ -14,9 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
- 
-     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-         const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
--        vec2 d = vec2(data_a[ib0 + i].d);
--        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
--        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-+        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
- 
-         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-@@ -113,7 +111,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
-               fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
-               fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
-                   (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
--            temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
-+            temp[j][n] = fma(dm.x, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dm.y, smin, temp[j][n]));
-         }
-     }
- }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
-index a20788c4b..d260969f0 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
-@@ -120,81 +120,11 @@ shared FLOAT_TYPE_VEC2 buf_b[BN * SHMEM_STRIDE];
- 
- #define NUM_WARPS (BLOCK_SIZE / WARP)
- 
--#ifdef MUL_MAT_ID
--shared u16vec2 row_ids[BN];
--uint _ne1;
--
--#ifdef MUL_MAT_ID_USE_SUBGROUPS
--shared uvec4 ballots_sh[NUM_WARPS];
--
--void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
--    _ne1 = 0;
--    uint num_elements = p.nei1 * p.nei0;
--    uint nei0shift = findLSB(p.nei0);
--
--    uint ids[16];
--    uint iter = 0;
--
--    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
--        // prefetch up to 16 elements
--        if (iter == 0) {
--            [[unroll]] for (uint k = 0; k < 16; ++k) {
--                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
--                bool in_range = i < num_elements;
--                uint ii1;
--                if (nei0_is_pow2) {
--                    ii1 = i >> nei0shift;
--                } else {
--                    ii1 = i / p.nei0;
--                }
--                uint ii0 = i - ii1 * p.nei0;
--                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
--            }
--        }
--        uint i = j + gl_LocalInvocationIndex;
--        bool in_range = i < num_elements;
--        uint ii1;
--        if (nei0_is_pow2) {
--            ii1 = i >> nei0shift;
--        } else {
--            ii1 = i / p.nei0;
--        }
--        uint ii0 = i - ii1 * p.nei0;
--        uint id = ids[iter++];
--        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
--
--        ballots_sh[gl_SubgroupID] = ballot;
--        barrier();
--
--        uint subgroup_base = 0;
--        uint total = 0;
--        for (uint k = 0; k < gl_NumSubgroups; ++k) {
--            if (k == gl_SubgroupID) {
--                subgroup_base = total;
--            }
--            total += subgroupBallotBitCount(ballots_sh[k]);
--        }
--        barrier();
--
--        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
--        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
--            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
--        }
--        _ne1 += total;
--        iter &= 15;
--        if (_ne1 >= (ic + 1) * BN) {
--            break;
--        }
--    }
--    barrier();
--}
--#endif // MUL_MAT_ID_USE_SUBGROUPS
--#endif // MUL_MAT_ID
--
- #ifdef COOPMAT
- shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
- #endif
- 
-+#include "mul_mm_id_funcs.glsl"
- #include "mul_mm_funcs.glsl"
- 
- void main() {
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl
-index 0ebfbd646..ee5ded2e8 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl
-@@ -134,15 +134,15 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
-             const uint ib = idx / 128;                         // 2 values per idx
-             const uint iqs = idx % 128;                        // 0..127
- 
--            const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
-+            const uint qsi = (iqs / 64) * 16 + (iqs % 16);     // 0..15
-             const uint scalesi = iqs / 8;                      // 0..15
-             const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
- 
--            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
-+            const uvec2 qs = uvec2(unpack8(data_a_packed16[ib].qs[qsi]));
-             const uint scales = data_a[ib].scales[scalesi];
--            const vec2 d = vec2(data_a[ib].d);
-+            const vec2 dm = vec2(data_a[ib].dm);
- 
--            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
-+            const vec2 v = dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
- 
-             buf_a[buf_idx] = FLOAT_TYPE_VEC2(v.xy);
- #elif defined(DATA_A_Q3_K)
-@@ -179,7 +179,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
-             const uint is = 2 * n + b;                 // 0..7
-             const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
- 
--            const vec2 loadd = vec2(data_a[ib].d);
-+            const vec2 loadd = vec2(data_a[ib].dm);
- 
-             const uint scidx0 = (is < 4) ? is : (is + 4);
-             const uint scidx1 = (is < 4) ? is : (is - 4);
-@@ -215,7 +215,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
- 
-             const uint8_t hm = uint8_t(1 << (iqs / 16));
- 
--            const vec2 loadd = vec2(data_a[ib].d);
-+            const vec2 loadd = vec2(data_a[ib].dm);
- 
-             const uint scidx0 = (is < 4) ? is : (is + 4);
-             const uint scidx1 = (is < 4) ? is : (is - 4);
-@@ -468,7 +468,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
-             const uint ib = idx / 8;
-             const uint iqs = (idx & 0x07) * 2;
- 
--            const float d = e8m0_to_fp32(data_a[ib].e);
-+            const float d = e8m0_to_fp32(data_a[ib].e) * 0.5;
-             const uint vui = uint(data_a[ib].qs[iqs]);
-             const uint vui2 = uint(data_a[ib].qs[iqs+1]);
- 
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl
-new file mode 100644
-index 000000000..1d0e84ac9
---- /dev/null
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl
-@@ -0,0 +1,70 @@
-+#ifdef MUL_MAT_ID
-+shared u16vec2 row_ids[BN];
-+uint _ne1;
-+
-+#ifdef MUL_MAT_ID_USE_SUBGROUPS
-+shared uvec4 ballots_sh[NUM_WARPS];
-+
-+void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
-+    _ne1 = 0;
-+    uint num_elements = p.nei1 * p.nei0;
-+    uint nei0shift = findLSB(p.nei0);
-+
-+    uint ids[16];
-+    uint iter = 0;
-+
-+    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
-+        // prefetch up to 16 elements
-+        if (iter == 0) {
-+            [[unroll]] for (uint k = 0; k < 16; ++k) {
-+                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
-+                bool in_range = i < num_elements;
-+                uint ii1;
-+                if (nei0_is_pow2) {
-+                    ii1 = i >> nei0shift;
-+                } else {
-+                    ii1 = i / p.nei0;
-+                }
-+                uint ii0 = i - ii1 * p.nei0;
-+                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-+            }
-+        }
-+        uint i = j + gl_LocalInvocationIndex;
-+        bool in_range = i < num_elements;
-+        uint ii1;
-+        if (nei0_is_pow2) {
-+            ii1 = i >> nei0shift;
-+        } else {
-+            ii1 = i / p.nei0;
-+        }
-+        uint ii0 = i - ii1 * p.nei0;
-+        uint id = ids[iter++];
-+        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-+
-+        ballots_sh[gl_SubgroupID] = ballot;
-+        barrier();
-+
-+        uint subgroup_base = 0;
-+        uint total = 0;
-+        for (uint k = 0; k < gl_NumSubgroups; ++k) {
-+            if (k == gl_SubgroupID) {
-+                subgroup_base = total;
-+            }
-+            total += subgroupBallotBitCount(ballots_sh[k]);
-+        }
-+        barrier();
-+
-+        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
-+        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
-+            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
-+        }
-+        _ne1 += total;
-+        iter &= 15;
-+        if (_ne1 >= (ic + 1) * BN) {
-+            break;
-+        }
-+    }
-+    barrier();
-+}
-+#endif // MUL_MAT_ID_USE_SUBGROUPS
-+#endif // MUL_MAT_ID
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
-index b5d761c0b..8b238ac4b 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
-@@ -10,10 +10,9 @@
- #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
- #endif
- 
--#ifdef COOPMAT
--#extension GL_KHR_cooperative_matrix : enable
--#extension GL_KHR_memory_scope_semantics : enable
-+#if defined(MUL_MAT_ID_USE_SUBGROUPS)
- #extension GL_KHR_shader_subgroup_basic : enable
-+#extension GL_KHR_shader_subgroup_ballot : enable
- #endif
- 
- #ifdef MUL_MAT_ID
-@@ -24,7 +23,10 @@
- 
- layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
- 
--layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
-+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-+#if defined(A_TYPE_PACKED16)
-+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
-+#endif
- #if defined(A_TYPE_PACKED32)
- layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
- #endif
-@@ -76,40 +78,27 @@ layout (constant_id = 10) const uint WARP = 32;
- 
- #define BK 32
- 
--#ifdef COOPMAT
--#define SHMEM_STRIDE (BK / 4 + 4)
--#else
--#define SHMEM_STRIDE (BK / 4 + 1)
--#endif
-+#define MMQ_SHMEM
- 
--shared int32_t buf_a_qs[BM * SHMEM_STRIDE];
-+#include "mul_mmq_shmem_types.glsl"
- 
--#ifndef COOPMAT
--#if QUANT_AUXF == 1
--shared FLOAT_TYPE buf_a_dm[BM];
--#else
--shared FLOAT_TYPE_VEC2 buf_a_dm[BM];
--#endif
-+#ifndef BK_STEP
-+#define BK_STEP 4
- #endif
- 
--shared int32_t buf_b_qs[BN * SHMEM_STRIDE];
--#ifndef COOPMAT
--shared FLOAT_TYPE_VEC2 buf_b_ds[BN];
--#endif
-+// Shared memory cache
-+shared block_a_cache buf_a[BM * BK_STEP];
-+shared block_b_cache buf_b[BN * BK_STEP];
-+// Register cache
-+block_a_cache cache_a[WMITER * TM];
-+block_b_cache cache_b;
- 
--#define LOAD_VEC_A (4 * QUANT_R)
-+#define LOAD_VEC_A (4 * QUANT_R_MMQ)
- #define LOAD_VEC_B 16
- 
--#ifdef MUL_MAT_ID
--shared u16vec2 row_ids[4096];
--#endif // MUL_MAT_ID
--
- #define NUM_WARPS (BLOCK_SIZE / WARP)
- 
--#ifdef COOPMAT
--shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
--#endif
--
-+#include "mul_mm_id_funcs.glsl"
- #include "mul_mmq_funcs.glsl"
- 
- void main() {
-@@ -139,26 +128,12 @@ void main() {
-     const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
-     const uint WSUBM = WM / WMITER;
-     const uint WSUBN = WN / WNITER;
--
--#ifdef COOPMAT
--    const uint warp_i = gl_SubgroupID;
--
--    const uint tiw = gl_SubgroupInvocationID;
--
--    const uint cms_per_row = WM / TM;
--    const uint cms_per_col = WN / TN;
--
--    const uint storestride = WARP / TM;
--    const uint store_r = tiw % TM;
--    const uint store_c = tiw / TM;
--#else
-     const uint warp_i = gl_LocalInvocationID.x / WARP;
- 
-     const uint tiw = gl_LocalInvocationID.x % WARP;
- 
-     const uint tiwr = tiw % (WSUBM / TM);
-     const uint tiwc = tiw / (WSUBM / TM);
--#endif
- 
-     const uint warp_r = warp_i % (BM / WM);
-     const uint warp_c = warp_i / (BM / WM);
-@@ -172,17 +147,27 @@ void main() {
-     const uint loadstride_b = BLOCK_SIZE * LOAD_VEC_B / BK;
- 
- #ifdef MUL_MAT_ID
--    uint _ne1 = 0;
--    for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
--        for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
-+#ifdef MUL_MAT_ID_USE_SUBGROUPS
-+    if (bitCount(p.nei0) == 1) {
-+        load_row_ids(expert_idx, true, ic);
-+    } else {
-+        load_row_ids(expert_idx, false, ic);
-+    }
-+#else
-+    _ne1 = 0;
-+    for (uint ii1 = 0; ii1 < p.nei1 && _ne1 < (ic + 1) * BN; ii1++) {
-+        for (uint ii0 = 0; ii0 < p.nei0 && _ne1 < (ic + 1) * BN; ii0++) {
-             if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
--                row_ids[_ne1] = u16vec2(ii0, ii1);
-+                if (_ne1 >= ic * BN) {
-+                    row_ids[_ne1 - ic * BN] = u16vec2(ii0, ii1);
-+                }
-                 _ne1++;
-             }
-         }
-     }
- 
-     barrier();
-+#endif
- 
-     // Workgroup has no work
-     if (ic * BN >= _ne1) return;
-@@ -209,159 +194,70 @@ void main() {
-     uint pos_b_ib = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / BK;
- #endif
- 
--#ifdef COOPMAT
--    coopmat<int8_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
--    coopmat<int8_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
--    coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_result;
--
--    coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> factors[cms_per_row * cms_per_col];
--
--    coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
--
--    [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
--        sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
--    }
--#else
--    int32_t cache_a_qs[WMITER * TM * BK / 4];
--
--    int32_t cache_b_qs[TN * BK / 4];
--
-     ACC_TYPE sums[WMITER * TM * WNITER * TN];
- 
-     [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
-         sums[i] = ACC_TYPE(0.0f);
-     }
--#endif
- 
--#if QUANT_AUXF == 1
--    FLOAT_TYPE cache_a_dm[WMITER * TM];
--#else
--    FLOAT_TYPE_VEC2 cache_a_dm[WMITER * TM];
--#endif
--
--    FLOAT_TYPE_VEC2 cache_b_ds[TN];
--
--    for (uint block = start_k; block < end_k; block += BK) {
-+    for (uint block = start_k; block < end_k; block += BK * BK_STEP) {
-         [[unroll]] for (uint l = 0; loadc_a + l < BM; l += loadstride_a) {
--            const uint ib = pos_a_ib + (loadc_a + l) * p.stride_a / BK;
--            const uint iqs = loadr_a;
-             const uint buf_ib = loadc_a + l;
-+            const uint ib = pos_a_ib + buf_ib * p.stride_a / BK;
-+            const uint iqs = loadr_a;
- 
--            if (iqs == 0) {
--#if QUANT_AUXF == 1
--                buf_a_dm[buf_ib] = get_d(ib);
--#else
--                buf_a_dm[buf_ib] = get_dm(ib);
--#endif
-+            [[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
-+                block_a_to_shmem(k_step * BM + buf_ib, ib + k_step, iqs);
-             }
--#if QUANT_R == 1
--            buf_a_qs[buf_ib * SHMEM_STRIDE + iqs] = repack(ib, iqs);
--#else
--            const i32vec2 vals = repack(ib, iqs);
--            buf_a_qs[buf_ib * SHMEM_STRIDE + iqs    ] = vals.x;
--            buf_a_qs[buf_ib * SHMEM_STRIDE + iqs + 4] = vals.y;
--#endif
-         }
-         [[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) {
-+            const uint buf_ib = loadc_b + l;
-+
- #ifdef MUL_MAT_ID
--            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
--            const uint idx = pos_b_ib + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
--            const uint ib = idx / 8;
--            const uint iqs = idx & 0x7;
-+            const u16vec2 row_idx = row_ids[buf_ib];
-+            const uint ib = pos_b_ib + row_idx.y * p.batch_stride_b / BK + (row_idx.x % p.ne11) * p.stride_b / BK;
- #else
--            const uint ib = pos_b_ib + (loadc_b + l) * p.stride_b / BK;
--            const uint ib_outer = ib / 4;
--            const uint ib_inner = ib % 4;
--
--            const uint iqs = loadr_b;
-+            const uint ib = pos_b_ib + buf_ib * p.stride_b / BK;
- #endif
-+            const uint iqs = loadr_b;
- 
--            const uint buf_ib = loadc_b + l;
--
--            if (iqs == 0) {
--                buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
-+            [[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
-+                block_b_to_shmem(k_step * BN + buf_ib, ib + k_step, iqs);
-             }
--            const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
--            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4    ] = values.x;
--            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 1] = values.y;
--            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 2] = values.z;
--            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 3] = values.w;
-         }
- 
-         barrier();
- 
--        pos_a_ib += 1;
--        pos_b_ib += 1;
-+        pos_a_ib += BK_STEP;
-+        pos_b_ib += BK_STEP;
- 
--#ifdef COOPMAT
--        [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
--            const uint ib_a = warp_r * WM + cm_row * TM;
-+        for (uint k_step = 0; k_step < BK_STEP; k_step++) {
-             // Load from shared into cache
--            coopMatLoad(cache_a, buf_a_qs, ib_a * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);
--
--            // TODO: only cache values that are actually needed
--            [[unroll]] for (uint t_idx = 0; t_idx < TM; t_idx++) {
--                cache_a_dm[t_idx] = buf_a_dm[ib_a + t_idx];
--            }
--
--            [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
--                const uint ib_b = warp_c * WN + cm_col * TN;
--                coopMatLoad(cache_b, buf_b_qs, ib_b * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);
--
--                // TODO: only cache values that are actually needed
--                [[unroll]] for (uint t_idx = 0; t_idx < TN; t_idx++) {
--                    cache_b_dm[t_idx] = buf_b_d[ib_b + t_idx];
--                }
--
--                cm_result = coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0);
--                cm_result = coopMatMulAdd(cache_a, cache_b, cm_result);
--
--                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
--                    coopmat_stage[warp_i * TM * TN + (store_c + col) * TM + store_r] = ACC_TYPE(float(cache_a_d[store_r]) * float(cache_b_d[store_c + col]));
--                }
--
--                coopMatLoad(factors, coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
--                sums[cm_col * cms_per_row + cm_row] += factors * coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(cm_result);
--            }
--        }
--#else
--        // Load from shared into cache
--        [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
--            [[unroll]] for (uint cr = 0; cr < TM; cr++) {
--                const uint ib = warp_r * WM + wsir * WSUBM + tiwr * TM + cr;
--                cache_a_dm[wsir * TM + cr] = buf_a_dm[ib];
--                [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
--                    cache_a_qs[(wsir * TM + cr) * (BK / 4) + idx_k] = buf_a_qs[ib * SHMEM_STRIDE + idx_k];
--                }
--            }
--        }
-+            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
-+                [[unroll]] for (uint cr = 0; cr < TM; cr++) {
-+                    const uint reg_ib = wsir * TM + cr;
-+                    const uint buf_ib = warp_r * WM + wsir * WSUBM + tiwr * TM + cr;
- 
--        [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
--            [[unroll]] for (uint cc = 0; cc < TN; cc++) {
--                const uint ib = warp_c * WN + wsic * WSUBN + tiwc * TN + cc;
--                cache_b_ds[cc] = buf_b_ds[ib];
--                [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
--                    cache_b_qs[cc * (BK / 4) + idx_k] = buf_b_qs[ib * SHMEM_STRIDE + idx_k];
-+                    block_a_to_registers(reg_ib, k_step * BM + buf_ib);
-                 }
-             }
- 
--            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
-+            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
-                 [[unroll]] for (uint cc = 0; cc < TN; cc++) {
--                    [[unroll]] for (uint cr = 0; cr < TM; cr++) {
--                        const uint cache_a_idx = wsir * TM + cr;
--                        const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
--                        int32_t q_sum = 0;
--                        [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
--                            q_sum += dotPacked4x8EXT(cache_a_qs[cache_a_idx * (BK / 4) + idx_k],
--                                                     cache_b_qs[cc * (BK / 4) + idx_k]);
--                        }
-+                    const uint ib = k_step * BN + warp_c * WN + wsic * WSUBN + tiwc * TN + cc;
-+                    block_b_to_registers(ib);
- 
--                        sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc], 1);
-+                    [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
-+                        [[unroll]] for (uint cr = 0; cr < TM; cr++) {
-+                            const uint cache_a_idx = wsir * TM + cr;
-+                            const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
-+
-+                            sums[sums_idx] += mmq_dot_product(cache_a_idx);
-+                        }
-                     }
-                 }
-             }
-         }
--#endif
- 
-         barrier();
-     }
-@@ -373,54 +269,6 @@ void main() {
-     const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
- #endif
- 
--#ifdef COOPMAT
--#ifdef MUL_MAT_ID
--    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
--        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
--            coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
--
--            [[unroll]] for (uint col = 0; col < BN; col += storestride) {
--                const uint row_i = dc + cm_col * TN + col + store_c;
--                if (row_i >= _ne1) break;
--
--                const u16vec2 row_idx = row_ids[row_i];
--
--                data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
--            }
--        }
--    }
--#else
--    const bool is_aligned = p.stride_d % 4 == 0;  // Assumption: D_TYPE == float
--
--    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
--        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
--            const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N;
--
--            if (is_aligned && is_in_bounds) {
--                // Full coopMat is within bounds and stride_d is aligned with 16B
--                coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_dtype = coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(sums[cm_col * cms_per_row + cm_row]);
--                coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor);
--            } else if (is_in_bounds) {
--                // Full coopMat is within bounds, but stride_d is not aligned
--                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
--
--                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
--                    data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
--                }
--            } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) {
--                // Partial coopMat is within bounds
--                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
--
--                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
--                    if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) {
--                        data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
--                    }
--                }
--            }
--        }
--    }
--#endif // MUL_MAT_ID
--#else
-     [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
-         [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
- 
-@@ -431,19 +279,21 @@ void main() {
-                 const uint row_i = dc_warp + cc;
-                 if (row_i >= _ne1) break;
- 
--                const u16vec2 row_idx = row_ids[row_i];
-+                const u16vec2 row_idx = row_ids[row_i - ic * BN];
- #endif // MUL_MAT_ID
-                 [[unroll]] for (uint cr = 0; cr < TM; cr++) {
-+                    const uint sums_idx = (wsic * TN + cc) * WMITER * TM + wsir * TM + cr;
- #ifdef MUL_MAT_ID
--                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
-+                    if (dr_warp + cr < p.M) {
-+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[sums_idx].x);
-+                    }
- #else
-                     if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
--                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
-+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[sums_idx].x);
-                     }
- #endif // MUL_MAT_ID
-                 }
-             }
-         }
-     }
--#endif // COOPMAT
- }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
-index fe71eb131..c0c03fedc 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
-@@ -6,41 +6,89 @@
- 
- // Each iqs value maps to a 32-bit integer
- 
--#if defined(DATA_A_Q4_0)
-+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q4_1)
-+// 2-byte loads for Q4_0 blocks (18 bytes)
-+// 4-byte loads for Q4_1 blocks (20 bytes)
- i32vec2 repack(uint ib, uint iqs) {
--    // Use 2-byte loads since a q4_0 block (18 bytes) is not divisible by 4
--    const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2    ],
--                                   data_a[ib].qs[iqs * 2 + 1]);
-+#ifdef DATA_A_Q4_0
-+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
-     const uint32_t vui = pack32(quants);
-     return i32vec2( vui       & 0x0F0F0F0F,
-                    (vui >> 4) & 0x0F0F0F0F);
-+#else // DATA_A_Q4_1
-+    const uint32_t vui = data_a_packed32[ib].qs[iqs];
-+    return i32vec2( vui       & 0x0F0F0F0F,
-+                   (vui >> 4) & 0x0F0F0F0F);
-+#endif
- }
- 
-+#ifdef DATA_A_Q4_0
- ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-     return ACC_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
- }
-+#else // DATA_A_Q4_1
-+ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-+    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-+}
- #endif
- 
--#if defined(DATA_A_Q4_1)
--i32vec2 repack(uint ib, uint iqs) {
--    // Use 4-byte loads since a q4_1 block (20 bytes) is divisible by 4
--    const uint32_t vui = data_a_packed32[ib].qs[iqs];
--    return i32vec2( vui       & 0x0F0F0F0F,
--                   (vui >> 4) & 0x0F0F0F0F);
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+#ifdef DATA_A_Q4_0
-+    buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
-+                                           data_a_packed16[ib].qs[iqs * 2 + 1]));
-+
-+    if (iqs == 0) {
-+        buf_a[buf_ib].dm = FLOAT_TYPE(data_a_packed16[ib].d);
-+    }
-+#else // DATA_A_Q4_1
-+    buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
-+
-+    if (iqs == 0) {
-+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
-+    }
-+#endif
- }
- 
--ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
--    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
- }
--#endif
- 
--#if defined(DATA_A_Q5_0)
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    int32_t q_sum = 0;
-+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
-+        const uint32_t vui = cache_a[ib_a].qs[iqs];
-+        const i32vec2 qs_a = i32vec2( vui       & 0x0F0F0F0F,
-+                                     (vui >> 4) & 0x0F0F0F0F);
-+
-+        const int32_t qs_b0 = cache_b.qs[iqs];
-+        const int32_t qs_b1 = cache_b.qs[iqs + 4];
-+
-+        q_sum += dotPacked4x8EXT(qs_a.x, qs_b0);
-+        q_sum += dotPacked4x8EXT(qs_a.y, qs_b1);
-+    }
-+
-+    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
-+}
-+#endif // MMQ_SHMEM
-+
-+#elif defined(DATA_A_Q5_0) || defined(DATA_A_Q5_1)
-+// 2-byte loads for Q5_0 blocks (22 bytes)
-+// 4-byte loads for Q5_1 blocks (24 bytes)
- i32vec2 repack(uint ib, uint iqs) {
--    // Use 2-byte loads since a q5_0 block (22 bytes) is not divisible by 4
--    const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2    ],
--                                   data_a[ib].qs[iqs * 2 + 1]);
-+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
-     const uint32_t vui = pack32(quants);
--    const int32_t qh = int32_t((uint32_t(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]) >> (4 * iqs));
-+#ifdef DATA_A_Q5_0
-+    const int32_t qh = int32_t((uint32_t(data_a_packed16[ib].qh[1]) << 16 | data_a_packed16[ib].qh[0]) >> (4 * iqs));
-+#else // DATA_A_Q5_1
-+    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
-+#endif
-     const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
-                      | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
- 
-@@ -50,40 +98,457 @@ i32vec2 repack(uint ib, uint iqs) {
-     return i32vec2(v0, v1);
- }
- 
-+#ifdef DATA_A_Q5_0
- ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-     return ACC_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
- }
-+#else // DATA_A_Q5_1
-+ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-+    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-+}
- #endif
- 
--#if defined(DATA_A_Q5_1)
--i32vec2 repack(uint ib, uint iqs) {
--    // Use 4-byte loads since a q5_1 block (24 bytes) is divisible by 4
--    const uint32_t vui = data_a_packed32[ib].qs[iqs];
--    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
--    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
--                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+#ifdef DATA_A_Q5_0
-+    buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
-+                                           data_a_packed16[ib].qs[iqs * 2 + 1]));
- 
--    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
--                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
-+    if (iqs == 0) {
-+        buf_a[buf_ib].dm = FLOAT_TYPE(data_a_packed16[ib].d);
-+        buf_a[buf_ib].qh = pack32(u16vec2(data_a_packed16[ib].qh[0], data_a_packed16[ib].qh[1]));
-+    }
-+#else // DATA_A_Q5_1
-+    buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
- 
--    return i32vec2(v0, v1);
-+    if (iqs == 0) {
-+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
-+        buf_a[buf_ib].qh = data_a_packed32[ib].qh;
-+    }
-+#endif
- }
- 
--ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
--    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
-+    cache_a[reg_ib].qh = buf_a[buf_ib].qh;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
- }
-+
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    int32_t q_sum = 0;
-+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
-+        const uint32_t vui = cache_a[ib_a].qs[iqs];
-+        const int32_t qh = int32_t(cache_a[ib_a].qh >> (4 * iqs));
-+        const int32_t qs_a0 = int32_t(vui & 0x0F0F0F0F)
-+                         | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
-+        const int32_t qs_a1 = int32_t((vui >> 4) & 0x0F0F0F0F)
-+                         | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
-+
-+        const int32_t qs_b0 = cache_b.qs[iqs];
-+        const int32_t qs_b1 = cache_b.qs[iqs + 4];
-+
-+        q_sum += dotPacked4x8EXT(qs_a0, qs_b0);
-+        q_sum += dotPacked4x8EXT(qs_a1, qs_b1);
-+    }
-+
-+    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
-+}
-+#endif // MMQ_SHMEM
- #endif
- 
- #if defined(DATA_A_Q8_0)
-+// 2-byte loads for Q8_0 blocks (34 bytes)
- int32_t repack(uint ib, uint iqs) {
--    // Use 2-byte loads since a q8_0 block (34 bytes) is not divisible by 4
--    return pack32(i16vec2(data_a[ib].qs[iqs * 2    ],
--                          data_a[ib].qs[iqs * 2 + 1]));
-+    return pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-+                          data_a_packed16[ib].qs[iqs * 2 + 1]));
- }
- 
- ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-     return ACC_TYPE(float(q_sum) * da * dsb.x);
- }
-+
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+    buf_a[buf_ib].qs[iqs] = pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2],
-+                                           data_a_packed16[ib].qs[iqs * 2 + 1]));
-+
-+    if (iqs == 0) {
-+        buf_a[buf_ib].dm = FLOAT_TYPE(data_a_packed16[ib].d);
-+    }
-+}
-+
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
-+}
-+
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    int32_t q_sum = 0;
-+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
-+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
-+        const int32_t qs_b = cache_b.qs[iqs];
-+
-+        q_sum += dotPacked4x8EXT(qs_a, qs_b);
-+    }
-+
-+    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
-+}
-+#endif // MMQ_SHMEM
-+#endif
-+
-+#if defined(DATA_A_MXFP4)
-+// 1-byte loads for mxfp4 blocks (17 bytes)
-+i32vec2 repack(uint ib, uint iqs) {
-+    const uint32_t quants = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
-+                                          data_a[ib].qs[iqs * 4 + 1],
-+                                          data_a[ib].qs[iqs * 4 + 2],
-+                                          data_a[ib].qs[iqs * 4 + 3]));
-+
-+    return i32vec2( quants       & 0x0F0F0F0F,
-+                   (quants >> 4) & 0x0F0F0F0F);
-+}
-+
-+ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-+    return ACC_TYPE(da * dsb.x * float(q_sum));
-+}
-+
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+    const uint32_t qs = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
-+                                      data_a[ib].qs[iqs * 4 + 1],
-+                                      data_a[ib].qs[iqs * 4 + 2],
-+                                      data_a[ib].qs[iqs * 4 + 3]));
-+
-+    const u8vec4 i_a0 = unpack8( qs       & 0x0F0F0F0F);
-+    const u8vec4 i_a1 = unpack8((qs >> 4) & 0x0F0F0F0F);
-+
-+    buf_a[buf_ib].qs[iqs    ] = pack32(i8vec4(kvalues_mxfp4[i_a0.x], kvalues_mxfp4[i_a0.y], kvalues_mxfp4[i_a0.z], kvalues_mxfp4[i_a0.w]));
-+    buf_a[buf_ib].qs[iqs + 4] = pack32(i8vec4(kvalues_mxfp4[i_a1.x], kvalues_mxfp4[i_a1.y], kvalues_mxfp4[i_a1.z], kvalues_mxfp4[i_a1.w]));
-+
-+    if (iqs == 0) {
-+        buf_a[buf_ib].d = FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e) * 0.5);
-+    }
-+}
-+
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].d = buf_a[buf_ib].d;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
-+}
-+
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    int32_t q_sum = 0;
-+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
-+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
-+
-+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
-+    }
-+
-+    return mul_q8_1(q_sum, cache_a[ib_a].d, cache_b.ds, 1);
-+}
-+#endif // MMQ_SHMEM
-+#endif
-+
-+// For k-quants, ib and iqs still assume 32-wide blocks, but k-quants are 256-wide
-+// iqs still refers to a 32-bit integer, meaning 0..7 for 32-wide quants
-+#if defined(DATA_A_Q2_K)
-+// 4-byte loads for Q2_K blocks (84 bytes)
-+int32_t repack(uint ib, uint iqs) {
-+    const uint ib_k = ib / 8;
-+    const uint iqs_k = (ib % 8) * 8 + iqs;
-+
-+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
-+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
-+
-+    return int32_t((data_a_packed32[ib_k].qs[qs_idx] >> qs_shift) & 0x03030303);
-+}
-+
-+uint8_t get_scale(uint ib, uint iqs) {
-+    const uint ib_k = ib / 8;
-+    const uint iqs_k = (ib % 8) * 8 + iqs;
-+
-+    return data_a[ib_k].scales[iqs_k / 4];
-+}
-+
-+ACC_TYPE mul_q8_1(const int32_t sum_d, const int32_t sum_m, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-+    return ACC_TYPE(dsb.x * (dma.x * float(sum_d) - dma.y * float(sum_m)));
-+}
-+
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+    const uint ib_k = ib / 8;
-+    const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
-+
-+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
-+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
-+
-+    // Repack 4x4 quants into one int
-+    const uint32_t vals0 = (data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x03030303;
-+    const uint32_t vals1 = (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x03030303;
-+    const uint32_t vals2 = (data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x03030303;
-+    const uint32_t vals3 = (data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x03030303;
-+
-+    buf_a[buf_ib].qs[iqs] = vals0 | (vals1 << 2) | (vals2 << 4) | (vals3 << 6);
-+
-+    if (iqs == 0) {
-+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
-+        buf_a[buf_ib].scales = unpack8(data_a_packed16[ib_k].scales[iqs_k / 8]);
-+    }
-+}
-+
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
-+    cache_a[reg_ib].scales = buf_a[buf_ib].scales;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 2; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
-+}
-+
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    int32_t sum_d = 0;
-+    int32_t sum_m = 0;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
-+        const uint8_t scale = cache_a[ib_a].scales[iqs / 4];
-+        const int32_t scale_m = int32_t(scale >> 4) * 0x01010101; // Duplicate 8-bit value across 32-bits.
-+        const int32_t qs_a = int32_t((cache_a[ib_a].qs[iqs / 4] >> ((iqs % 4) * 2)) & 0x03030303);
-+
-+        sum_d += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]) * (scale & 0xF);
-+        sum_m += dotPacked4x8EXT(scale_m, cache_b.qs[iqs]);
-+    }
-+
-+    return mul_q8_1(sum_d, sum_m, cache_a[ib_a].dm, cache_b.ds, 1);
-+}
-+#endif // MMQ_SHMEM
-+#endif
-+
-+#if defined(DATA_A_Q3_K)
-+// 2-byte loads for Q3_K blocks (110 bytes)
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+    const uint ib_k = ib / 8;
-+    const uint hm_idx = iqs * QUANT_R_MMQ;
-+    const uint iqs_k = (ib % 8) * 8 + hm_idx;
-+
-+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
-+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
-+    const uint hm_shift = iqs_k / 8;
-+
-+    // Repack 2x4 quants into one int
-+    // Add the 3rd bit instead of subtracting it to allow packing the quants
-+    const i8vec2 vals00 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2        ] >> qs_shift) & uint16_t(0x0303))) |
-+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2    ] >> hm_shift) & uint16_t(0x0101)) << 2));
-+    const i8vec2 vals01 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 1    ] >> qs_shift) & uint16_t(0x0303))) |
-+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 1] >> hm_shift) & uint16_t(0x0101)) << 2));
-+    const i8vec2 vals10 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 2    ] >> qs_shift) & uint16_t(0x0303))) |
-+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 2] >> hm_shift) & uint16_t(0x0101)) << 2));
-+    const i8vec2 vals11 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 3    ] >> qs_shift) & uint16_t(0x0303))) |
-+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 3] >> hm_shift) & uint16_t(0x0101)) << 2));
-+    buf_a[buf_ib].qs[iqs] = pack32(u8vec4(vals00.x, vals00.y, vals01.x, vals01.y)) |
-+                           (pack32(u8vec4(vals10.x, vals10.y, vals11.x, vals11.y)) << 4);
-+
-+    if (iqs == 0) {
-+        const uint is = iqs_k / 4;
-+        const i8vec2 scales = i8vec2(unpack8(((data_a_packed16[ib_k].scales[(is % 8      ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
-+                                            (((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4)));
-+
-+        buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales - 32);
-+    }
-+}
-+
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].d_scales = buf_a[buf_ib].d_scales;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
-+}
-+
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    float result = 0.0;
-+    int32_t q_sum = 0;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
-+        // Subtract 4 from the quants to correct the 3rd bit offset
-+        const int32_t qs_a = pack32(unpack8(int32_t((cache_a[ib_a].qs[iqs / 2] >> ((iqs % 2) * 4)) & 0x0F0F0F0F)) - int8_t(4));
-+
-+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
-+    }
-+    result += float(cache_a[ib_a].d_scales[0]) * float(q_sum);
-+    q_sum = 0;
-+
-+    [[unroll]] for (uint iqs = 4; iqs < 8; iqs++) {
-+        const int32_t qs_a = pack32(unpack8(int32_t((cache_a[ib_a].qs[iqs / 2] >> ((iqs % 2) * 4)) & 0x0F0F0F0F)) - int8_t(4));
-+
-+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
-+    }
-+    result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
-+
-+    return ACC_TYPE(cache_b.ds.x * result);
-+}
-+#endif // MMQ_SHMEM
-+#endif
-+
-+#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
-+// 4-byte loads for Q4_K blocks (144 bytes) and Q5_K blocks (176 bytes)
-+ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-+    return ACC_TYPE(dsb.x * dma.x * float(q_sum) - dma.y * dsb.y);
-+}
-+
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+    const uint ib_k = ib / 8;
-+    const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
-+
-+    const uint qs_idx = (iqs_k / 16) * 8 + (iqs_k % 8);
-+    const uint qs_shift = ((iqs_k % 16) / 8) * 4;
-+
-+    // Repack 2x4 quants into one int
-+#if defined(DATA_A_Q4_K)
-+    const uint32_t vals0 = (data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x0F0F0F0F;
-+    const uint32_t vals1 = (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x0F0F0F0F;
-+
-+    buf_a[buf_ib].qs[iqs] = vals0 | (vals1 << 4);
-+#else // defined(DATA_A_Q5_K)
-+    const uint qh_idx = iqs * QUANT_R_MMQ;
-+    const uint qh_shift = iqs_k / 8;
-+
-+    buf_a[buf_ib].qs[iqs] = int32_t(((data_a_packed32[ib_k].qs[qs_idx] >> qs_shift) & 0x0F0F0F0F) |
-+                                   (((data_a_packed32[ib_k].qh[qh_idx] >> qh_shift) & 0x01010101) << 4));
-+#endif
-+
-+
-+    if (iqs == 0) {
-+        // Scale index
-+        const uint is = iqs_k / 8;
-+        u8vec2 scale_dm;
-+        if (is < 4) {
-+            scale_dm = u8vec2(data_a[ib_k].scales[is] & 0x3F, data_a[ib_k].scales[is + 4] & 0x3F);
-+        } else {
-+            scale_dm = u8vec2((data_a[ib_k].scales[is+4] & 0xF) | ((data_a[ib_k].scales[is-4] & 0xC0) >> 2),
-+                              (data_a[ib_k].scales[is+4] >>  4) | ((data_a[ib_k].scales[is  ] & 0xC0) >> 2));
-+        }
-+
-+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
-+    }
-+}
-+
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].dm = buf_a[buf_ib].dm;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 8 / QUANT_R_MMQ; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
-+}
-+
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    int32_t q_sum = 0;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
-+#if defined(DATA_A_Q4_K)
-+        const int32_t qs_a = int32_t((cache_a[ib_a].qs[iqs / 2] >> ((iqs % 2) * 4)) & 0x0F0F0F0F);
-+#else // defined(DATA_A_Q5_K)
-+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
-+#endif
-+
-+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
-+    }
-+
-+    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
-+}
-+#endif // MMQ_SHMEM
-+#endif
-+
-+#ifdef MMQ_SHMEM
-+void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+    const uint ib_outer = ib / 4;
-+    const uint ib_inner = ib % 4;
-+
-+    if (iqs == 0) {
-+        buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
-+    }
-+
-+    const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
-+    buf_b[buf_ib].qs[iqs * 4    ] = values.x;
-+    buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
-+    buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
-+    buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
-+}
-+
-+void block_b_to_registers(const uint ib) {
-+    cache_b.ds = buf_b[ib].ds;
-+    [[unroll]] for (uint iqs = 0; iqs < BK / 4; iqs++) {
-+        cache_b.qs[iqs] = buf_b[ib].qs[iqs];
-+    }
-+}
-+#endif
-+
-+#if defined(DATA_A_Q6_K)
-+// 2-byte loads for Q6_K blocks (210 bytes)
-+#ifdef MMQ_SHMEM
-+void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-+    const uint ib_k = ib / 8;
-+    const uint iqs_k = (ib % 8) * 8 + iqs;
-+
-+    const uint ql_idx = (iqs_k / 32) * 16 + iqs_k % 16;
-+    const uint ql_shift = ((iqs_k % 32) / 16) * 4;
-+
-+    const uint qh_idx = (iqs_k / 32) * 8 + iqs;
-+    const uint qh_shift = ((iqs_k % 32) / 8) * 2;
-+
-+    const i8vec2 vals00 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))) |
-+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-+    const i8vec2 vals01 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))) |
-+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-+    buf_a[buf_ib].qs[iqs] = pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y));
-+
-+    if (iqs == 0) {
-+        const uint is = iqs_k / 4;
-+        const i8vec2 scales = unpack8(data_a_packed16[ib_k].scales[is / 2]);
-+
-+        buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales);
-+    }
-+}
-+
-+void block_a_to_registers(const uint reg_ib, const uint buf_ib) {
-+    cache_a[reg_ib].d_scales = buf_a[buf_ib].d_scales;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 8; iqs++) {
-+        cache_a[reg_ib].qs[iqs] = buf_a[buf_ib].qs[iqs];
-+    }
-+}
-+
-+ACC_TYPE mmq_dot_product(const uint ib_a) {
-+    float result = 0.0;
-+    int32_t q_sum = 0;
-+
-+    [[unroll]] for (uint iqs = 0; iqs < 4; iqs++) {
-+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
-+
-+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
-+    }
-+    result += float(cache_a[ib_a].d_scales[0]) * float(q_sum);
-+    q_sum = 0;
-+
-+    [[unroll]] for (uint iqs = 4; iqs < 8; iqs++) {
-+        const int32_t qs_a = cache_a[ib_a].qs[iqs];
-+
-+        q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
-+    }
-+    result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
-+
-+    return ACC_TYPE(cache_b.ds.x * result);
-+}
-+#endif // MMQ_SHMEM
- #endif
- 
- #if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
-@@ -103,3 +568,10 @@ FLOAT_TYPE_VEC2 get_dm(uint ib) {
-     return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
- }
- #endif
-+
-+#if defined(DATA_A_Q2_K)
-+FLOAT_TYPE_VEC2 get_dm(uint ib) {
-+    const uint ib_k = ib / 8;
-+    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
-+}
-+#endif
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl
-new file mode 100644
-index 000000000..72fec4404
---- /dev/null
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl
-@@ -0,0 +1,78 @@
-+#if defined(DATA_A_Q4_0)
-+#define QUANT_R_MMQ 2
-+struct block_a_cache {
-+    uint32_t qs[16/4];
-+    FLOAT_TYPE dm;
-+};
-+#elif defined(DATA_A_Q4_1)
-+#define QUANT_R_MMQ 2
-+struct block_a_cache {
-+    uint32_t qs[16/4];
-+    FLOAT_TYPE_VEC2 dm;
-+};
-+#elif defined(DATA_A_Q5_0)
-+#define QUANT_R_MMQ 2
-+struct block_a_cache {
-+    uint32_t qs[16/4];
-+    uint32_t qh;
-+    FLOAT_TYPE dm;
-+};
-+#elif defined(DATA_A_Q5_1)
-+#define QUANT_R_MMQ 2
-+struct block_a_cache {
-+    uint32_t qs[16/4];
-+    uint32_t qh;
-+    FLOAT_TYPE_VEC2 dm;
-+};
-+#elif defined(DATA_A_Q8_0)
-+#define QUANT_R_MMQ 1
-+// AMD likes 4, Intel likes 1 and Nvidia likes 2
-+#define BK_STEP 1
-+struct block_a_cache {
-+    int32_t qs[32/4];
-+    FLOAT_TYPE dm;
-+};
-+#elif defined(DATA_A_MXFP4)
-+#define QUANT_R_MMQ 2
-+struct block_a_cache {
-+    int32_t qs[8];
-+    FLOAT_TYPE d;
-+};
-+#elif defined(DATA_A_Q2_K)
-+#define QUANT_R_MMQ 4
-+struct block_a_cache {
-+    uint32_t qs[2];
-+    u8vec2 scales;
-+    FLOAT_TYPE_VEC2 dm;
-+};
-+#elif defined(DATA_A_Q3_K)
-+#define QUANT_R_MMQ 2
-+struct block_a_cache {
-+    uint32_t qs[4];
-+    FLOAT_TYPE_VEC2 d_scales;
-+};
-+#elif defined(DATA_A_Q4_K)
-+#define QUANT_R_MMQ 2
-+struct block_a_cache {
-+    uint32_t qs[4];
-+    FLOAT_TYPE_VEC2 dm;
-+};
-+#elif defined(DATA_A_Q5_K)
-+#define QUANT_R_MMQ 1
-+struct block_a_cache {
-+    int32_t qs[8];
-+    FLOAT_TYPE_VEC2 dm;
-+};
-+#elif defined(DATA_A_Q6_K)
-+#define QUANT_R_MMQ 1
-+struct block_a_cache {
-+    int32_t qs[8];
-+    FLOAT_TYPE_VEC2 d_scales;
-+};
-+#endif
-+
-+struct block_b_cache
-+{
-+    int32_t qs[8];
-+    FLOAT_TYPE_VEC2 ds;
-+};
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl
-index 2fa54ce51..02578c77c 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl
-@@ -66,6 +66,7 @@ struct block_q4_0_packed16
- #define QUANT_AUXF 1
- #define A_TYPE block_q4_0
- #define A_TYPE_PACKED16 block_q4_0_packed16
-+#define DATA_A_QUANT_LEGACY
- #endif
- 
- #define QUANT_K_Q4_1 32
-@@ -98,6 +99,7 @@ struct block_q4_1_packed32
- #define A_TYPE block_q4_1
- #define A_TYPE_PACKED16 block_q4_1_packed16
- #define A_TYPE_PACKED32 block_q4_1_packed32
-+#define DATA_A_QUANT_LEGACY
- #endif
- 
- #define QUANT_K_Q5_0 32
-@@ -123,6 +125,7 @@ struct block_q5_0_packed16
- #define QUANT_AUXF 1
- #define A_TYPE block_q5_0
- #define A_TYPE_PACKED16 block_q5_0_packed16
-+#define DATA_A_QUANT_LEGACY
- #endif
- 
- #define QUANT_K_Q5_1 32
-@@ -158,6 +161,7 @@ struct block_q5_1_packed32
- #define A_TYPE block_q5_1
- #define A_TYPE_PACKED16 block_q5_1_packed16
- #define A_TYPE_PACKED32 block_q5_1_packed32
-+#define DATA_A_QUANT_LEGACY
- #endif
- 
- #define QUANT_K_Q8_0 32
-@@ -186,6 +190,7 @@ struct block_q8_0_packed32
- #define A_TYPE block_q8_0
- #define A_TYPE_PACKED16 block_q8_0_packed16
- #define A_TYPE_PACKED32 block_q8_0_packed32
-+#define DATA_A_QUANT_LEGACY
- #endif
- 
- #define QUANT_K_Q8_1 32
-@@ -226,21 +231,21 @@ struct block_q2_K
- {
-     uint8_t scales[QUANT_K_Q2_K/16];
-     uint8_t qs[QUANT_K_Q2_K/4];
--    f16vec2 d;
-+    f16vec2 dm;
- };
- 
- struct block_q2_K_packed16
- {
-     uint16_t scales[QUANT_K_Q2_K/16/2];
-     uint16_t qs[QUANT_K_Q2_K/4/2];
--    f16vec2 d;
-+    f16vec2 dm;
- };
- 
- struct block_q2_K_packed32
- {
-     uint32_t scales[QUANT_K_Q2_K/16/4];
-     uint32_t qs[QUANT_K_Q2_K/4/4];
--    f16vec2 d;
-+    f16vec2 dm;
- };
- 
- #if defined(DATA_A_Q2_K)
-@@ -249,6 +254,8 @@ struct block_q2_K_packed32
- #define A_TYPE block_q2_K
- #define A_TYPE_PACKED16 block_q2_K_packed16
- #define A_TYPE_PACKED32 block_q2_K_packed32
-+#define SCALES_PER_32 2
-+#define DATA_A_QUANT_K
- #endif
- 
- #define QUANT_K_Q3_K 256
-@@ -274,27 +281,28 @@ struct block_q3_K_packed16
- #define QUANT_R 1
- #define A_TYPE block_q3_K
- #define A_TYPE_PACKED16 block_q3_K_packed16
-+#define DATA_A_QUANT_K
- #endif
- 
- #define QUANT_K_Q4_K 256
- 
- struct block_q4_K
- {
--    f16vec2 d;
-+    f16vec2 dm;
-     uint8_t scales[3*QUANT_K_Q4_K/64];
-     uint8_t qs[QUANT_K_Q4_K/2];
- };
- 
- struct block_q4_K_packed16
- {
--    f16vec2 d;
-+    f16vec2 dm;
-     uint16_t scales[3*QUANT_K_Q4_K/64/2];
-     uint16_t qs[QUANT_K_Q4_K/2/2];
- };
- 
- struct block_q4_K_packed32
- {
--    f16vec2 d;
-+    f16vec2 dm;
-     uint32_t scales[3*QUANT_K_Q4_K/64/4];
-     uint32_t qs[QUANT_K_Q4_K/2/4];
- };
-@@ -310,13 +318,14 @@ struct block_q4_K_packed128
- #define A_TYPE block_q4_K
- #define A_TYPE_PACKED16 block_q4_K_packed16
- #define A_TYPE_PACKED32 block_q4_K_packed32
-+#define DATA_A_QUANT_K
- #endif
- 
- #define QUANT_K_Q5_K 256
- 
- struct block_q5_K
- {
--    f16vec2 d;
-+    f16vec2 dm;
-     uint8_t scales[12];
-     uint8_t qh[QUANT_K_Q5_K/8];
-     uint8_t qs[QUANT_K_Q5_K/2];
-@@ -324,12 +333,20 @@ struct block_q5_K
- 
- struct block_q5_K_packed16
- {
--    f16vec2 d;
-+    f16vec2 dm;
-     uint16_t scales[12/2];
-     uint16_t qh[QUANT_K_Q5_K/8/2];
-     uint16_t qs[QUANT_K_Q5_K/2/2];
- };
- 
-+struct block_q5_K_packed32
-+{
-+    f16vec2 dm;
-+    uint32_t scales[12/4];
-+    uint32_t qh[QUANT_K_Q5_K/8/4];
-+    uint32_t qs[QUANT_K_Q5_K/2/4];
-+};
-+
- struct block_q5_K_packed128
- {
-     uvec4 q5k[11];
-@@ -340,6 +357,8 @@ struct block_q5_K_packed128
- #define QUANT_R 1
- #define A_TYPE block_q5_K
- #define A_TYPE_PACKED16 block_q5_K_packed16
-+#define A_TYPE_PACKED32 block_q5_K_packed32
-+#define DATA_A_QUANT_K
- #endif
- 
- #define QUANT_K_Q6_K 256
-@@ -356,7 +375,7 @@ struct block_q6_K_packed16
- {
-     uint16_t ql[QUANT_K_Q6_K/2/2];
-     uint16_t qh[QUANT_K_Q6_K/4/2];
--    int8_t scales[QUANT_K_Q6_K/16];
-+    int16_t scales[QUANT_K_Q6_K/16/2];
-     float16_t d;
- };
- 
-@@ -365,6 +384,7 @@ struct block_q6_K_packed16
- #define QUANT_R 1
- #define A_TYPE block_q6_K
- #define A_TYPE_PACKED16 block_q6_K_packed16
-+#define DATA_A_QUANT_K
- #endif
- 
- // IQuants
-@@ -1363,18 +1383,11 @@ struct block_mxfp4
-     uint8_t qs[QUANT_K_MXFP4/2];
- };
- 
--//struct block_mxfp4_packed16
--//{
--//    uint8_t e;
--//    uint16_t qs[QUANT_K_MXFP4/2/2];
--//};
--
- #if defined(DATA_A_MXFP4)
- #define QUANT_K QUANT_K_MXFP4
- #define QUANT_R QUANT_R_MXFP4
- #define QUANT_AUXF 1
- #define A_TYPE block_mxfp4
--//#define A_TYPE_PACKED16 block_mxfp4_packed16
- #endif
- 
- #if defined(DATA_A_IQ4_NL) || defined(DATA_A_IQ4_XS)
-@@ -1397,12 +1410,12 @@ void init_iq_shmem(uvec3 wgsize)
- #endif
- 
- #if defined(DATA_A_MXFP4)
--const FLOAT_TYPE kvalues_mxfp4_const[16] = {
--    FLOAT_TYPE(0.0f), FLOAT_TYPE(0.5f), FLOAT_TYPE(1.0f), FLOAT_TYPE(1.5f), FLOAT_TYPE(2.0f), FLOAT_TYPE(3.0f), FLOAT_TYPE(4.0f), FLOAT_TYPE(6.0f),
--    FLOAT_TYPE(-0.0f), FLOAT_TYPE(-0.5f), FLOAT_TYPE(-1.0f), FLOAT_TYPE(-1.5f), FLOAT_TYPE(-2.0f), FLOAT_TYPE(-3.0f), FLOAT_TYPE(-4.0f), FLOAT_TYPE(-6.0f)
-+const int8_t kvalues_mxfp4_const[16] = {
-+    int8_t(0), int8_t(1), int8_t(2), int8_t(3), int8_t(4), int8_t(6), int8_t(8), int8_t(12),
-+    int8_t(0), int8_t(-1), int8_t(-2), int8_t(-3), int8_t(-4), int8_t(-6), int8_t(-8), int8_t(-12),
- };
- 
--shared FLOAT_TYPE kvalues_mxfp4[16];
-+shared int8_t kvalues_mxfp4[16];
- 
- #define NEEDS_INIT_IQ_SHMEM
- void init_iq_shmem(uvec3 wgsize)
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-index 0f25ba345..03fa01639 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-@@ -566,7 +566,8 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
-         }
- 
- #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
--        if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && is_legacy_quant(tname)) {
-+        // Integer dot mmq performs better with f32 accumulators
-+        if (!f16acc && !coopmat && !coopmat2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
-             string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
-         }
- #endif
-@@ -574,7 +575,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
- }
- 
- void process_shaders() {
--    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
-+    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
- 
-     // matmul
-     for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
diff --git a/llama/patches/0036-win-exit-instead-of-abort.patch b/llama/patches/0030-win-exit-instead-of-abort.patch
similarity index 95%
rename from llama/patches/0036-win-exit-instead-of-abort.patch
rename to llama/patches/0030-win-exit-instead-of-abort.patch
index ab78d631..9f1a65ea 100644
--- a/llama/patches/0036-win-exit-instead-of-abort.patch
+++ b/llama/patches/0030-win-exit-instead-of-abort.patch
@@ -8,7 +8,7 @@ Subject: [PATCH] win: exit instead of abort
  1 file changed, 6 insertions(+), 1 deletion(-)
 
 diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
-index 9be35c1be..923c33d05 100644
+index b99345a2e..1c9e0bc05 100644
 --- a/ggml/src/ggml.c
 +++ b/ggml/src/ggml.c
 @@ -229,8 +229,13 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
diff --git a/llama/patches/0031-fix-bakllava-regression.patch b/llama/patches/0031-fix-bakllava-regression.patch
new file mode 100644
index 00000000..9481f87a
--- /dev/null
+++ b/llama/patches/0031-fix-bakllava-regression.patch
@@ -0,0 +1,25 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Daniel Hiltgen <daniel@ollama.com>
+Date: Tue, 11 Nov 2025 11:39:43 -0800
+Subject: [PATCH] fix bakllava regression
+
+Rever to prior logic of assuming an empty projector type is mlp
+---
+ tools/mtmd/clip.cpp | 4 ++++
+ 1 file changed, 4 insertions(+)
+
+diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
+index f4c4d2c48..3334ff25b 100644
+--- a/tools/mtmd/clip.cpp
++++ b/tools/mtmd/clip.cpp
+@@ -2648,6 +2648,10 @@ struct clip_model_loader {
+             if (proj_type.empty()) {
+                 if (modality == CLIP_MODALITY_VISION) {
+                     get_string(KEY_VISION_PROJ_TYPE, proj_type, false);
++                    if (proj_type.empty()) {
++                        // Assume MLP if no projector type listed
++                        proj_type = "mlp";
++                    }
+                 } else if (modality == CLIP_MODALITY_AUDIO) {
+                     get_string(KEY_AUDIO_PROJ_TYPE, proj_type, false);
+                 } else {
diff --git a/llama/patches/0031-vulkan-Update-topk_moe-fusion-to-handle-gpt-s-late-s.patch b/llama/patches/0031-vulkan-Update-topk_moe-fusion-to-handle-gpt-s-late-s.patch
deleted file mode 100644
index f48e25bb..00000000
--- a/llama/patches/0031-vulkan-Update-topk_moe-fusion-to-handle-gpt-s-late-s.patch
+++ /dev/null
@@ -1,657 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Jeff Bolz <jbolz@nvidia.com>
-Date: Wed, 29 Oct 2025 08:44:29 -0500
-Subject: [PATCH] vulkan: Update topk_moe fusion to handle gpt's late softmax
- (#16656)
-
-* vulkan: Update topk_moe fusion to handle gpt's late softmax
-
-Based on #16649.
-
-* Add ggml_check_edges
-
-* Add sync logging to show fusion effects
-
-* handle clamp added in #16655
-
-* Update ggml/src/ggml-impl.h
-
-Co-authored-by: Diego Devesa <slarengh@gmail.com>
----
- ggml/src/ggml-impl.h                          |  16 +
- ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 304 +++++++++++-------
- .../ggml-vulkan/vulkan-shaders/topk_moe.comp  |  90 ++++--
- 3 files changed, 272 insertions(+), 138 deletions(-)
-
-diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
-index 639d551a2..e5c446d1d 100644
---- a/ggml/src/ggml-impl.h
-+++ b/ggml/src/ggml-impl.h
-@@ -693,6 +693,7 @@ GGML_API void ggml_dxgi_pdh_release();
- #endif
- 
- #ifdef __cplusplus
-+#include <array>
- #include <initializer_list>
- #include <vector>
- 
-@@ -708,6 +709,21 @@ inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph *          cgraph,
-     return ggml_can_fuse_subgraph(cgraph, start_idx, ops.size(), ops.begin(), outputs.begin(), outputs.size());
- }
- 
-+// Return true if the edges in the graph match expectations.
-+inline bool ggml_check_edges(const struct ggml_cgraph *                cgraph,
-+                             int                                       start_idx,
-+                             std::initializer_list<std::array<int, 3>> edges) {
-+    for (const auto & edge : edges) {
-+        int dst_node = edge[0];
-+        int src_idx  = edge[1];
-+        int src_node = edge[2];
-+        if (cgraph->nodes[start_idx + dst_node]->src[src_idx] != cgraph->nodes[start_idx + src_node]) {
-+            return false;
-+        }
-+    }
-+    return true;
-+}
-+
- // expose GGUF internals for test code
- GGML_API size_t gguf_type_size(enum gguf_type type);
- GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);
-diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 7669ed206..63a762ec2 100644
---- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -387,12 +387,76 @@ static constexpr uint32_t num_argsort_pipelines = 11;
- static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
- static constexpr uint32_t num_topk_moe_pipelines = 10;
- 
--static constexpr std::array topk_moe_norm{ GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
--                                           GGML_OP_VIEW,     GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
--                                           GGML_OP_SUM_ROWS, GGML_OP_DIV,      GGML_OP_RESHAPE };
--static constexpr std::array topk_moe     { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
--                                           GGML_OP_VIEW,     GGML_OP_GET_ROWS };
-+static constexpr std::initializer_list<ggml_op> topk_moe_early_softmax_norm{ GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
-+                                                                             GGML_OP_VIEW,     GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
-+                                                                             GGML_OP_SUM_ROWS, GGML_OP_CLAMP,    GGML_OP_DIV,
-+                                                                             GGML_OP_RESHAPE };
-+static constexpr std::initializer_list<ggml_op> topk_moe_early_softmax     { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
-+                                                                             GGML_OP_VIEW,     GGML_OP_GET_ROWS };
-+static constexpr std::initializer_list<ggml_op> topk_moe_late_softmax      { GGML_OP_ARGSORT,  GGML_OP_VIEW,
-+                                                                             GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
-+                                                                             GGML_OP_SOFT_MAX, GGML_OP_RESHAPE };
-+
-+//node #978 (  SOFT_MAX):     ffn_moe_probs-15 (   0K) [Vulka         ] use=2:    ffn_moe_logits-15 (   0K) [Vulka         ]
-+//node #979 (   RESHAPE): ffn_moe_probs-15 (re (   0K) [Vulka         ] use=1:     ffn_moe_probs-15 (   0K) [Vulka         ]
-+//node #980 (   ARGSORT):   ffn_moe_argsort-15 (   0K) [Vulka         ] use=1:     ffn_moe_probs-15 (   0K) [Vulka         ]
-+//node #981 (      VIEW):      ffn_moe_topk-15 (   0K) [Vulka         ] use=4:   ffn_moe_argsort-15 (   0K) [Vulka         ]
-+//node #982 (  GET_ROWS):   ffn_moe_weights-15 (   0K) [Vulka         ] use=1: ffn_moe_probs-15 (re (   0K) [Vulka         ]      ffn_moe_topk-15 (   0K) [Vulka         ]
-+//node #983 (   RESHAPE): ffn_moe_weights-15 ( (   0K) [Vulka         ] use=2:   ffn_moe_weights-15 (   0K) [Vulka         ]
-+//node #984 (  SUM_ROWS): ffn_moe_weights_sum- (   0K) [Vulka         ] use=1: ffn_moe_weights-15 ( (   0K) [Vulka         ]
-+//node #985 (     CLAMP): ffn_moe_weights_sum_ (   0K) [Vulka         ] use=1: ffn_moe_weights_sum- (   0K) [Vulka         ]
-+//node #986 (       DIV): ffn_moe_weights_norm (   0K) [Vulka         ] use=1: ffn_moe_weights-15 ( (   0K) [Vulka         ] ffn_moe_weights_sum_ (   0K) [Vulka         ]
-+//node #987 (   RESHAPE): ffn_moe_weights_norm (   0K) [Vulka         ] use=1: ffn_moe_weights_norm (   0K) [Vulka         ]
-+static constexpr std::initializer_list<std::array<int, 3>> topk_moe_early_softmax_norm_edges {
-+    { 1, 0, 0 }, // reshape->src[0]  == softmax
-+    { 2, 0, 0 }, // argsort->src[0]  == softmax
-+    { 3, 0, 2 }, // view->src[0]     == argsort
-+    { 4, 0, 1 }, // get_rows->src[0] == reshape
-+    { 4, 1, 3 }, // get_rows->src[1] == view
-+    { 5, 0, 4 }, // reshape->src[0]  == get_rows
-+    { 6, 0, 5 }, // sum_rows->src[0] == reshape
-+    { 7, 0, 6 }, // clamp->src[0]    == sum_rows
-+    { 8, 0, 5 }, // div->src[0]      == reshape
-+    { 8, 1, 7 }, // div->src[1]      == clamp
-+    { 9, 0, 8 }, // reshape->src[0]  == div
-+};
-+
-+// same as early_softmax_norm but ending after the get_rows
-+static constexpr std::initializer_list<std::array<int, 3>> topk_moe_early_softmax_edges {
-+    { 1, 0, 0 }, // reshape->src[0]  == softmax
-+    { 2, 0, 0 }, // argsort->src[0]  == softmax
-+    { 3, 0, 2 }, // view->src[0]     == argsort
-+    { 4, 0, 1 }, // get_rows->src[0] == reshape
-+    { 4, 1, 3 }, // get_rows->src[1] == view
-+};
- 
-+//node #652 (   ARGSORT):   ffn_moe_argsort-11 (   0K) [Vulka         ] use=1:     ffn_moe_probs-11 (   0K) [Vulka         ]
-+//node #653 (      VIEW):      ffn_moe_topk-11 (   0K) [Vulka         ] use=7:   ffn_moe_argsort-11 (   0K) [Vulka         ]
-+//node #654 (  GET_ROWS):   ffn_moe_weights-11 (   0K) [Vulka         ] use=1: ffn_moe_probs-11 (re (   0K) [Vulka         ]      ffn_moe_topk-11 (   0K) [Vulka         ]
-+//node #655 (   RESHAPE): ffn_moe_weights-11 ( (   0K) [Vulka         ] use=1:   ffn_moe_weights-11 (   0K) [Vulka         ]
-+//node #656 (  SOFT_MAX):             node_656 (   0K) [Vulka         ] use=1: ffn_moe_weights-11 ( (   0K) [Vulka         ]
-+//node #657 (   RESHAPE): ffn_moe_weights_soft (   0K) [Vulka         ] use=1:             node_656 (   0K) [Vulka         ]
-+static constexpr std::initializer_list<std::array<int, 3>> topk_moe_late_softmax_edges {
-+    { 1, 0, 0 }, // view->src[0]     == argsort
-+    { 2, 1, 1 }, // get_rows->src[1] == view
-+    { 3, 0, 2 }, // reshape->src[0]  == get_rows
-+    { 4, 0, 3 }, // soft_max->src[0] == reshape
-+    { 5, 0, 4 }, // reshape->src[0]  == soft_max
-+};
-+
-+enum topk_moe_mode {
-+    TOPK_MOE_EARLY_SOFTMAX,
-+    TOPK_MOE_EARLY_SOFTMAX_NORM,
-+    TOPK_MOE_LATE_SOFTMAX,
-+    TOPK_MOE_COUNT,
-+};
-+
-+static topk_moe_mode ggml_vk_num_additional_ops_to_topk_moe_mode(uint32_t num) {
-+    topk_moe_mode mode = num == topk_moe_early_softmax_norm.size() - 1 ? TOPK_MOE_EARLY_SOFTMAX_NORM :
-+                         num == topk_moe_early_softmax.size() - 1      ? TOPK_MOE_EARLY_SOFTMAX :
-+                                                                         TOPK_MOE_LATE_SOFTMAX;
-+    return mode;
-+}
- 
- struct vk_device_struct {
-     std::recursive_mutex mutex;
-@@ -607,8 +671,7 @@ struct vk_device_struct {
- 
-     vk_pipeline pipeline_flash_attn_split_k_reduce;
- 
--    // [2] is {!norm, norm}
--    vk_pipeline pipeline_topk_moe[num_topk_moe_pipelines][2];
-+    vk_pipeline pipeline_topk_moe[num_topk_moe_pipelines][TOPK_MOE_COUNT];
- 
-     std::vector<vk_pipeline_ref> all_pipelines;
- 
-@@ -956,6 +1019,8 @@ static_assert(sizeof(vk_op_multi_add_push_constants) <= 256);
- struct vk_op_topk_moe_push_constants {
-     uint32_t n_rows;
-     uint32_t n_expert_used;
-+    float clamp_min;
-+    float clamp_max;
- };
- 
- struct vk_op_add_id_push_constants {
-@@ -3806,8 +3871,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
-     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f16_f32, "conv2d_dw_cwhn_f16_f32", conv2d_dw_cwhn_f16_f32_len, conv2d_dw_cwhn_f16_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
- 
-     for (uint32_t i = 0; i < num_topk_moe_pipelines; ++i) {
--        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][0], "topk_moe_f32_"+std::to_string(i),   topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0}, 1, true, true);
--        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][1], "topk_moe_f32_"+std::to_string(i),   topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 1}, 1, true, true);
-+        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX],      "topk_moe_f32_early_softmax_"+std::to_string(i),       topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 0}, 1, true, true);
-+        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX_NORM], "topk_moe_f32_early_softmax_norm"+std::to_string(i),   topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 1, 0}, 1, true, true);
-+        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_LATE_SOFTMAX],       "topk_moe_f32_late_softmax"+std::to_string(i),         topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 1}, 1, true, true);
-     }
- 
-     for (auto &c : compiles) {
-@@ -8085,8 +8151,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
-         if (ctx->num_additional_fused_ops) {
-             uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
-             GGML_ASSERT(idx < num_topk_moe_pipelines);
--            bool with_norm = ctx->num_additional_fused_ops == topk_moe_norm.size() - 1;
--            return ctx->device->pipeline_topk_moe[idx][with_norm];
-+            topk_moe_mode mode = ggml_vk_num_additional_ops_to_topk_moe_mode(ctx->num_additional_fused_ops);
-+            return ctx->device->pipeline_topk_moe[idx][mode];
-         }
- 
-         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
-@@ -8141,6 +8207,13 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
-             return nullptr;
-         }
-     case GGML_OP_ARGSORT:
-+        if (ctx->num_additional_fused_ops) {
-+            uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
-+            GGML_ASSERT(idx < num_topk_moe_pipelines);
-+            topk_moe_mode mode = ggml_vk_num_additional_ops_to_topk_moe_mode(ctx->num_additional_fused_ops);
-+            return ctx->device->pipeline_topk_moe[idx][mode];
-+        }
-+
-         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
-             uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
-             return ctx->device->pipeline_argsort_f32[idx];
-@@ -9676,10 +9749,12 @@ static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& sub
- 
- static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
- 
--    bool with_norm = ctx->num_additional_fused_ops == topk_moe_norm.size() - 1;
-+    topk_moe_mode mode = ggml_vk_num_additional_ops_to_topk_moe_mode(ctx->num_additional_fused_ops);
-     ggml_tensor * logits = cgraph->nodes[node_idx + 0]->src[0];
--    ggml_tensor * weights = with_norm ? cgraph->nodes[node_idx + 8] : cgraph->nodes[node_idx + 4];
--    ggml_tensor * ids = cgraph->nodes[node_idx + 3];
-+    ggml_tensor * weights = (mode == TOPK_MOE_EARLY_SOFTMAX_NORM) ? cgraph->nodes[node_idx + 9] :
-+                            (mode == TOPK_MOE_EARLY_SOFTMAX)      ? cgraph->nodes[node_idx + 4] :
-+                                                                    cgraph->nodes[node_idx + 5];
-+    ggml_tensor * ids = (mode == TOPK_MOE_LATE_SOFTMAX) ? cgraph->nodes[node_idx + 1] : cgraph->nodes[node_idx + 3];
- 
-     GGML_ASSERT(logits->type == GGML_TYPE_F32);
-     GGML_ASSERT(weights->type == GGML_TYPE_F32);
-@@ -9738,9 +9813,14 @@ static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx,
-         GGML_ASSERT(d_ids != nullptr);
-     }
- 
--    vk_op_topk_moe_push_constants pc;
-+    vk_op_topk_moe_push_constants pc {};
-     pc.n_rows = n_rows;
-     pc.n_expert_used = n_expert_used;
-+    if (mode == TOPK_MOE_EARLY_SOFTMAX_NORM) {
-+        ggml_tensor * clamp = cgraph->nodes[node_idx + 7];
-+        pc.clamp_min = ggml_get_op_params_f32(clamp, 0);
-+        pc.clamp_max = ggml_get_op_params_f32(clamp, 1);
-+    }
- 
-     GGML_ASSERT(n_expert_used <= n_experts);
- 
-@@ -11335,7 +11415,13 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
-                 }
-             }
-         }
-+
-+#define ENABLE_SYNC_LOGGING 0
-+
-         if (need_sync) {
-+#if ENABLE_SYNC_LOGGING
-+            std::cerr <<  "sync" << std::endl;
-+#endif
-             ctx->unsynced_nodes_written.clear();
-             ctx->unsynced_nodes_read.clear();
-             ggml_vk_sync_buffers(ctx, compute_ctx);
-@@ -11353,6 +11439,18 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
-             }
-         }
-     }
-+#if ENABLE_SYNC_LOGGING
-+    if (!dryrun) {
-+        for (int i = 0; i < ctx->num_additional_fused_ops + 1; ++i) {
-+            auto *n = cgraph->nodes[node_idx + i];
-+            std::cerr << node_idx + i << " " << ggml_op_name(n->op) << " " <<  n->name;
-+            if (n->op == GGML_OP_GLU) {
-+                std::cerr << " " << ggml_glu_op_name(ggml_get_glu_op(n)) << " " << (n->src[1] ? "split" : "single") << " ";
-+            }
-+            std::cerr << std::endl;
-+        }
-+    }
-+#endif
- 
-     switch (node->op) {
-     case GGML_OP_REPEAT:
-@@ -11531,7 +11629,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
- 
-         break;
-     case GGML_OP_ARGSORT:
--        ggml_vk_argsort(ctx, compute_ctx, src0, node, dryrun);
-+        if (ctx->num_additional_fused_ops) {
-+            ggml_vk_topk_moe(ctx, compute_ctx, cgraph, node_idx, dryrun);
-+        } else {
-+            ggml_vk_argsort(ctx, compute_ctx, src0, node, dryrun);
-+        }
- 
-         break;
-     case GGML_OP_SUM:
-@@ -12329,30 +12431,27 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
- }
- 
- static bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph,
--                                      int node_idx, bool with_norm) {
-+                                      int node_idx, topk_moe_mode mode) {
- 
--    if (with_norm) {
--        if (node_idx + (int)topk_moe_norm.size() > cgraph->n_nodes) {
--            return false;
--        }
--        for (size_t i = 0; i < topk_moe_norm.size(); ++i) {
--            if (cgraph->nodes[node_idx + i]->op != topk_moe_norm[i]) {
--                return false;
--            }
--        }
--    } else {
--        if (node_idx + (int)topk_moe.size() > cgraph->n_nodes) {
--            return false;
--        }
--        for (size_t i = 0; i < topk_moe.size(); ++i) {
--            if (cgraph->nodes[node_idx + i]->op != topk_moe[i]) {
--                return false;
--            }
--        }
--    }
-+    const ggml_tensor * softmax;
-+    const ggml_tensor * weights;
- 
--    const ggml_tensor * softmax =  cgraph->nodes[node_idx + 0];
--    const ggml_tensor * weights = with_norm ? cgraph->nodes[node_idx + 8] : cgraph->nodes[node_idx + 4];
-+    switch (mode) {
-+    case TOPK_MOE_EARLY_SOFTMAX_NORM:
-+        softmax = cgraph->nodes[node_idx + 0];
-+        weights = cgraph->nodes[node_idx + 9];
-+        break;
-+    case TOPK_MOE_EARLY_SOFTMAX:
-+        softmax = cgraph->nodes[node_idx + 0];
-+        weights = cgraph->nodes[node_idx + 4];
-+        break;
-+    case TOPK_MOE_LATE_SOFTMAX:
-+        softmax = cgraph->nodes[node_idx + 4];
-+        weights = cgraph->nodes[node_idx + 5];
-+        break;
-+    default:
-+        return false;
-+    }
- 
-     const float * op_params = (const float *)softmax->op_params;
- 
-@@ -12378,60 +12477,6 @@ static bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struc
-         return false;
-     }
- 
--    // Check that the nodes don't have any unexpected uses
--    const ggml_tensor * reshape1 =  cgraph->nodes[node_idx + 1];
--    const ggml_tensor * argsort =   cgraph->nodes[node_idx + 2];
--    const ggml_tensor * view =      cgraph->nodes[node_idx + 3];
--    const ggml_tensor * get_rows =  cgraph->nodes[node_idx + 4];
--    const ggml_tensor * reshape5 =  with_norm ? cgraph->nodes[node_idx + 5] : nullptr;
--    const ggml_tensor * sum_rows =  with_norm ? cgraph->nodes[node_idx + 6] : nullptr;
--    const ggml_tensor * div =       with_norm ? cgraph->nodes[node_idx + 7] : nullptr;
--    const ggml_tensor * reshape8 =  with_norm ? cgraph->nodes[node_idx + 8] : nullptr;
--
--    // softmax is used by reshape and argsort
--    if (ggml_node_get_use_count(cgraph, node_idx) != 2 ||
--        reshape1->src[0] != softmax ||
--        argsort->src[0] != softmax) {
--        return false;
--    }
--    // reshape is used by get_rows
--    if (ggml_node_get_use_count(cgraph, node_idx + 1) != 1 ||
--        get_rows->src[0] != reshape1) {
--        return false;
--    }
--    // argsort is used by view
--    if (ggml_node_get_use_count(cgraph, node_idx + 2) != 1 ||
--        view->src[0] != argsort) {
--        return false;
--    }
--    // view is written (via argsort), we can skip checking it
--
--    if (with_norm) {
--        // get_rows is used by reshape
--        if (ggml_node_get_use_count(cgraph, node_idx + 4) != 1 ||
--            reshape5->src[0] != get_rows) {
--            return false;
--        }
--
--        // reshape is used by sum_rows and div
--        if (ggml_node_get_use_count(cgraph, node_idx + 5) != 2 ||
--            sum_rows->src[0] != reshape5 ||
--            div->src[0] != reshape5) {
--            return false;
--        }
--
--        // sum_rows is used by div
--        if (ggml_node_get_use_count(cgraph, node_idx + 6) != 1 ||
--            div->src[1] != sum_rows) {
--            return false;
--        }
--
--        // div/reshape are written
--        if (reshape8->src[0] != div) {
--            return false;
--        }
--    }
--
-     if (!ctx->device->subgroup_arithmetic ||
-         !ctx->device->subgroup_shuffle ||
-         !ctx->device->subgroup_require_full_support ||
-@@ -12517,10 +12562,18 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
-                 ctx->num_additional_fused_ops = num_adds - 1;
-             } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-                 ctx->num_additional_fused_ops = 1;
--            } else if (ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, true)) {
--                ctx->num_additional_fused_ops = topk_moe_norm.size() - 1;
--            } else if (ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, false)) {
--                ctx->num_additional_fused_ops = topk_moe.size() - 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax_norm, { i + 3, i + 9 }) &&
-+                       ggml_check_edges(cgraph, i, topk_moe_early_softmax_norm_edges) &&
-+                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX_NORM)) {
-+                ctx->num_additional_fused_ops = topk_moe_early_softmax_norm.size() - 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax, { i + 3, i + 4 }) &&
-+                       ggml_check_edges(cgraph, i, topk_moe_early_softmax_edges) &&
-+                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX)) {
-+                ctx->num_additional_fused_ops = topk_moe_early_softmax.size() - 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_late_softmax, { i + 1, i + 5 }) &&
-+                       ggml_check_edges(cgraph, i, topk_moe_late_softmax_edges) &&
-+                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_LATE_SOFTMAX)) {
-+                ctx->num_additional_fused_ops = topk_moe_late_softmax.size() - 1;
-             }
-         }
-         ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
-@@ -12618,10 +12671,18 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
-                 ctx->num_additional_fused_ops = num_adds - 1;
-             } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-                 ctx->num_additional_fused_ops = 1;
--            } else if (ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, true)) {
--                ctx->num_additional_fused_ops = topk_moe_norm.size() - 1;
--            } else if (ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, false)) {
--                ctx->num_additional_fused_ops = topk_moe.size() - 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax_norm, { i + 3, i + 9 }) &&
-+                       ggml_check_edges(cgraph, i, topk_moe_early_softmax_norm_edges) &&
-+                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX_NORM)) {
-+                ctx->num_additional_fused_ops = topk_moe_early_softmax_norm.size() - 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax, { i + 3, i + 4 }) &&
-+                       ggml_check_edges(cgraph, i, topk_moe_early_softmax_edges) &&
-+                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX)) {
-+                ctx->num_additional_fused_ops = topk_moe_early_softmax.size() - 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_late_softmax, { i + 1, i + 5 }) &&
-+                       ggml_check_edges(cgraph, i, topk_moe_late_softmax_edges) &&
-+                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_LATE_SOFTMAX)) {
-+                ctx->num_additional_fused_ops = topk_moe_late_softmax.size() - 1;
-             }
-         }
- 
-@@ -12754,25 +12815,44 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
-     while (first_unused < graph->n_nodes) {
-         std::vector<int> current_set;
- 
--        // Avoid reordering topk_moe_norm
--        if (first_unused + (int)topk_moe_norm.size() <= graph->n_nodes) {
--            bool is_topk_moe_norm = true;
--            for (size_t j = 0; j < topk_moe_norm.size(); ++j) {
--                if (graph->nodes[first_unused + j]->op != topk_moe_norm[j] || used[first_unused + j]) {
--                    is_topk_moe_norm = false;
-+        // Check for fusion patterns and avoid reordering them
-+        auto const &match_pattern = [&](const std::initializer_list<ggml_op> &pattern, int start) -> bool {
-+            if (start + (int)pattern.size() <= graph->n_nodes) {
-+                bool is_pattern = true;
-+                for (size_t j = 0; j < pattern.size(); ++j) {
-+                    if (graph->nodes[start + j]->op != pattern.begin()[j] || used[start + j]) {
-+                        is_pattern = false;
-+                    }
-                 }
-+                return is_pattern;
-             }
--            if (is_topk_moe_norm) {
--                for (size_t j = 0; j < topk_moe_norm.size(); ++j) {
-+            return false;
-+        };
-+
-+        auto const &keep_pattern = [&](const std::initializer_list<ggml_op> &pattern) -> bool {
-+            if (match_pattern(pattern, first_unused)) {
-+                for (size_t j = 0; j < pattern.size(); ++j) {
-                     new_order.push_back(graph->nodes[first_unused + j]);
-                     used[first_unused + j] = true;
-                 }
-                 while (first_unused < graph->n_nodes && used[first_unused]) {
-                     first_unused++;
-                 }
--                continue;
-+                return true;
-             }
-+            return false;
-+        };
-+
-+        if (keep_pattern(topk_moe_early_softmax_norm)) {
-+            continue;
-+        }
-+        if (keep_pattern(topk_moe_early_softmax)) {
-+            continue;
-         }
-+        if (keep_pattern(topk_moe_late_softmax)) {
-+            continue;
-+        }
-+
-         // First, grab the next unused node.
-         current_set.push_back(first_unused);
- 
-@@ -12790,6 +12870,12 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
-             if (is_empty(graph->nodes[j])) {
-                 continue;
-             }
-+            // Don't pull forward nodes from fusion patterns
-+            if (match_pattern(topk_moe_early_softmax_norm, j) ||
-+                match_pattern(topk_moe_early_softmax, j) ||
-+                match_pattern(topk_moe_late_softmax, j)) {
-+                continue;
-+            }
-             bool ok = true;
-             for (int c = first_unused; c < j; ++c) {
-                 if (!used[c] &&
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp b/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp
-index 9e56d5f8a..bc1c278bf 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp
-@@ -11,6 +11,8 @@ layout (push_constant) uniform parameter
- {
-     uint n_rows;
-     uint n_expert_used;
-+    float clamp_min;
-+    float clamp_max;
- };
- 
- layout(local_size_x_id = 0, local_size_y = 4, local_size_z = 1) in;
-@@ -18,6 +20,7 @@ layout(local_size_x_id = 0, local_size_y = 4, local_size_z = 1) in;
- layout(constant_id = 0) const uint WARP_SIZE = 32;
- layout(constant_id = 1) const uint n_experts = 512;
- layout(constant_id = 2) const bool with_norm = true;
-+layout(constant_id = 3) const bool late_softmax = false;
- 
- const uint experts_per_thread = (n_experts > WARP_SIZE) ? n_experts / WARP_SIZE : 1;
- 
-@@ -25,53 +28,72 @@ layout (binding = 0, std430) readonly buffer Logits {float logits[];};
- layout (binding = 1, std430) writeonly buffer Weights {float weights[];};
- layout (binding = 2, std430) writeonly buffer Ids {uint ids[];};
- 
--void main() {
--    const uint row = gl_WorkGroupID.x * gl_WorkGroupSize.y + gl_LocalInvocationID.y;
--    if (row >= n_rows) {
--        return;
--    }
-+const float INFINITY = 1.0 / 0.0;
- 
--    const uint logits_offset = n_experts * row;
--    const uint weights_offset = n_expert_used * row;
--    const uint ids_offset = n_experts * row;
--
--    float logits_r[experts_per_thread];
--
--    const float INFINITY = 1.0 / 0.0;
-+// Warp-local softmax used for both the pre-top-k logits and the post-top-k delayed path.
-+void softmax_warp_inplace(inout float vals[experts_per_thread], const uint limit, const uint lane, const bool use_limit) {
-+    float max_val = -INFINITY;
- 
-     [[unroll]]
--    for (uint i = 0; i < n_experts; i += WARP_SIZE) {
--        const uint expert        = i + gl_LocalInvocationID.x;
--        logits_r[i / WARP_SIZE] = n_experts % WARP_SIZE == 0 || expert < n_experts ? logits[logits_offset + expert] : -INFINITY;
-+    for (int i = 0; i < experts_per_thread; i++) {
-+        const uint idx       = lane + i * WARP_SIZE;
-+        const bool is_active = !use_limit || (idx < limit);
-+        if (is_active) {
-+            max_val = max(max_val, vals[i]);
-+        }
-     }
- 
--    float max_val = logits_r[0];
-+    max_val = subgroupMax(max_val);
-+
-+    float sum = 0.f;
- 
-     [[unroll]]
--    for (int i = 1; i < experts_per_thread; i++) {
--        const float val = logits_r[i];
--        max_val         = max(val, max_val);
-+    for (int i = 0; i < experts_per_thread; i++) {
-+        const uint idx       = lane + i * WARP_SIZE;
-+        const bool is_active = !use_limit || (idx < limit);
-+        if (is_active) {
-+            const float val = exp(vals[i] - max_val);
-+            vals[i]         = val;
-+            sum += val;
-+        } else {
-+            vals[i] = 0.f;
-+        }
-     }
- 
--    max_val = subgroupMax(max_val);
-+    sum = subgroupAdd(sum);
- 
--    float wt[experts_per_thread];
--    float tmp = 0.f;
-+    const float inv_sum = 1.0f / sum;
- 
-     [[unroll]]
-     for (int i = 0; i < experts_per_thread; i++) {
--        const float val = logits_r[i];
--        wt[i]           = exp(val - max_val);
--        tmp += wt[i];
-+        const uint idx       = lane + i * WARP_SIZE;
-+        const bool is_active = !use_limit || (idx < limit);
-+        if (is_active) {
-+            vals[i] *= inv_sum;
-+        }
-     }
-+}
- 
--    tmp = subgroupAdd(tmp);
-+void main() {
-+    const uint row = gl_WorkGroupID.x * gl_WorkGroupSize.y + gl_LocalInvocationID.y;
-+    if (row >= n_rows) {
-+        return;
-+    }
- 
--    const float inv_sum = 1.0f / tmp;
-+    const uint logits_offset = n_experts * row;
-+    const uint weights_offset = n_expert_used * row;
-+    const uint ids_offset = n_experts * row;
-+
-+    float wt[experts_per_thread];
- 
-     [[unroll]]
--    for (int i = 0; i < experts_per_thread; i++) {
--        wt[i] = wt[i] * inv_sum;
-+    for (uint i = 0; i < n_experts; i += WARP_SIZE) {
-+        const uint expert = i + gl_LocalInvocationID.x;
-+        wt[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? logits[logits_offset + expert] : -INFINITY;
-+    }
-+
-+    if (!late_softmax) {
-+        softmax_warp_inplace(wt, n_experts, gl_LocalInvocationID.x, false);
-     }
- 
-     // at this point, each thread holds a portion of softmax,
-@@ -82,6 +104,11 @@ void main() {
- 
-     float output_weights[experts_per_thread];
- 
-+    [[unroll]]
-+    for (int i = 0; i < experts_per_thread; i++) {
-+        output_weights[i] = 0.f;
-+    }
-+
-     for (int k = 0; k < n_expert_used; k++) {
-         float max_val    = wt[0];
-         uint   max_expert = gl_LocalInvocationID.x;
-@@ -121,6 +148,7 @@ void main() {
- 
-     if (with_norm) {
-         wt_sum              = subgroupAdd(wt_sum);
-+        wt_sum              = clamp(wt_sum, clamp_min, clamp_max);
-         const float inv_sum = 1.0f / wt_sum;
- 
-         [[unroll]]
-@@ -129,6 +157,10 @@ void main() {
-         }
-     }
- 
-+    if (late_softmax) {
-+        softmax_warp_inplace(output_weights, n_expert_used, gl_LocalInvocationID.x, true);
-+    }
-+
-     [[unroll]]
-     for (uint i = 0; i < experts_per_thread; ++i) {
-         uint idx = i * WARP_SIZE + gl_LocalInvocationID.x;
diff --git a/llama/patches/0032-vulkan-Fuse-rope-set_rows-16769.patch b/llama/patches/0032-vulkan-Fuse-rope-set_rows-16769.patch
deleted file mode 100644
index 27e342dc..00000000
--- a/llama/patches/0032-vulkan-Fuse-rope-set_rows-16769.patch
+++ /dev/null
@@ -1,1242 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Jeff Bolz <jbolz@nvidia.com>
-Date: Wed, 29 Oct 2025 15:13:10 -0500
-Subject: [PATCH] vulkan: Fuse rope+set_rows (#16769)
-
-This pattern appears in a lot of models, the rope operation is applied right
-before storing into the KV cache (usually on the K tensor).
-
-Add a path to some of the rope shaders that computes the destination address
-based on the set_rows tensor. Compile variants of the shader with D_TYPE of
-f16 (the usual KV cache type).
-
-Add a src3 operand to ggml_vk_op_f32 - sometimes rope uses three srcs and needs
-the fourth for the row indices.
-
-Add fused_ops_write_mask to indicate which intermediate tensors need to write
-their results to memory. Skipping writing the roped K value helps to allow more
-nodes to run concurrently.
-
-Add logic to ggml_vk_graph_optimize to make ROPE+VIEW+SET_ROWS consecutive. It
-rarely starts out that way in the graph.
-
-Add new backend tests.
----
- ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 334 +++++++++++++-----
- .../ggml-vulkan/vulkan-shaders/rope_head.glsl |   2 +
- .../ggml-vulkan/vulkan-shaders/rope_neox.comp |  13 +-
- .../ggml-vulkan/vulkan-shaders/rope_norm.comp |  13 +-
- .../vulkan-shaders/vulkan-shaders-gen.cpp     |   4 +
- tests/test-backend-ops.cpp                    | 122 +++++--
- 6 files changed, 371 insertions(+), 117 deletions(-)
-
-diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 63a762ec2..db92a7901 100644
---- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -458,6 +458,11 @@ static topk_moe_mode ggml_vk_num_additional_ops_to_topk_moe_mode(uint32_t num) {
-     return mode;
- }
- 
-+static constexpr std::initializer_list<std::array<int, 3>> rope_view_set_rows_edges {
-+    { 1, 0, 0 }, // view->src[0]     == rope
-+    { 2, 0, 1 }, // set_rows->src[0] == view
-+};
-+
- struct vk_device_struct {
-     std::recursive_mutex mutex;
- 
-@@ -640,8 +645,8 @@ struct vk_device_struct {
-     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
-     vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
-     vk_pipeline pipeline_soft_max_back_f32;
--    vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
--    vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
-+    vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16, pipeline_rope_norm_f32_f16;
-+    vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16, pipeline_rope_neox_f32_f16;
-     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
-     vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
-     vk_pipeline pipeline_argsort_f32[num_argsort_pipelines];
-@@ -1054,6 +1059,7 @@ struct vk_op_rope_push_constants {
-     uint32_t s2;
-     int32_t sections[4];
-     uint32_t is_back;
-+    uint32_t set_rows_stride;
- };
- 
- struct vk_op_soft_max_push_constants {
-@@ -1563,6 +1569,10 @@ struct ggml_backend_vk_context {
-     // number of additional consecutive nodes that are being fused with the
-     // node currently being processed
-     int num_additional_fused_ops {};
-+    // Bitmask of which fused ops need to write an intermediate value to memory.
-+    // Bit 'i' means nodes[start_of_fusion + i] writes to memory.
-+    // If there's no fusion, bit 0 is still set.
-+    int fused_ops_write_mask {};
- };
- 
- static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
-@@ -3697,21 +3707,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
-     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
-     ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1, true);
- 
--    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--    ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f32, "rope_multi_f32", rope_multi_f32_len, rope_multi_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--    ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f32, "rope_vision_f32", rope_vision_f32_len, rope_vision_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+    ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f32, "rope_multi_f32", rope_multi_f32_len, rope_multi_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+    ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f32, "rope_vision_f32", rope_vision_f32_len, rope_vision_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
- 
-     if (device->float_controls_rte_fp16) {
--        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_rte_len, rope_multi_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_rte_len, rope_vision_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_rte_len, rope_multi_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_rte_len, rope_vision_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32_f16, "rope_norm_f32_f16", rope_norm_f32_f16_rte_len, rope_norm_f32_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32_f16, "rope_neox_f32_f16", rope_neox_f32_f16_rte_len, rope_neox_f32_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-     } else {
--        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_len, rope_multi_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
--        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_len, rope_multi_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32_f16, "rope_norm_f32_f16", rope_norm_f32_f16_len, rope_norm_f32_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32_f16, "rope_neox_f32_f16", rope_neox_f32_f16_len, rope_neox_f32_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-     }
- 
-     for (uint32_t i = 0; i < num_argsort_pipelines; ++i) {
-@@ -8170,7 +8186,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
-     case GGML_OP_ROPE:
-     case GGML_OP_ROPE_BACK:
-         {
--            const int mode = ((const int32_t *) dst->op_params)[2];
-+            const ggml_tensor *rope = ctx->num_additional_fused_ops == 2 ? dst->src[0]->src[0] : dst;
-+            const int mode = ((const int32_t *) rope->op_params)[2];
-             const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-             const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
-             const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
-@@ -8179,6 +8196,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
-                 if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-                     return ctx->device->pipeline_rope_neox_f32;
-                 }
-+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
-+                    return ctx->device->pipeline_rope_neox_f32_f16;
-+                }
-                 if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-                     return ctx->device->pipeline_rope_neox_f16;
-                 }
-@@ -8200,6 +8220,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
-                 if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-                     return ctx->device->pipeline_rope_norm_f32;
-                 }
-+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
-+                    return ctx->device->pipeline_rope_norm_f32_f16;
-+                }
-                 if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-                     return ctx->device->pipeline_rope_norm_f16;
-                 }
-@@ -8409,20 +8432,22 @@ static uint32_t get_misalign_bytes(ggml_backend_vk_context * ctx, const ggml_ten
-     return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
- }
- 
--template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
-+template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-     GGML_UNUSED(p);
-     GGML_UNUSED(src0);
-     GGML_UNUSED(src1);
-     GGML_UNUSED(src2);
-+    GGML_UNUSED(src3);
-     GGML_UNUSED(dst);
-     static_assert(!std::is_const<T>::value, "unexpected type");
-     GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0);
-     GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0);
-     GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0);
-+    GGML_ASSERT(!src3 || get_misalign_bytes(ctx, src3) == 0);
-     GGML_ASSERT(!dst  || get_misalign_bytes(ctx, dst) == 0);
- }
- 
--template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
-+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
-     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
- 
-@@ -8430,9 +8455,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
- 
-     GGML_UNUSED(src1);
-     GGML_UNUSED(src2);
-+    GGML_UNUSED(src3);
- }
- 
--template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_sum_rows_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
-+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_sum_rows_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
-     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
- 
-@@ -8440,9 +8466,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
- 
-     GGML_UNUSED(src1);
-     GGML_UNUSED(src2);
-+    GGML_UNUSED(src3);
- }
- 
--template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_pad_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
-+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_pad_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
-     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
- 
-@@ -8450,9 +8477,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
- 
-     GGML_UNUSED(src1);
-     GGML_UNUSED(src2);
-+    GGML_UNUSED(src3);
- }
- 
--template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_im2col_3d_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
-+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_im2col_3d_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-     const uint32_t a_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
-     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
- 
-@@ -8460,9 +8488,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
- 
-     GGML_UNUSED(src0);
-     GGML_UNUSED(src2);
-+    GGML_UNUSED(src3);
- }
- 
--template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
-+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
-     const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
-     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
-@@ -8472,9 +8501,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
-     p.misalign_offsets = (a_offset << 16) | (b_offset << 8) | d_offset;
- 
-     GGML_UNUSED(src2);
-+    GGML_UNUSED(src3);
- }
- 
--template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
-+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
-     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
- 
-@@ -8483,10 +8513,11 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
- 
-     GGML_UNUSED(src1);
-     GGML_UNUSED(src2);
-+    GGML_UNUSED(src3);
- }
- 
- template<typename PC>
--static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
-+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
-     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-     if (src1 != nullptr) {
-         std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-@@ -8494,6 +8525,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-     if (src2 != nullptr) {
-         std::cerr << "), (" << src2 << ", name=" << src2->name << ", type=" << src2->type << ", ne0=" << src2->ne[0] << ", ne1=" << src2->ne[1] << ", ne2=" << src2->ne[2] << ", ne3=" << src2->ne[3] << ", nb0=" << src2->nb[0] << ", nb1=" << src2->nb[1] << ", nb2=" << src2->nb[2] << ", nb3=" << src2->nb[3];
-     }
-+    if (src3 != nullptr) {
-+        std::cerr << "), (" << src3 << ", name=" << src3->name << ", type=" << src3->type << ", ne0=" << src3->ne[0] << ", ne1=" << src3->ne[1] << ", ne2=" << src3->ne[2] << ", ne3=" << src3->ne[3] << ", nb0=" << src3->nb[0] << ", nb1=" << src3->nb[1] << ", nb2=" << src3->nb[2] << ", nb3=" << src3->nb[3];
-+    }
-     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
-     std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")");
-     GGML_ASSERT(op == GGML_OP_GET_ROWS || op == GGML_OP_CPY || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
-@@ -8520,6 +8554,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-     const uint64_t ne23 = use_src2 ? src2->ne[3] : 0;
-     const uint64_t ne2 = ne20 * ne21;
- 
-+    const bool use_src3 = src3 != nullptr;
-+    const uint64_t ne30 = use_src3 ? src3->ne[0] : 0;
-+    const uint64_t ne31 = use_src3 ? src3->ne[1] : 0;
-+    const uint64_t ne32 = use_src3 ? src3->ne[2] : 0;
-+    const uint64_t ne33 = use_src3 ? src3->ne[3] : 0;
-+    const uint64_t ne3 = ne30 * ne31;
-+
-     const uint64_t ned0 = dst->ne[0];
-     const uint64_t ned1 = dst->ne[1];
-     const uint64_t ned2 = dst->ne[2];
-@@ -8550,6 +8591,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
-     ggml_backend_vk_buffer_context * src1_buf_ctx = use_src1 ? (ggml_backend_vk_buffer_context *)src1->buffer->context : nullptr;
-     ggml_backend_vk_buffer_context * src2_buf_ctx = use_src2 ? (ggml_backend_vk_buffer_context *)src2->buffer->context : nullptr;
-+    ggml_backend_vk_buffer_context * src3_buf_ctx = use_src3 ? (ggml_backend_vk_buffer_context *)src3->buffer->context : nullptr;
- 
-     vk_buffer d_X = nullptr;
-     size_t x_buf_offset = 0;
-@@ -8557,10 +8599,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-     size_t y_buf_offset = 0;
-     vk_buffer d_Z = nullptr;
-     size_t z_buf_offset = 0;
-+    vk_buffer d_W = nullptr;
-+    size_t w_buf_offset = 0;
- 
-     bool src0_uma = false;
-     bool src1_uma = false;
-     bool src2_uma = false;
-+    bool src3_uma = false;
- 
-     if (ctx->device->uma) {
-         ggml_vk_host_get(ctx->device, src0->data, d_X, x_buf_offset);
-@@ -8573,6 +8618,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-             ggml_vk_host_get(ctx->device, src2->data, d_Z, z_buf_offset);
-             src2_uma = d_Z != nullptr;
-         }
-+        if (use_src3) {
-+            ggml_vk_host_get(ctx->device, src3->data, d_W, w_buf_offset);
-+            src3_uma = d_W != nullptr;
-+        }
-     }
- 
-     vk_buffer d_D = dst_buf_ctx->dev_buffer;
-@@ -8594,11 +8643,17 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-         z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
-         GGML_ASSERT(d_Z != nullptr);
-     }
-+    if (use_src3 && !src3_uma) {
-+        d_W = src3_buf_ctx->dev_buffer;
-+        w_buf_offset = vk_tensor_offset(src3) + src3->view_offs;
-+        GGML_ASSERT(d_W != nullptr);
-+    }
-     // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets.
--    init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, dst);
-+    init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, src3, dst);
-     x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-     y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-     z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-+    w_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-     d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
- 
-     std::array<uint32_t, 3> elements;
-@@ -8799,12 +8854,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-         break;
-     }
- 
--    uint64_t x_sz, y_sz, z_sz, d_sz;
-+    uint64_t x_sz, y_sz, z_sz, w_sz, d_sz;
- 
-     if (op_supports_incontiguous) {
-         x_sz = ggml_nbytes(src0) + get_misalign_bytes(ctx, src0);
-         y_sz = use_src1 ? ggml_nbytes(src1) + get_misalign_bytes(ctx, src1) : 0;
-         z_sz = use_src2 ? ggml_nbytes(src2) + get_misalign_bytes(ctx, src2) : 0;
-+        w_sz = use_src3 ? ggml_nbytes(src3) + get_misalign_bytes(ctx, src3) : 0;
-         d_sz = ggml_nbytes(dst) + get_misalign_bytes(ctx, dst);
- 
-         if (x_buf_offset + x_sz >= d_X->size) {
-@@ -8816,6 +8872,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-         if (use_src2 && z_buf_offset + z_sz >= d_Z->size) {
-             z_sz = ggml_vk_get_max_buffer_range(ctx, d_Z, z_buf_offset);
-         }
-+        if (use_src3 && w_buf_offset + w_sz >= d_W->size) {
-+            w_sz = ggml_vk_get_max_buffer_range(ctx, d_W, w_buf_offset);
-+        }
-         if (d_buf_offset + d_sz >= d_D->size) {
-             d_sz = ggml_vk_get_max_buffer_range(ctx, d_D, d_buf_offset);
-         }
-@@ -8823,6 +8882,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-         x_sz = ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0 * ne02 * ne03;
-         y_sz = use_src1 ? ggml_type_size(src1->type) * ne1 * ne12 * ne13 : 0;
-         z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 * ne22 * ne23 : 0;
-+        w_sz = use_src3 ? ggml_type_size(src3->type) * ne3 * ne32 * ne33 : 0;
-         d_sz = ggml_type_size(dst->type) * ned * ned2 * ned3;
-     }
- 
-@@ -8864,14 +8924,19 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
-     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
-         // Empty src2 is possible in rope, but the shader needs a buffer
--        vk_subbuffer subbuf_z;
-+        vk_subbuffer subbuf_z, subbuf_w;
-         if (use_src2) {
-             subbuf_z = { d_Z, z_buf_offset, z_sz };
-         } else {
-             subbuf_z = { d_X, 0, x_sz };
-         }
-+        if (use_src3) {
-+            subbuf_w = { d_W, w_buf_offset, w_sz };
-+        } else {
-+            subbuf_w = { d_X, 0, x_sz };
-+        }
- 
--        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
-+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz }, subbuf_w }, pc, elements);
-     } else if (op == GGML_OP_IM2COL || op == GGML_OP_IM2COL_3D) {
-         if (ctx->device->shader_int64 && ctx->device->buffer_device_address) {
-             // buffer device address path doesn't use dst buffer
-@@ -8887,6 +8952,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-     } else if (op == GGML_OP_OPT_STEP_SGD) {
-         // OPT_STEP_SGD works on src0, it does not need dst
-         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz } }, pc, elements);
-+    } else if (use_src3) {
-+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_W, w_buf_offset, w_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
-     } else if (use_src2) {
-         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
-     } else if (use_src1) {
-@@ -8901,7 +8968,7 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
-     const uint32_t src1_type_size = ggml_type_size(src1->type);
-     const uint32_t dst_type_size = ggml_type_size(dst->type);
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GET_ROWS, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_GET_ROWS, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -8921,7 +8988,7 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
-     // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-     int offset = dst->op_params[3] / 4; // offset in bytes
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ACC, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ACC, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9046,7 +9113,7 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const
-     const uint32_t src1_type_size = ggml_type_size(src1->type);
-     const uint32_t dst_type_size = ggml_type_size(dst->type);
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ADD, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ADD, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9061,7 +9128,7 @@ static void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const
-     const uint32_t src1_type_size = ggml_type_size(src1->type);
-     const uint32_t dst_type_size = ggml_type_size(dst->type);
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SUB, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SUB, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9076,7 +9143,7 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const
-     const uint32_t src1_type_size = ggml_type_size(src1->type);
-     const uint32_t dst_type_size = ggml_type_size(dst->type);
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_MUL, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_MUL, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9091,7 +9158,7 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
-     const uint32_t src1_type_size = ggml_type_size(src1->type);
-     const uint32_t dst_type_size = ggml_type_size(dst->type);
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_DIV, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_DIV, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9106,7 +9173,7 @@ static void ggml_vk_add_id(ggml_backend_vk_context * ctx, vk_context& subctx, co
-     const uint32_t src1_type_size = ggml_type_size(src1->type);
-     const uint32_t src2_type_size = ggml_type_size(src2->type);
- 
--    ggml_vk_op_f32<vk_op_add_id_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ADD_ID, {
-+    ggml_vk_op_f32<vk_op_add_id_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_ADD_ID, {
-         (uint32_t)dst->ne[0],
-         (uint32_t)dst->ne[1],
-         (uint32_t)src0->nb[1] / src0_type_size,
-@@ -9339,7 +9406,7 @@ static void ggml_vk_ssm_conv(ggml_backend_vk_context * ctx, vk_context& subctx,
-     const ggml_tensor * src0 = dst->src[0];
-     const ggml_tensor * src1 = dst->src[1];
- 
--    ggml_vk_op_f32<vk_op_ssm_conv_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SSM_CONV, {
-+    ggml_vk_op_f32<vk_op_ssm_conv_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SSM_CONV, {
-         (uint32_t)src0->nb[1], (uint32_t)src0->nb[2],
-         (uint32_t)src1->nb[1],
-         (uint32_t)dst->nb[0], (uint32_t)dst->nb[1], (uint32_t)dst->nb[2],
-@@ -9457,7 +9524,7 @@ static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& su
- static void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
-     const size_t n = ggml_nelements(dst->src[0]);
- 
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
- }
- 
- static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-@@ -9467,7 +9534,7 @@ static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, co
-     const uint32_t src1_type_size = ggml_type_size(src1->type);
-     const uint32_t dst_type_size = ggml_type_size(dst->type);
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONCAT, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONCAT, {
-         (uint32_t)ggml_nelements(dst),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9491,7 +9558,7 @@ static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, c
-         sf1 = (float)(dst->ne[1] - 1) / (src0->ne[1] - 1);
-     }
- 
--    ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
-+    ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
-         (uint32_t)ggml_nelements(dst), 0, 0,
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1],
-         (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-@@ -9505,23 +9572,23 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, con
-     p.param1 = ggml_get_op_params_f32(dst, 0);
-     p.param2 = ggml_get_op_params_f32(dst, 1);
- 
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SCALE, std::move(p), dryrun);
- }
- 
- static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
- }
- 
- static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
- }
- 
- static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
- }
- 
- static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_COS, vk_op_unary_push_constants_init(src0, dst), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_COS, vk_op_unary_push_constants_init(src0, dst), dryrun);
- }
- 
- static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-@@ -9529,12 +9596,12 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, con
-     p.param1 = ggml_get_op_params_f32(dst, 0);
-     p.param2 = ggml_get_op_params_f32(dst, 1);
- 
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CLAMP, std::move(p), dryrun);
- }
- 
- static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     vk_op_pad_push_constants p = vk_op_pad_push_constants_init(src0, dst);
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p), dryrun);
- }
- 
- static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-@@ -9549,17 +9616,17 @@ static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, cons
-     memcpy(&p.param1, &s01_packed, sizeof(float));
-     memcpy(&p.param2, &s23_packed, sizeof(float));
- 
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ROLL, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ROLL, std::move(p), dryrun);
- }
- 
- static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT, std::move(p), dryrun);
- }
- 
- static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, std::move(p), dryrun);
- }
- 
- static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-@@ -9575,7 +9642,7 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
-     }
- 
-     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ne);
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CPY, std::move(p), dryrun);
- }
- 
- static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-@@ -9590,7 +9657,7 @@ static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
-         return;
-     }
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SET_ROWS, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SET_ROWS, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9601,13 +9668,13 @@ static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
- }
- 
- static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
- }
- 
- static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     float * op_params = (float *)dst->op_params;
- 
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
- }
- 
- static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-@@ -9618,7 +9685,7 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
-     const float eps = float_op_params[1];
-     const uint32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
- 
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
- }
- 
- static uint32_t ggml_vk_rms_num_partials(ggml_backend_vk_context * ctx, const ggml_tensor *node) {
-@@ -9641,7 +9708,7 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
- 
-     uint32_t param3 = ctx->do_add_rms_partials ? ggml_vk_rms_num_partials(ctx, dst) : 0;
- 
--    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM, {
-+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM, {
-         (uint32_t)ggml_nelements(src0),
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
-@@ -9658,16 +9725,16 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
- 
- static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-     float * op_params = (float *)dst->op_params;
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
- }
- 
- static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     float * op_params = (float *)dst->op_params;
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
- }
- 
- static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
- }
- 
- static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-@@ -9690,7 +9757,7 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
- 
-     const uint32_t mode = split ? 2 : (swapped ? 1 : 0);
- 
--    ggml_vk_op_f32<vk_op_glu_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GLU,
-+    ggml_vk_op_f32<vk_op_glu_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_GLU,
-         {
-             (uint32_t)ggml_nelements(dst),
-             (uint32_t)src0->ne[0],
-@@ -9703,7 +9770,7 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
- 
- static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     int32_t * op_params = (int32_t *)dst->op_params;
--    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
-+    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
- }
- 
- static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
-@@ -9728,7 +9795,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
-     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
- 
--    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_SOFT_MAX, {
-+    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_SOFT_MAX, {
-         ncols,
-         src1 != nullptr ? nrows_y : (uint32_t)0,
-         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
-@@ -9744,7 +9811,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
- 
- static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-     float * op_params = (float *)dst->op_params;
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), op_params[0], op_params[1] }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), op_params[0], op_params[1] }, dryrun);
- }
- 
- static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
-@@ -9835,7 +9902,12 @@ static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx,
-         }, pc, elements);
- }
- 
--static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool backprop, bool dryrun = false) {
-+static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_cgraph * cgraph, int node_idx, bool backprop, bool dryrun = false) {
-+    ggml_tensor * dst = cgraph->nodes[node_idx];
-+    const ggml_tensor * src0 = dst->src[0];
-+    const ggml_tensor * src1 = dst->src[1];
-+    const ggml_tensor * src2 = dst->src[2];
-+    const ggml_tensor * src3 = nullptr;
-     const int n_dims        = ((int32_t *) dst->op_params)[1];
-     const int mode          = ((int32_t *) dst->op_params)[2];
-     // const int n_ctx         = ((int32_t *) dst->op_params)[3];
-@@ -9859,11 +9931,20 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
-     uint32_t s1 = src0->nb[1] / ggml_type_size(src0->type);
-     uint32_t s2 = src0->nb[2] / ggml_type_size(src0->type);
- 
--    ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ROPE, {
-+    uint32_t set_rows_stride = 0;
-+    // Fused rope + view + set_rows passes the set_rows destination stride in set_rows_stride
-+    // and overrides the dst and sets src3=row_indices
-+    if (ctx->num_additional_fused_ops > 0) {
-+        set_rows_stride = cgraph->nodes[node_idx + 2]->nb[1] / ggml_type_size(cgraph->nodes[node_idx + 2]->type);
-+        src3 = cgraph->nodes[node_idx + 2]->src[1];
-+        dst = cgraph->nodes[node_idx + 2];
-+    }
-+
-+    ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, src3, dst, GGML_OP_ROPE, {
-         (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
-         freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
-         src2 != nullptr, (uint32_t)src0->ne[2], s1, s2,
--        { sections[0], sections[1], sections[2], sections[3] }, backprop
-+        { sections[0], sections[1], sections[2], sections[3] }, backprop, set_rows_stride,
-     }, dryrun);
- }
- 
-@@ -9872,7 +9953,7 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
- 
-     uint32_t ncols = src0->ne[0];
- 
--    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
-+    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
-         ncols,
-         op_params[0],
-     }, dryrun);
-@@ -9880,26 +9961,26 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
- 
- static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, ggml_nelements(src0));
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, p, dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM, p, dryrun);
- }
- 
- static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p, dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p, dryrun);
- }
- 
- static void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
-     p.weight = 1.0f / (float)src0->ne[0];
--    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_MEAN, p, dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_MEAN, p, dryrun);
- }
- 
- static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
- }
- 
- static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_COUNT_EQUAL, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_COUNT_EQUAL, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
- }
- 
- static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-@@ -9932,7 +10013,7 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co
- 
-     const vk::DeviceAddress dst_addr = d_buf->bda_addr + vk_tensor_offset(dst) + dst->view_offs;
- 
--    ggml_vk_op_f32<vk_op_im2col_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_IM2COL, {
-+    ggml_vk_op_f32<vk_op_im2col_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_IM2COL, {
-         dst_addr,
-         batch_offset, offset_delta,
-         IC, IW, IH, OW, OH, KW, KH,
-@@ -10005,7 +10086,7 @@ static void ggml_vk_im2col_3d(ggml_backend_vk_context * ctx, vk_context& subctx,
-     pc.OH_OW_IC_KD_KH_KW = OH*OW*IC*KD*KH*KW;
-     pc.OW_IC_KD_KH_KW = OW*IC*KD*KH*KW;
- 
--    ggml_vk_op_f32<vk_op_im2col_3d_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_IM2COL_3D, std::move(pc), dryrun);
-+    ggml_vk_op_f32<vk_op_im2col_3d_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_IM2COL_3D, std::move(pc), dryrun);
- }
- 
- static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-@@ -10013,7 +10094,7 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
-     const uint32_t max_period = dst->op_params[1];
-     const uint32_t nb1 = dst->nb[1] / ggml_type_size(dst->type);
- 
--    ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, {
-+    ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, {
-         nb1, dim, max_period,
-     }, dryrun);
- }
-@@ -10046,7 +10127,7 @@ static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context&
-     p.nb1 = static_cast<uint32_t>(nb1 / nb0);
-     p.s0 = static_cast<uint32_t>(s0);
- 
--    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
- }
- 
- static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-@@ -10069,7 +10150,7 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
- 
-     const uint32_t parallel_elements = N * OC * OH * OW;
- 
--    ggml_vk_op_f32<vk_op_pool2d_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_POOL_2D, {
-+    ggml_vk_op_f32<vk_op_pool2d_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_POOL_2D, {
-         IW, IH, OW, OH, OC,
-         parallel_elements,
-         op,
-@@ -10123,7 +10204,7 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
-     GGML_ASSERT(ne03 == ne2);
-     GGML_ASSERT(ne02 == ne12);
- 
--    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
- }
- 
- static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
-@@ -10172,7 +10253,7 @@ static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context
-     GGML_ASSERT(ne02 == ne2);
-     GGML_ASSERT(ne03 == ne12);
- 
--    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p), dryrun);
- }
- 
- static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-@@ -10196,12 +10277,12 @@ static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx
-     GGML_ASSERT(src0->ne[3] == p.channels);
-     GGML_ASSERT(src1->ne[3] == p.batches);
- 
--    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
-+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
- }
- 
- static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-     const float * op_params = (const float *)dst->op_params;
--    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
-+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
- }
- 
- #ifdef GGML_VULKAN_RUN_TESTS
-@@ -11327,7 +11408,6 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
-         case GGML_OP_DIAG_MASK_INF:
-         case GGML_OP_SOFT_MAX:
-         case GGML_OP_SOFT_MAX_BACK:
--        case GGML_OP_ROPE:
-         case GGML_OP_ROPE_BACK:
-         case GGML_OP_ARGSORT:
-         case GGML_OP_SUM:
-@@ -11401,9 +11481,12 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
-         // nodes require synchronization.
-         for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1 && !need_sync; ++i) {
-             const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
--            if (overlaps_unsynced(cur_node, ctx->unsynced_nodes_read) || overlaps_unsynced(cur_node, ctx->unsynced_nodes_written)) {
--                need_sync = true;
--                break;
-+            // If the node actually writes to memory, then check if it needs to sync
-+            if (ctx->fused_ops_write_mask & (1 << i)) {
-+                if (overlaps_unsynced(cur_node, ctx->unsynced_nodes_read) || overlaps_unsynced(cur_node, ctx->unsynced_nodes_written)) {
-+                    need_sync = true;
-+                    break;
-+                }
-             }
-             for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
-                 if (!cur_node->src[j]) {
-@@ -11430,7 +11513,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
-         for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1; ++i) {
-             const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
-             // Multiple outputs could be written, e.g. in topk_moe. Add them all to the list.
--            ctx->unsynced_nodes_written.push_back(cur_node);
-+            if (ctx->fused_ops_write_mask & (1 << i)) {
-+                ctx->unsynced_nodes_written.push_back(cur_node);
-+            }
-             for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
-                 if (!cur_node->src[j]) {
-                     continue;
-@@ -11621,11 +11706,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
- 
-         break;
-     case GGML_OP_ROPE:
--        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, false, dryrun);
-+        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, false, dryrun);
- 
-         break;
-     case GGML_OP_ROPE_BACK:
--        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, true, dryrun);
-+        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, true, dryrun);
- 
-         break;
-     case GGML_OP_ARGSORT:
-@@ -12487,6 +12572,41 @@ static bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struc
-     return true;
- }
- 
-+static bool ggml_vk_can_fuse_rope_set_rows(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph,
-+                                           int node_idx) {
-+    GGML_UNUSED(ctx);
-+    const ggml_tensor *rope = cgraph->nodes[node_idx + 0];
-+    const ggml_tensor *view = cgraph->nodes[node_idx + 1];
-+    const ggml_tensor *set_rows = cgraph->nodes[node_idx + 2];
-+
-+    // ne3 not tested
-+    if (rope->src[0]->ne[3] != 1) {
-+        return false;
-+    }
-+
-+    if (set_rows->type != GGML_TYPE_F32 && set_rows->type != GGML_TYPE_F16) {
-+        return false;
-+    }
-+
-+    if (set_rows->src[1]->type != GGML_TYPE_I64) {
-+        return false;
-+    }
-+
-+    // The view should flatten two dims of rope into one dim
-+    if (!ggml_is_contiguous(view) ||
-+        view->ne[0] != rope->ne[0] * rope->ne[1]) {
-+        return false;
-+    }
-+
-+    // Only norm/neox shaders have the fusion code
-+    const int mode = ((const int32_t *) rope->op_params)[2];
-+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
-+        return false;
-+    }
-+
-+    return true;
-+}
-+
- static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
- 
-     const ggml_tensor *first_node = cgraph->nodes[node_idx];
-@@ -12562,6 +12682,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
-                 ctx->num_additional_fused_ops = num_adds - 1;
-             } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-                 ctx->num_additional_fused_ops = 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 2 }) &&
-+                       ggml_check_edges(cgraph, i, rope_view_set_rows_edges) &&
-+                       ggml_vk_can_fuse_rope_set_rows(ctx, cgraph, i)) {
-+                ctx->num_additional_fused_ops = 2;
-             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax_norm, { i + 3, i + 9 }) &&
-                        ggml_check_edges(cgraph, i, topk_moe_early_softmax_norm_edges) &&
-                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX_NORM)) {
-@@ -12671,20 +12795,31 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
-                 ctx->num_additional_fused_ops = num_adds - 1;
-             } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-                 ctx->num_additional_fused_ops = 1;
-+            } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 2 }) &&
-+                       ggml_check_edges(cgraph, i, rope_view_set_rows_edges) &&
-+                       ggml_vk_can_fuse_rope_set_rows(ctx, cgraph, i)) {
-+                ctx->num_additional_fused_ops = 2;
-             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax_norm, { i + 3, i + 9 }) &&
-                        ggml_check_edges(cgraph, i, topk_moe_early_softmax_norm_edges) &&
-                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX_NORM)) {
-                 ctx->num_additional_fused_ops = topk_moe_early_softmax_norm.size() - 1;
-+                // view of argsort writes to memory
-+                ctx->fused_ops_write_mask |= 1 << 3;
-             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax, { i + 3, i + 4 }) &&
-                        ggml_check_edges(cgraph, i, topk_moe_early_softmax_edges) &&
-                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX)) {
-                 ctx->num_additional_fused_ops = topk_moe_early_softmax.size() - 1;
-+                // view of argsort writes to memory
-+                ctx->fused_ops_write_mask |= 1 << 3;
-             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_late_softmax, { i + 1, i + 5 }) &&
-                        ggml_check_edges(cgraph, i, topk_moe_late_softmax_edges) &&
-                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_LATE_SOFTMAX)) {
-                 ctx->num_additional_fused_ops = topk_moe_late_softmax.size() - 1;
-+                // view of argsort writes to memory
-+                ctx->fused_ops_write_mask |= 1 << 1;
-             }
-         }
-+        ctx->fused_ops_write_mask |= 1 << ctx->num_additional_fused_ops;
- 
-         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
-         bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5;
-@@ -12730,6 +12865,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
-         }
-         i += ctx->num_additional_fused_ops;
-         ctx->num_additional_fused_ops = 0;
-+        ctx->fused_ops_write_mask = 0;
-     }
- 
-     if (vk_perf_logger_enabled) {
-@@ -12887,6 +13023,32 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
-             }
-             if (ok) {
-                 current_set.push_back(j);
-+                // Look for ROPE + VIEW + SET_ROWS and make them consecutive
-+                if (graph->nodes[j]->op == GGML_OP_ROPE) {
-+                    int view_idx = -1;
-+                    int set_rows_idx = -1;
-+                    for (int k = j+1; k < std::min(j + 10, graph->n_nodes); ++k) {
-+                        if (view_idx == -1 &&
-+                            graph->nodes[k]->op == GGML_OP_VIEW &&
-+                            graph->nodes[k]->src[0] == graph->nodes[j]) {
-+                            view_idx = k;
-+                            continue;
-+                        }
-+                        if (view_idx != -1 &&
-+                            set_rows_idx == -1 &&
-+                            graph->nodes[k]->op == GGML_OP_SET_ROWS &&
-+                            graph->nodes[k]->src[0] == graph->nodes[view_idx]) {
-+                            set_rows_idx = k;
-+                            break;
-+                        }
-+                    }
-+                    if (set_rows_idx != -1) {
-+                        current_set.push_back(view_idx);
-+                        current_set.push_back(set_rows_idx);
-+                        used[view_idx] = true;
-+                        used[set_rows_idx] = true;
-+                    }
-+                }
-             }
-         }
-         // Second pass grabs view nodes.
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl
-index 50fc1f1e2..0eda186c8 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl
-@@ -10,6 +10,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
- layout (binding = 1) readonly buffer Y {int data_pos[];};
- layout (binding = 2) readonly buffer Z {float data_ff[];};
- layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
-+layout (binding = 4) readonly buffer I {uvec2 data_i[];}; // indices for set_rows
- 
- layout (push_constant) uniform parameter {
-     uint ncols;
-@@ -27,6 +28,7 @@ layout (push_constant) uniform parameter {
-     uint s2;
-     int sections[4];
-     uint is_back;
-+    uint set_rows_stride;
- } p;
- 
- float rope_yarn_ramp(const float low, const float high, const uint i0) {
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
-index 06e095bef..9f4538155 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
-@@ -16,12 +16,19 @@ void main() {
-     const uint row_x     = row_dst % ne1;
-     const uint channel_x = row_dst / ne1;
- 
--    const uint idst = row_dst*ne0 + i0/2;
-+    uint idst = row_dst*ne0 + i0/2;
-     const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
- 
-+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
-+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
-+    if (p.set_rows_stride != 0) {
-+        idst = row_x*ne0 + i0/2;
-+        idst += data_i[channel_x].x * p.set_rows_stride;
-+    }
-+
-     if (i0 >= p.n_dims) {
--        data_d[idst + i0/2 + 0] = data_a[ix + i0/2 + 0];
--        data_d[idst + i0/2 + 1] = data_a[ix + i0/2 + 1];
-+        data_d[idst + i0/2 + 0] = D_TYPE(data_a[ix + i0/2 + 0]);
-+        data_d[idst + i0/2 + 1] = D_TYPE(data_a[ix + i0/2 + 1]);
- 
-         return;
-     }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
-index 6ba957540..f4209ed95 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
-@@ -16,12 +16,19 @@ void main() {
-     const uint row_x     = row_dst % ne1;
-     const uint channel_x = row_dst / ne1;
- 
--    const uint idst = row_dst*ne0 + i0;
-+    uint idst = row_dst*ne0 + i0;
-     const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0;
- 
-+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
-+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
-+    if (p.set_rows_stride != 0) {
-+        idst = row_x*ne0 + i0;
-+        idst += data_i[channel_x].x * p.set_rows_stride;
-+    }
-+
-     if (i0 >= p.n_dims) {
--        data_d[idst + 0] = data_a[ix + 0];
--        data_d[idst + 1] = data_a[ix + 1];
-+        data_d[idst + 0] = D_TYPE(data_a[ix + 0]);
-+        data_d[idst + 1] = D_TYPE(data_a[ix + 1]);
- 
-         return;
-     }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-index 03fa01639..e6ec589fb 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-@@ -842,10 +842,14 @@ void process_shaders() {
-     string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-     string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-     string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
-+    string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
-+    string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
- 
-     string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-     string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-     string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
-+    string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
-+    string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
- 
-     string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-     string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
-index 9eb2b6687..657b6cc2f 100644
---- a/tests/test-backend-ops.cpp
-+++ b/tests/test-backend-ops.cpp
-@@ -2105,6 +2105,34 @@ struct test_get_rows_back : public test_case {
-     }
- };
- 
-+static void init_set_rows_row_ids(ggml_tensor * t, int num_rows) {
-+    std::random_device rd;
-+    std::default_random_engine rng(rd());
-+    for (int i2 = 0; i2 < t->ne[2]; i2++) {
-+        for (int i1 = 0; i1 < t->ne[1]; i1++) {
-+            // generate a shuffled subset of row indices
-+            std::vector<int64_t> data(num_rows);
-+            for (int i = 0; i < num_rows; i++) {
-+                data[i] = i;
-+            }
-+            std::shuffle(data.begin(), data.end(), rng);
-+            data.resize(t->ne[0]);
-+
-+            const size_t offs = i1*t->nb[1] + i2*t->nb[2];
-+            if (t->type == GGML_TYPE_I32) {
-+                // TODO: Make a template or something
-+                std::vector<int32_t> data_i32(t->ne[0]);
-+                for (int i = 0; i < t->ne[0]; i++) {
-+                    data_i32[i] = static_cast<int32_t>(data[i]);
-+                }
-+                ggml_backend_tensor_set(t, data_i32.data(), offs, t->ne[0]*sizeof(int32_t));
-+            } else {
-+                ggml_backend_tensor_set(t, data.data(), offs, t->ne[0]*sizeof(int64_t));
-+            }
-+        }
-+    }
-+}
-+
- // GGML_OP_SET_ROWS
- struct test_set_rows : public test_case {
-     const ggml_type type;
-@@ -2148,37 +2176,13 @@ struct test_set_rows : public test_case {
-     }
- 
-     void initialize_tensors(ggml_context * ctx) override {
--        std::random_device rd;
--        std::default_random_engine rng(rd());
-         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-             if (t->type == GGML_TYPE_I64 || t->type == GGML_TYPE_I32) {
-                 if (ggml_is_view_op(t->op)) {
-                     continue;
-                 }
- 
--                for (int i2 = 0; i2 < t->ne[2]; i2++) {
--                    for (int i1 = 0; i1 < t->ne[1]; i1++) {
--                        // generate a shuffled subset of row indices
--                        std::vector<int64_t> data(ne[1]);
--                        for (int i = 0; i < ne[1]; i++) {
--                            data[i] = i;
--                        }
--                        std::shuffle(data.begin(), data.end(), rng);
--                        data.resize(t->ne[0]);
--
--                        const size_t offs = i1*t->nb[1] + i2*t->nb[2];
--                        if (t->type == GGML_TYPE_I32) {
--                            // TODO: Make a template or something
--                            std::vector<int32_t> data_i32(t->ne[0]);
--                            for (int i = 0; i < t->ne[0]; i++) {
--                                data_i32[i] = static_cast<int32_t>(data[i]);
--                            }
--                            ggml_backend_tensor_set(t, data_i32.data(), offs, t->ne[0]*sizeof(int32_t));
--                        } else {
--                            ggml_backend_tensor_set(t, data.data(), offs, t->ne[0]*sizeof(int64_t));
--                        }
--                    }
--                }
-+                init_set_rows_row_ids(t, ne[1]);
-             } else {
-                 init_tensor_uniform(t);
-             }
-@@ -2207,6 +2211,67 @@ struct test_set_rows : public test_case {
-     }
- };
- 
-+// GGML_OP_ROPE + GGML_OP_VIEW + GGML_OP_SET_ROWS
-+struct test_rope_set_rows : public test_case {
-+    const ggml_type type;
-+    const ggml_type type_idx;
-+    const std::array<int64_t, 4> ne;
-+    int mode;
-+
-+    std::string vars() override {
-+        return VARS_TO_STR4(type, type_idx, ne, mode);
-+    }
-+
-+    std::string op_desc(ggml_tensor * t) override {
-+        GGML_UNUSED(t);
-+        return "ROPE_SET_ROWS";
-+    }
-+
-+    bool run_whole_graph() override { return true; }
-+
-+    test_rope_set_rows(ggml_type type,
-+            ggml_type type_idx,
-+            std::array<int64_t, 4> ne,
-+            int mode)
-+        : type(type), type_idx(type_idx), ne(ne), mode(mode) {}
-+
-+    ggml_tensor * build_graph(ggml_context * ctx) override {
-+        ggml_tensor * src = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], ne[1], ne[2], 1);
-+        ggml_set_name(src, "src");
-+
-+        ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]);
-+
-+        ggml_tensor * rope = ggml_rope(ctx, src, pos, ne[0], mode);
-+
-+        ggml_tensor * view = ggml_view_2d(ctx, rope, ne[0] * ne[1], ne[2], rope->nb[2], 0);
-+
-+        ggml_tensor * dst = ggml_new_tensor_4d(ctx, type, ne[0] * ne[1], ne[2] * ne[3], 1, 1);
-+        ggml_set_name(dst, "dst");
-+
-+        ggml_tensor * row_idxs = ggml_new_tensor_3d(ctx, type_idx, ne[2], 1, 1);
-+        ggml_set_name(row_idxs, "row_idxs");
-+
-+        ggml_tensor * out = ggml_set_rows(ctx, dst, view, row_idxs);
-+        ggml_set_name(out, "out");
-+
-+        return out;
-+    }
-+
-+    void initialize_tensors(ggml_context * ctx) override {
-+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-+            if (t->type == GGML_TYPE_I64 || t->type == GGML_TYPE_I32) {
-+                if (ggml_is_view_op(t->op)) {
-+                    continue;
-+                }
-+
-+                init_set_rows_row_ids(t, ne[2]);
-+            } else {
-+                init_tensor_uniform(t);
-+            }
-+        }
-+    }
-+};
-+
- // GGML_OP_ARGMAX
- struct test_argmax : public test_case {
-     const ggml_type type;
-@@ -6008,6 +6073,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
-         }
-     }
- 
-+    for (int mode : { GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX }) {
-+        for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
-+            test_cases.emplace_back(new test_rope_set_rows(type, GGML_TYPE_I64, { 128, 32, 1, 100 }, mode));
-+            test_cases.emplace_back(new test_rope_set_rows(type, GGML_TYPE_I64, { 128, 32, 512, 1 }, mode));
-+        }
-+    }
-+
-     for (ggml_type type_input : {GGML_TYPE_F32}) {
-         for (ggml_op_pool pool_type : {GGML_OP_POOL_AVG, GGML_OP_POOL_MAX}) {
-             for (int k0 : {1, 3}) {
diff --git a/llama/patches/0033-vulkan-Handle-argsort-with-a-large-number-of-rows-16.patch b/llama/patches/0033-vulkan-Handle-argsort-with-a-large-number-of-rows-16.patch
deleted file mode 100644
index a7048e7f..00000000
--- a/llama/patches/0033-vulkan-Handle-argsort-with-a-large-number-of-rows-16.patch
+++ /dev/null
@@ -1,85 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Jeff Bolz <jbolz@nvidia.com>
-Date: Thu, 30 Oct 2025 01:27:41 -0500
-Subject: [PATCH] vulkan: Handle argsort with a large number of rows (#16851)
-
----
- ggml/src/ggml-vulkan/ggml-vulkan.cpp             |  4 ++++
- ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp | 16 ++++++++++++----
- 2 files changed, 16 insertions(+), 4 deletions(-)
-
-diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index db92a7901..e959674d1 100644
---- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -1084,6 +1084,7 @@ struct vk_op_soft_max_push_constants {
- 
- struct vk_op_argsort_push_constants {
-     uint32_t ncols;
-+    uint32_t nrows;
-     int32_t order;
- };
- 
-@@ -8710,6 +8711,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
-         break;
-     case GGML_OP_ARGSORT:
-         elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 };
-+        elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
-         break;
-     case GGML_OP_IM2COL:
-         {
-@@ -9952,9 +9954,11 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
-     int32_t * op_params = (int32_t *)dst->op_params;
- 
-     uint32_t ncols = src0->ne[0];
-+    uint32_t nrows = ggml_nrows(src0);
- 
-     ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
-         ncols,
-+        nrows,
-         op_params[0],
-     }, dryrun);
- }
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp b/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
-index c81b84452..c4e68bc02 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
-@@ -14,6 +14,7 @@ layout (binding = 1)          buffer D {int data_d[];};
- 
- layout (push_constant) uniform parameter {
-     uint ncols;
-+    uint nrows;
-     uint order;
- } p;
- 
-@@ -26,10 +27,9 @@ void swap(uint idx0, uint idx1) {
-     dst_row[idx1] = tmp;
- }
- 
--void argsort(bool needs_bounds_check) {
-+void argsort(bool needs_bounds_check, const uint row) {
-     // bitonic sort
-     const int col = int(gl_LocalInvocationID.x);
--    const uint row = gl_WorkGroupID.y;
- 
-     const uint row_offset = row * p.ncols;
- 
-@@ -72,8 +72,16 @@ void argsort(bool needs_bounds_check) {
- 
- void main() {
-     if (p.ncols == BLOCK_SIZE) {
--        argsort(false);
-+        uint row = gl_WorkGroupID.y;
-+        while (row < p.nrows) {
-+            argsort(false, row);
-+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
-+        }
-     } else {
--        argsort(true);
-+        uint row = gl_WorkGroupID.y;
-+        while (row < p.nrows) {
-+            argsort(true, row);
-+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
-+        }
-     }
- }
diff --git a/llama/patches/0034-vulkan-fix-shmem-overrun-in-mmq-id-shader-16873.patch b/llama/patches/0034-vulkan-fix-shmem-overrun-in-mmq-id-shader-16873.patch
deleted file mode 100644
index 73dad676..00000000
--- a/llama/patches/0034-vulkan-fix-shmem-overrun-in-mmq-id-shader-16873.patch
+++ /dev/null
@@ -1,77 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Ruben Ortlam <picard12@live.de>
-Date: Fri, 31 Oct 2025 08:14:49 +0100
-Subject: [PATCH] vulkan: fix shmem overrun in mmq id shader (#16873)
-
-* vulkan: fix shmem overrun in mmq id shader
-
-* metal : fix mul_mm_id
-
----------
-
-Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
----
- ggml/src/ggml-metal/ggml-metal-device.cpp                    | 2 +-
- ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp             | 4 ++++
- ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl | 2 +-
- tests/test-backend-ops.cpp                                   | 3 +++
- 4 files changed, 9 insertions(+), 2 deletions(-)
-
-diff --git a/ggml/src/ggml-metal/ggml-metal-device.cpp b/ggml/src/ggml-metal/ggml-metal-device.cpp
-index 758116342..c78082ac3 100644
---- a/ggml/src/ggml-metal/ggml-metal-device.cpp
-+++ b/ggml/src/ggml-metal/ggml-metal-device.cpp
-@@ -677,7 +677,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0(ggml_metal_
-     char name[256];
- 
-     snprintf(base, 256, "kernel_mul_mm_id_map0_ne20_%d", ne20);
--    snprintf(name, 256, "%s", base);
-+    snprintf(name, 256, "%s_ne02=%d", base, ne02);
- 
-     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
-     if (res) {
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
-index 8b238ac4b..d955b4fc7 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
-@@ -82,9 +82,13 @@ layout (constant_id = 10) const uint WARP = 32;
- 
- #include "mul_mmq_shmem_types.glsl"
- 
-+#ifdef MUL_MAT_ID
-+#define BK_STEP 1
-+#else
- #ifndef BK_STEP
- #define BK_STEP 4
- #endif
-+#endif
- 
- // Shared memory cache
- shared block_a_cache buf_a[BM * BK_STEP];
-diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl
-index 72fec4404..1c0f5306f 100644
---- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl
-+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl
-@@ -27,7 +27,7 @@ struct block_a_cache {
- #elif defined(DATA_A_Q8_0)
- #define QUANT_R_MMQ 1
- // AMD likes 4, Intel likes 1 and Nvidia likes 2
--#define BK_STEP 1
-+// #define BK_STEP 1
- struct block_a_cache {
-     int32_t qs[32/4];
-     FLOAT_TYPE dm;
-diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
-index 657b6cc2f..1f8dda383 100644
---- a/tests/test-backend-ops.cpp
-+++ b/tests/test-backend-ops.cpp
-@@ -6722,6 +6722,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
-     test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 1, 1, false, 8, 16, 1));
-     test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, false, 32, 32, 32, 3));
- 
-+    // gpt-oss issue with Vulkan mmq_id
-+    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_MXFP4, GGML_TYPE_F32, 32, 2, false, 2880, 32, 2880));
-+
-     for (ggml_type type_a : base_types) {
-         for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
-             for (int n_mats : {4, 8}) {
diff --git a/llama/patches/0035-vulkan-Fix-crash-when-FP16-mul_mat-accumulation-is-n.patch b/llama/patches/0035-vulkan-Fix-crash-when-FP16-mul_mat-accumulation-is-n.patch
deleted file mode 100644
index d7c4def1..00000000
--- a/llama/patches/0035-vulkan-Fix-crash-when-FP16-mul_mat-accumulation-is-n.patch
+++ /dev/null
@@ -1,80 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Masato Nakasaka <masato.nakasaka@intel.com>
-Date: Fri, 31 Oct 2025 16:18:59 +0900
-Subject: [PATCH] vulkan: Fix crash when FP16 mul_mat accumulation is not
- supported (#16796)
-
-* Experimenting crash fix
-
-* added assert for aborting and fixed comment
-
-* changed to check if a pipeline is empty or not
-
-* Moved function in class definition
-
-* replaced with is_empty
-
-* Modified is_empty to check only unaligned pipelines
----
- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 20 +++++++++++++-------
- 1 file changed, 13 insertions(+), 7 deletions(-)
-
-diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index e959674d1..903050b0b 100644
---- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -146,8 +146,13 @@ static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
- struct vk_matmul_pipeline_struct {
-     vk_pipeline l, m, s;
-     vk_pipeline a_l, a_m, a_s;
-+    // Returns true when all unaligned pipelines are null.
-+    // We only check for unaligned variants since one of the unaligned pipelines must exist
-+    // while aligned pipelines are optional
-+    bool is_empty() const {
-+        return l == nullptr && m == nullptr && s == nullptr;
-+    }
- };
--
- typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
- 
- struct vk_matmul_pipeline2 {
-@@ -5080,7 +5085,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
-     if (src1_type == GGML_TYPE_Q8_1) {
-         vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f32acc;
- 
--        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
-+        if (pipelines->is_empty()) {
-             return nullptr;
-         }
- 
-@@ -5229,7 +5234,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
-     if (src1_type == GGML_TYPE_Q8_1) {
-         vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_id_q8_1[src0_type].f32acc;
- 
--        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
-+        if (pipelines->is_empty()) {
-             return nullptr;
-         }
- 
-@@ -5264,16 +5269,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
-             return nullptr;
-     }
- 
-+    vk_matmul_pipeline2& mmp = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type];
-     // XXX TODO 'prec' is not actually allowed in mul_mat_id.
-     bool prefer_fp16acc = ctx->device->fp16 /*&& prec == GGML_PREC_DEFAULT*/;
--    bool support_fp16acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc != nullptr;
--    bool support_fp32acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc != nullptr;
-+    bool support_fp16acc = !mmp.f16acc->is_empty();
-+    bool support_fp32acc = !mmp.f32acc->is_empty();
- 
-     if (support_fp16acc && (prefer_fp16acc || !support_fp32acc)) {
--        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc;
-+        return mmp.f16acc;
-     } else {
-         GGML_ASSERT(support_fp32acc);
--        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
-+        return mmp.f32acc;
-     }
- }
- 
diff --git a/ml/backend/ggml/ggml.go b/ml/backend/ggml/ggml.go
index 520d95cb..39457939 100644
--- a/ml/backend/ggml/ggml.go
+++ b/ml/backend/ggml/ggml.go
@@ -1702,7 +1702,7 @@ func (t *Tensor) Duplicate(ctx ml.Context) ml.Tensor {
 func (t *Tensor) TopK(ctx ml.Context, k int) ml.Tensor {
 	return &Tensor{
 		b: t.b,
-		t: C.ggml_top_k(ctx.(*Context).ctx, t.t, C.int(k)),
+		t: C.ggml_argsort_top_k(ctx.(*Context).ctx, t.t, C.int(k)),
 	}
 }
 
diff --git a/ml/backend/ggml/ggml/include/ggml-rpc.h b/ml/backend/ggml/ggml/include/ggml-rpc.h
index e6dca3f6..832c26c6 100644
--- a/ml/backend/ggml/ggml/include/ggml-rpc.h
+++ b/ml/backend/ggml/ggml/include/ggml-rpc.h
@@ -8,7 +8,7 @@ extern "C" {
 #endif
 
 #define RPC_PROTO_MAJOR_VERSION    3
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MINOR_VERSION    5
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 
diff --git a/ml/backend/ggml/ggml/include/ggml.h b/ml/backend/ggml/ggml/include/ggml.h
index d948b00c..4dbca868 100644
--- a/ml/backend/ggml/ggml/include/ggml.h
+++ b/ml/backend/ggml/ggml/include/ggml.h
@@ -242,6 +242,7 @@
 #define GGML_ROPE_TYPE_NEOX   2
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
+#define GGML_ROPE_TYPE_IMROPE 40 // binary: 101000
 
 #define GGML_MROPE_SECTIONS   4
 
@@ -474,6 +475,7 @@ extern "C" {
         GGML_OP_COS,
         GGML_OP_SUM,
         GGML_OP_SUM_ROWS,
+        GGML_OP_CUMSUM,
         GGML_OP_MEAN,
         GGML_OP_ARGMAX,
         GGML_OP_COUNT_EQUAL,
@@ -528,7 +530,10 @@ extern "C" {
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
+        GGML_OP_TOP_K,
         GGML_OP_LEAKY_RELU,
+        GGML_OP_TRI,
+        GGML_OP_FILL,
 
         GGML_OP_FLASH_ATTN_EXT,
         GGML_OP_FLASH_ATTN_BACK,
@@ -541,6 +546,7 @@ extern "C" {
         GGML_OP_RWKV_WKV6,
         GGML_OP_GATED_LINEAR_ATTN,
         GGML_OP_RWKV_WKV7,
+        GGML_OP_SOLVE_TRI,
 
         GGML_OP_UNARY,
 
@@ -575,6 +581,8 @@ extern "C" {
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
         GGML_UNARY_OP_EXP,
+        GGML_UNARY_OP_EXPM1,
+        GGML_UNARY_OP_SOFTPLUS,
         GGML_UNARY_OP_GELU_ERF,
         GGML_UNARY_OP_XIELU,
         GGML_UNARY_OP_FLOOR,
@@ -619,6 +627,13 @@ extern "C" {
         GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
+    enum ggml_tri_type {
+        GGML_TRI_TYPE_UPPER_DIAG = 0,
+        GGML_TRI_TYPE_UPPER      = 1,
+        GGML_TRI_TYPE_LOWER_DIAG = 2,
+        GGML_TRI_TYPE_LOWER      = 3
+    };
+
     struct ggml_init_params {
         // memory pool
         size_t mem_size;   // bytes
@@ -956,6 +971,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_expm1(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_expm1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_softplus(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_softplus_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_sin(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -982,6 +1013,10 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_cumsum(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
     // mean along rows
     GGML_API struct ggml_tensor * ggml_mean(
             struct ggml_context * ctx,
@@ -2107,6 +2142,7 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
+        GGML_SCALE_MODE_BICUBIC  = 2,
 
         GGML_SCALE_MODE_COUNT
     };
@@ -2185,6 +2221,23 @@ extern "C" {
             int                   shift2,
             int                   shift3);
 
+    // Convert matrix into a triangular one (upper, strict upper, lower or strict lower) by writing
+    // zeroes everywhere outside the masked area
+    GGML_API struct ggml_tensor * ggml_tri(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_tri_type    type);
+
+    // Fill tensor a with constant c
+    GGML_API struct ggml_tensor * ggml_fill(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 c);
+
+    GGML_API struct ggml_tensor * ggml_fill_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 c);
 
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
@@ -2206,18 +2259,25 @@ extern "C" {
             struct ggml_tensor  * a,
             enum ggml_sort_order  order);
 
+    // similar to ggml_top_k but implemented as `argsort` + `view`
+    GGML_API struct ggml_tensor * ggml_argsort_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
+    // top k elements per row
+    // note: the resulting top k indices are in no particular order
+    GGML_API struct ggml_tensor * ggml_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
     GGML_API struct ggml_tensor * ggml_arange(
             struct ggml_context * ctx,
             float                 start,
             float                 stop,
             float                 step);
 
-    // top k elements per row
-    GGML_API struct ggml_tensor * ggml_top_k(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   k);
-
 #define GGML_KQ_MASK_PAD 64
 
     // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
@@ -2354,6 +2414,27 @@ extern "C" {
             struct ggml_tensor  * b,
             struct ggml_tensor  * state);
 
+    /* Solves a specific equation of the form Ax=B, where A is a triangular matrix
+    *  without zeroes on the diagonal (i.e. invertible).
+    *  B can have any number of columns, but must have the same number of rows as A
+    *  If A is [n, n] and B is [n, m], then the result will be [n, m] as well
+    *  Has O(n^3) complexity (unlike most matrix ops out there), so use on cases
+    *  where n > 100 sparingly, pre-chunk if necessary.
+    *
+    *  If left = false, solves xA=B instead
+    *  If lower = false, assumes upper triangular instead
+    *  If uni = true, assumes diagonal of A to be all ones (will override actual values)
+    *
+    *  TODO: currently only lower, right, non-unitriangular variant is implemented
+    */
+    GGML_API struct ggml_tensor * ggml_solve_tri(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  left,
+        bool                  lower,
+        bool                  uni);
+
     // custom operators
 
     typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
diff --git a/ml/backend/ggml/ggml/src/CMakeLists.txt b/ml/backend/ggml/ggml/src/CMakeLists.txt
index 4b3e5efb..faa1beed 100644
--- a/ml/backend/ggml/ggml/src/CMakeLists.txt
+++ b/ml/backend/ggml/ggml/src/CMakeLists.txt
@@ -214,15 +214,29 @@ add_library(ggml-base
             mem_dxgi_pdh.cpp
             gguf.cpp)
 
+set_target_properties(ggml-base PROPERTIES
+    VERSION ${GGML_VERSION}
+    SOVERSION ${GGML_VERSION_MAJOR}
+)
+
 target_include_directories(ggml-base PRIVATE .)
 if (GGML_BACKEND_DL)
     target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL)
 endif()
 
+if (GGML_SCHED_NO_REALLOC)
+    target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC)
+endif()
+
 add_library(ggml
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
 
+set_target_properties(ggml PROPERTIES
+    VERSION ${GGML_VERSION}
+    SOVERSION ${GGML_VERSION_MAJOR}
+)
+
 if (GGML_BACKEND_DIR)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
@@ -262,6 +276,15 @@ function(ggml_add_backend_library backend)
         target_compile_definitions(${backend} PUBLIC  GGML_BACKEND_SHARED)
     endif()
 
+    # Set versioning properties for all backend libraries
+    # Building a MODULE library with a version is not supported on macOS (https://gitlab.kitware.com/cmake/cmake/-/issues/20782)
+    if (NOT (APPLE AND GGML_BACKEND_DL))
+        set_target_properties(${backend} PROPERTIES
+            VERSION ${GGML_VERSION}
+            SOVERSION ${GGML_VERSION_MAJOR}
+        )
+    endif()
+
     if(NOT GGML_AVAILABLE_BACKENDS)
         set(GGML_AVAILABLE_BACKENDS "${backend}"
             CACHE INTERNAL "List of backends for cmake package")
@@ -311,6 +334,18 @@ function(ggml_add_cpu_backend_variant tag_name)
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
+        foreach (feat VXE2 NNPA)
+            set(GGML_INTERNAL_${feat} OFF)
+        endforeach()
+
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        foreach (feat RVV)
+            set(GGML_INTERNAL_${feat} OFF)
+        endforeach()
+
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
@@ -378,12 +413,18 @@ if (GGML_CPU_ALL_VARIANTS)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            ggml_add_cpu_backend_variant(s390x_z15  Z15 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z16  Z16 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z17  Z17 VXE)
+            ggml_add_cpu_backend_variant(z15    Z15 VXE2)
+            ggml_add_cpu_backend_variant(z16    Z16 VXE2 NNPA)
         else()
             message(FATAL_ERROR "Unsupported s390x target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(riscv64_0)
+            ggml_add_cpu_backend_variant(riscv64_v   RVV)
+        else()
+            message(FATAL_ERROR "Unsupported RISC-V target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
     else()
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()
diff --git a/ml/backend/ggml/ggml/src/ggml-alloc.c b/ml/backend/ggml/ggml/src/ggml-alloc.c
index 36385387..06ee502a 100644
--- a/ml/backend/ggml/ggml/src/ggml-alloc.c
+++ b/ml/backend/ggml/ggml/src/ggml-alloc.c
@@ -226,16 +226,23 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
     }
 
     if (best_fit_block == -1) {
-        // no suitable block found, try the last block (this will grow a chunks size)
+        // no suitable block found, try the last block (this may grow a chunks size)
+        int64_t best_reuse = INT64_MIN;
         for (int c = 0; c < alloc->n_chunks; ++c) {
             struct tallocr_chunk * chunk = alloc->chunks[c];
             if (chunk->n_free_blocks > 0) {
                 struct free_block * block = &chunk->free_blocks[chunk->n_free_blocks - 1];
                 max_avail = MAX(max_avail, block->size);
-                if (block->size >= size) {
+                int64_t reuse_factor = chunk->max_size - block->offset - size;
+                // reuse_factor < 0 : amount of extra memory that needs to be allocated
+                // reuse_factor = 0 : allocated free space exactly matches tensor size
+                // reuse_factor > 0 : superfluous memory that will remain unused
+                bool better_reuse = best_reuse < 0 && reuse_factor > best_reuse;
+                bool better_fit = reuse_factor >= 0 && reuse_factor < best_reuse;
+                if (block->size >= size && (better_reuse || better_fit)) {
                     best_fit_chunk = c;
                     best_fit_block = chunk->n_free_blocks - 1;
-                    break;
+                    best_reuse = reuse_factor;
                 }
             }
         }
@@ -268,7 +275,7 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, addr, tensor);
     size_t cur_max = addr.offset + size;
-    if (cur_max > alloc->max_size[addr.chunk]) {
+    if (cur_max > chunk->max_size) {
         // sort allocated_tensors by chunk/offset
         for (int i = 0; i < 1024; i++) {
             for (int j = i + 1; j < 1024; j++) {
@@ -921,10 +928,15 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         }
         if (realloc) {
 #ifndef NDEBUG
-            size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
-            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+            {
+                size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
+                if (cur_size > 0) {
+                    GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n",
+                        __func__, ggml_backend_buft_name(galloc->bufts[i]),
+                        cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+                }
+            }
 #endif
-
             ggml_vbuffer_free(galloc->buffers[i]);
             galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
             if (galloc->buffers[i]) {
diff --git a/ml/backend/ggml/ggml/src/ggml-backend.cpp b/ml/backend/ggml/ggml/src/ggml-backend.cpp
index 9b0a9b91..8d2cc167 100644
--- a/ml/backend/ggml/ggml/src/ggml-backend.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-backend.cpp
@@ -1431,14 +1431,20 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 
     // allocate graph
     if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+#ifdef GGML_SCHED_NO_REALLOC
+        GGML_ABORT("%s: failed to allocate graph, but graph re-allocation is disabled by GGML_SCHED_NO_REALLOC\n", __func__);
+#endif
+
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+#endif
+
         // the re-allocation may cause the split inputs to be moved to a different address
         // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy
         for (int i = 0; i < sched->n_backends; i++) {
             ggml_backend_synchronize(sched->backends[i]);
         }
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
-#endif
+
         ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
         if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
             GGML_LOG_ERROR("%s: failed to allocate graph\n", __func__);
@@ -1757,8 +1763,6 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
     GGML_ASSERT(sched);
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
-    ggml_backend_sched_reset(sched);
-
     ggml_backend_sched_synchronize(sched);
 
     ggml_backend_sched_split_graph(sched, measure_graph);
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/CMakeLists.txt b/ml/backend/ggml/ggml/src/ggml-cpu/CMakeLists.txt
index 34323afa..7e53a57b 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/CMakeLists.txt
@@ -126,36 +126,48 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 )
                 if (NOT ARM_MCPU_RESULT)
                     string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
+                    string(REGEX MATCH "-march=[^ ']+" ARM_MARCH_FLAG "${ARM_MCPU}")
+
+                    # on some old GCC we need to read -march=
+                    if (ARM_MARCH_FLAG AND NOT "${ARM_MARCH_FLAG}" STREQUAL "-march=native")
+                        set(ARM_NATIVE_FLAG "${ARM_MARCH_FLAG}")
+                    elseif(ARM_MCPU_FLAG AND NOT "${ARM_MCPU_FLAG}" STREQUAL "-mcpu=native")
+                        set(ARM_NATIVE_FLAG "${ARM_MCPU_FLAG}")
+                    endif()
                 endif()
-                if ("${ARM_MCPU_FLAG}" STREQUAL "")
-                    set(ARM_MCPU_FLAG -mcpu=native)
-                    message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
+
+                if ("${ARM_NATIVE_FLAG}" STREQUAL "")
+                    set(ARM_NATIVE_FLAG -mcpu=native)
+                    message(WARNING "ARM -march/-mcpu not found, -mcpu=native will be used")
+                else()
+                    message(STATUS "ARM detected flags: ${ARM_NATIVE_FLAG}")
                 endif()
 
                 include(CheckCXXSourceRuns)
 
-                function(check_arm_feature tag code)
+                macro(check_arm_feature tag feature code)
                     set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-                    set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
+                    set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}")
                     check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
                     if (GGML_MACHINE_SUPPORTS_${tag})
-                        set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
+                        set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}")
                     else()
-                        set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
+                        set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+no${tag}")
                         check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
                         if (GGML_MACHINE_SUPPORTS_no${tag})
-                            set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
+                            set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}")
+                            list(APPEND ARCH_FLAGS -U__ARM_FEATURE_${feature})
                         endif()
                     endif()
                     set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
-                endfunction()
+                endmacro()
 
-                check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
-                check_arm_feature(i8mm    "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
-                check_arm_feature(sve     "#include <arm_sve.h>\nint main()  { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
-                check_arm_feature(sme     "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
+                check_arm_feature(dotprod DOTPROD     "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
+                check_arm_feature(i8mm    MATMUL_INT8 "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
+                check_arm_feature(sve     SVE         "#include <arm_sve.h>\nint main()  { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
+                check_arm_feature(sme     SME         "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
 
-                list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
+                list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}")
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
@@ -205,35 +217,28 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 endif()
             endif()
 
-            # show enabled features
-            if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
-                set(FEAT_INPUT_FILE "NUL")
-            else()
-                set(FEAT_INPUT_FILE "/dev/null")
-            endif()
+            message(STATUS "Checking for ARM features using flags:")
+            foreach(flag IN LISTS ARCH_FLAGS)
+                message(STATUS "  ${flag}")
+            endforeach()
 
-            execute_process(
-                COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
-                INPUT_FILE ${FEAT_INPUT_FILE}
-                OUTPUT_VARIABLE ARM_FEATURE
-                RESULT_VARIABLE ARM_FEATURE_RESULT
-            )
-            if (ARM_FEATURE_RESULT)
-                message(WARNING "Failed to get ARM features")
-            else()
-                foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
-                    string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
-                    if (NOT ${feature_pos} EQUAL -1)
-                        # Special handling for MATMUL_INT8 when machine doesn't support i8mm
-                        if ("${feature}" STREQUAL "MATMUL_INT8" AND GGML_MACHINE_SUPPORTS_noi8mm)
-                            message(STATUS "ARM feature ${feature} detected but unsetting due to machine not supporting i8mm")
-                            list(APPEND ARCH_FLAGS -U__ARM_FEATURE_MATMUL_INT8)
-                        else()
-                            message(STATUS "ARM feature ${feature} enabled")
-                        endif()
-                    endif()
-                endforeach()
-            endif()
+            include(CheckCXXSourceCompiles)
+            set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+            string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}")
+            set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}")
+            foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
+                set(ARM_FEATURE "HAVE_${feature}")
+                check_cxx_source_compiles(
+                    "
+                    #if !defined(__ARM_FEATURE_${feature})
+                    #  error \"Feature ${feature} is not defined\"
+                    #endif
+                    int main() { return 0; }
+                    "
+                    ${ARM_FEATURE}
+                )
+            endforeach()
+            set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
         message(STATUS "x86 detected")
@@ -388,9 +393,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)
-                list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
+                list(APPEND ARCH_FLAGS -mcpu=power10)
             elseif (EXTRACTED_NUMBER EQUAL 9)
-                list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
+                list(APPEND ARCH_FLAGS -mcpu=power9)
             elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
                 list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
             else()
@@ -448,22 +453,35 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ggml-cpu/spacemit/ime_kernels.h
             )
         endif()
-        set(MARCH_STR "rv64gc")
-        if (GGML_RV_ZFH)
-            string(APPEND MARCH_STR "_zfh")
-        endif()
-        if (GGML_XTHEADVECTOR)
-            string(APPEND MARCH_STR "_xtheadvector")
-        elseif (GGML_RVV)
-            string(APPEND MARCH_STR "_v")
-            if (GGML_RV_ZVFH)
-                string(APPEND MARCH_STR "_zvfh")
+        if(NOT GGML_CPU_ALL_VARIANTS)
+            set(MARCH_STR "rv64gc")
+            if (GGML_RV_ZFH)
+                string(APPEND MARCH_STR "_zfh")
             endif()
+            if (GGML_XTHEADVECTOR)
+                string(APPEND MARCH_STR "_xtheadvector")
+            elseif (GGML_RVV)
+                string(APPEND MARCH_STR "_v")
+                if (GGML_RV_ZVFH)
+                    string(APPEND MARCH_STR "_zvfh")
+                endif()
+            endif()
+            if (GGML_RV_ZICBOP)
+                string(APPEND MARCH_STR "_zicbop")
+            endif()
+            list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
+        else()
+            # Begin with the lowest baseline
+            set(ARCH_DEFINITIONS "")
+
+            if (GGML_INTERNAL_RVV)
+                message(STATUS "RVV enabled")
+                list(APPEND ARCH_DEFINITIONS GGML_USE_RVV)
+                list(APPEND ARCH_FLAGS -march=rv64gc_v -mabi=lp64d)
+            endif()
+
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} riscv ${ARCH_DEFINITIONS})
         endif()
-        if (GGML_RV_ZICBOP)
-            string(APPEND MARCH_STR "_zicbop")
-        endif()
-        list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
         list(APPEND GGML_CPU_SOURCES
@@ -504,11 +522,18 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             endforeach()
         endif()
 
-        if (GGML_VXE OR GGML_INTERNAL_VXE)
-            message(STATUS "VX/VXE/VXE2 enabled")
+        if (GGML_VXE OR GGML_INTERNAL_VXE2)
+            message(STATUS "VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
-            list(APPEND ARCH_DEFINITIONS GGML_VXE)
+            list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2)
         endif()
+
+        if (GGML_INTERNAL_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA)
+        endif()
+
+        ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS})
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
         list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
@@ -572,6 +597,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${KLEIDIAI_SRC}/kai/ukernels/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
 
@@ -590,23 +616,34 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
-            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)
 
         if (NOT DOTPROD_ENABLED MATCHES -1)
             list(APPEND GGML_KLEIDIAI_SOURCES
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
-                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
         endif()
 
         if (NOT I8MM_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
         endif()
 
         if (NOT SME_ENABLED MATCHES -1)
             list(APPEND GGML_KLEIDIAI_SOURCES
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/arch-fallback.h b/ml/backend/ggml/ggml/src/ggml-cpu/arch-fallback.h
index edfd7913..0775c87f 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/arch-fallback.h
@@ -33,10 +33,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -44,27 +46,30 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__POWERPC__) || defined(__powerpc__)
 // ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
@@ -76,10 +81,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -87,6 +94,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -101,10 +109,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -112,6 +122,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -134,15 +145,18 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -163,10 +177,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -174,6 +190,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
@@ -196,10 +213,12 @@
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
@@ -207,6 +226,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/arm.go b/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/arm.go
index 581801c0..bf21fad0 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/arm.go
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/arm.go
@@ -1,3 +1,5 @@
+//go:build arm64
+
 package arm
 
 // #cgo CXXFLAGS: -std=c++17
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/quants.c b/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/quants.c
index aadbb487..b390ab61 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/quants.c
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/quants.c
@@ -2044,6 +2044,26 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 }
 
+#ifdef __ARM_FEATURE_SVE
+static inline svuint32_t ggml_decode_q4scales_and_mins_for_mmla(const uint32_t * vx_scales) {
+    const svbool_t pg_all   = svptrue_pat_b32(SV_VL4);
+    const svbool_t pg_false = svpfalse_b();            // 0x0000
+    const svbool_t pg_lo_8  = svwhilelt_b8_s32(0,  8); // 0x00ff
+    const svbool_t pg_odd   = svzip1_b32(pg_false, pg_lo_8);
+
+    svuint32_t vutmp_hi, vutmp_lo;
+    svuint32_t vx01 = svld1_u32(pg_lo_8, vx_scales);
+    vutmp_hi = svzip1_u32(vx01, vx01);
+    vutmp_hi = svlsr_n_u32_m(pg_odd, vutmp_hi, 2);
+    vutmp_hi = svreinterpret_u32_u64(svand_n_u64_x(pg_all, svreinterpret_u64_u32(vutmp_hi), UINT64_C(0x303030303f3f3f3f)));
+    const svuint32_t vx2 = svdup_u32(vx_scales[2]);
+    vutmp_lo = svlsr_u32_x(pg_all, vx2, svreinterpret_u32_s32(svindex_s32(-2, 2)));
+    vutmp_lo = svand_n_u32_z(pg_odd, vutmp_lo, UINT32_C(0x0f0f0f0f));
+    svuint32_t vutmp = svorr_u32_z(pg_all, vutmp_hi, vutmp_lo);
+    return vutmp;
+}
+#endif
+
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
 #ifdef __ARM_FEATURE_MATMUL_INT8
@@ -2066,8 +2086,220 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     static const uint32_t kmask3 = 0x03030303;
 
     uint32_t utmp[4];
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+#endif
 
-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        svbool_t pg32_2 = svptrue_pat_b32(SV_VL2);
+
+        const block_q4_K * GGML_RESTRICT vx0 = vx;
+        const block_q8_K * GGML_RESTRICT vy0 = vy;
+        const block_q4_K * GGML_RESTRICT vx1 = (const block_q4_K *) ((const uint8_t*)vx + bx);
+        const block_q8_K * GGML_RESTRICT vy1 = (const block_q8_K *) ((const uint8_t*)vy + by);
+
+        union {
+            uint32_t u32[8];
+            uint64_t u64[4];
+        } new_utmp;
+
+        svfloat32_t sumf1 = svdup_n_f32(0);
+
+        switch (vector_length) {
+            case 128:
+                {
+                    svbool_t pg_false = svpfalse_b();
+                    svbool_t pg_lo_8  = svwhilelt_b8_s32(0,  8);
+                    svbool_t vmins_mask1= svzip1_b32(pg_lo_8, pg_false);
+                    svbool_t vmins_mask2 = svzip1_b32(pg_false, pg_lo_8);
+                    svbool_t pg128_all  = svptrue_pat_b8(SV_VL16);
+                    for (int i = 0; i < nb; ++i) {
+                        svfloat32_t vy_d = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg128_all, vy_d, vx_d);
+                        svfloat32_t vx_dmins = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].dmin)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].dmin)));
+                        svfloat32_t vy_dmins = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t svdmins = svmul_n_f32_x(pg128_all, svmul_f32_x(pg128_all, vy_dmins, vx_dmins), -1);
+                        const uint8_t * GGML_RESTRICT q4_0 = vx0[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_0 = vy0[i].qs;
+                        const uint8_t * GGML_RESTRICT q4_1 = vx1[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_1 = vy1[i].qs;
+                        svint16_t lo = svld1_s16(pg128_all, vy0[i].bsums + 0);
+                        svint16_t hi = svld1_s16(pg128_all, vy0[i].bsums + 8);
+                        svint16_t sum_tmp1 = svuzp1_s16(lo, hi);
+                        svint16_t sum_tmp2 = svuzp2_s16(lo, hi);
+                        svint16_t svq8sums_0 = svadd_s16_x(pg128_all, sum_tmp1, sum_tmp2);
+                        lo = svld1_s16(pg128_all, vy1[i].bsums + 0);
+                        hi = svld1_s16(pg128_all, vy1[i].bsums + 8);
+                        sum_tmp1 = svuzp1(lo, hi);
+                        sum_tmp2 = svuzp2(lo, hi);
+                        svint16_t svq8sums_1 = svadd_s16_x(pg128_all, sum_tmp1, sum_tmp2);
+                        svuint32_t decoded_scales0 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx0[i].scales);
+                        svuint32_t decoded_scales1 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx1[i].scales);
+                        svuint32x2_t decoded_scales = svcreate2_u32(decoded_scales0, decoded_scales1);
+                        svst2_u32(pg128_all, new_utmp.u32, decoded_scales);
+                        svint16_t svmins8_0 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u32(svuzp1_u32(svld1_u32(vmins_mask1, new_utmp.u32+4), svdup_n_u32(0)))));
+                        svint16_t svmins8_1 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u32(svuzp2_u32(svld1_u32(vmins_mask2, new_utmp.u32+4), svdup_n_u32(0)))));
+                        svint32_t svsumfs_tmp1 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_0, svmins8_0));
+                        svint32_t svsumfs_tmp2 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_0, svmins8_1));
+                        svint32_t svsumfs_tmp3 = svtrn1_s32(svsumfs_tmp1, svsumfs_tmp2);
+                        svint32_t svsumfs_tmp4 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_1, svmins8_0));
+                        svint32_t svsumfs_tmp5 = svreinterpret_s32_s64(svdot_s64(svdup_n_s64(0), svq8sums_1, svmins8_1));
+                        svint32_t svsumfs_tmp6 = svtrn1_s32(svsumfs_tmp4, svsumfs_tmp5);
+                        svint32_t svsumfs_tmp7 = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(svsumfs_tmp3), svreinterpret_s64_s32(svsumfs_tmp6)));
+                        svint32_t svsumfs_tmp8 = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(svsumfs_tmp3), svreinterpret_s64_s32(svsumfs_tmp6)));
+                        svint32_t svsumfs_tmp = svadd_s32_x(pg128_all, svsumfs_tmp7, svsumfs_tmp8);
+                        svint32_t svscales, sumi1, sumi2;
+                        svint32_t acc_sumif1 = svdup_n_s32(0);
+                        svint32_t acc_sumif2 = svdup_n_s32(0);
+                        svint8_t q4bytes_0_l, q4bytes_0_h, q4bytes_1_l, q4bytes_1_h, l0, l1, l2, l3,
+                                 q8bytes_0_h, q8bytes_0_l, q8bytes_1_h, q8bytes_1_l, r0, r1, r2, r3;
+#pragma GCC unroll 1
+                        for (int j = 0; j < QK_K/64; ++j) {
+                            q4bytes_0_l = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0), 0xf));
+                            q4bytes_1_l = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1), 0xf));
+                            q4bytes_0_h = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0+16), 0xf));
+                            q4bytes_1_h = svreinterpret_s8_u8(svand_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1+16), 0xf));
+                            l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            l3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            q8bytes_0_h = svld1_s8(pg128_all, q8_0);
+                            q8bytes_1_h = svld1_s8(pg128_all, q8_1);
+                            q8bytes_0_l = svld1_s8(pg128_all, q8_0+16);
+                            q8bytes_1_l = svld1_s8(pg128_all, q8_1+16);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            sumi1 = svmmla_s32(svmmla_s32(svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg128_all, svlsl_n_u32_x(pg128_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-1)), 24));
+                            acc_sumif1 = svmla_s32_x(pg128_all, acc_sumif1, svscales, sumi1);
+
+                            q4bytes_0_l = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0), 4));
+                            q4bytes_1_l = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1), 4));
+                            q4bytes_0_h = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_0+16), 4));
+                            q4bytes_1_h = svreinterpret_s8_u8(svlsr_n_u8_x(pg128_all, svld1_u8(pg128_all, q4_1+16), 4));
+                            l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_l), svreinterpret_s64_s8(q4bytes_1_l)));
+                            l2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            l3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q4bytes_0_h), svreinterpret_s64_s8(q4bytes_1_h)));
+                            q8bytes_0_h = svld1_s8(pg128_all, q8_0+32);
+                            q8bytes_1_h = svld1_s8(pg128_all, q8_1+32);
+                            q8bytes_0_l = svld1_s8(pg128_all, q8_0+48);
+                            q8bytes_1_l = svld1_s8(pg128_all, q8_1+48);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_h), svreinterpret_s64_s8(q8bytes_1_h)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0_l), svreinterpret_s64_s8(q8bytes_1_l)));
+                            sumi2 = svmmla_s32(svmmla_s32(svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg128_all, svlsl_n_u32_x(pg128_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-2)), 24));
+                            acc_sumif2 = svmla_s32_x(pg128_all, acc_sumif2, svscales, sumi2);
+                            q4_0 += 32; q4_1 += 32; q8_0 += 64; q8_1 += 64;
+                        }
+                        sumf1 = svmla_f32_x(pg128_all,
+                                svmla_f32_x(pg128_all,
+                                    sumf1,
+                                    svcvt_f32_x(pg128_all,
+                                        svadd_s32_x(pg128_all, acc_sumif1, acc_sumif2)),
+                                    svsuper_block_scales),
+                                svdmins,
+                                svcvt_f32_s32_x(pg128_all, svsumfs_tmp));
+                    }  //end of for nb
+                } // end of case 128
+                break;
+            case 256:
+            case 512:
+                {
+                    const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4);
+                    const svbool_t pg8_16 = svptrue_pat_b8(SV_VL16);
+                    const svbool_t pg256_all = svptrue_pat_b8(SV_ALL);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT q4_0 = vx0[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_0 = vy0[i].qs;
+                        const uint8_t * GGML_RESTRICT q4_1 = vx1[i].qs;
+                        const int8_t  * GGML_RESTRICT q8_1 = vy1[i].qs;
+                        svint32_t svscales, sumi1, sumi2;
+                        svint32_t acc_sumif1 = svdup_n_s32(0);
+                        svint32_t acc_sumif2 = svdup_n_s32(0);
+                        svint8_t l0, l1, l2, l3, r0, r1, r2, r3;
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat64_t vy_d_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_d = svreinterpret_f32_f64(svuzp1_f64(vy_d_tmp, vy_d_tmp));
+                        svfloat32_t svsuper_block_scales = svmul_f32_z(pg32_4, vy_d, vx_d);
+                        svfloat32_t vx_dmins = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].dmin)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].dmin)));
+                        svfloat64_t vy_dmins_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_dmins = svreinterpret_f32_f64(svuzp1_f64(vy_dmins_tmp, vy_dmins_tmp));
+                        svfloat32_t svdmins = svmul_n_f32_x(pg32_4, svmul_f32_x(pg32_4, vx_dmins, vy_dmins), -1);
+                        svint16_t rc1 = svuzp1_s16(svld1_s16(pg256_all, vy0[i].bsums), svld1_s16(pg256_all, vy1[i].bsums));
+                        svint16_t rc2 = svuzp2_s16(svld1_s16(pg256_all, vy0[i].bsums), svld1_s16(pg256_all, vy1[i].bsums));
+                        svint16_t svq8sums = svadd_s16_x(pg256_all, rc1, rc2);
+                        svuint32_t decoded_scales0 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx0[i].scales);
+                        svuint32_t decoded_scales1 = ggml_decode_q4scales_and_mins_for_mmla((const uint32_t *)vx1[i].scales);
+                        svuint32x2_t decoded_scales = svcreate2_u32(decoded_scales0, decoded_scales1);
+                        svst2_u32(pg8_16, new_utmp.u32, decoded_scales);
+                        svint16_t new_svq8sums_0 = svreinterpret_s16_u64(svtrn1_u64(svreinterpret_u64_s16(svq8sums), svreinterpret_u64_s16(svq8sums)));
+                        svint16_t new_svq8sums_1 = svreinterpret_s16_u64(svtrn2_u64(svreinterpret_u64_s16(svq8sums), svreinterpret_u64_s16(svq8sums)));
+                        svuint64_t new_mins_0 = svdup_u64(new_utmp.u64[2]);
+                        svuint64_t new_mins_1 = svdup_u64(new_utmp.u64[3]);
+                        svint16_t new_svmins8_0 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u64(new_mins_0)));
+                        svint16_t new_svmins8_1 = svreinterpret_s16_u16(svunpklo_u16(svreinterpret_u8_u64(new_mins_1)));
+                        svint64_t dot_prod_0 = svdot_s64(svdup_s64(0), new_svmins8_0, new_svq8sums_0);
+                        svint64_t dot_prod_1 = svdot_s64(dot_prod_0, new_svmins8_1, new_svq8sums_1);
+                        svfloat32_t converted_dot_prod_1 = svcvt_f32_s64_x(pg256_all, dot_prod_1);
+                        svfloat32_t svsumfs_tmp = svuzp1_f32(converted_dot_prod_1, converted_dot_prod_1);
+
+#pragma GCC unroll 1
+                        for (int j = 0; j < QK_K/64; ++j) {
+                            svuint8_t q4bytes_0 = svand_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_0), 0xf);
+                            svuint8_t q4bytes_1 = svand_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_1), 0xf);
+                            svuint8_t q4bytes_2 = svlsr_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_0), 4);
+                            svuint8_t q4bytes_3 = svlsr_n_u8_x(pg256_all, svld1_u8(pg256_all, q4_1), 4);
+                            l0 = svreinterpret_s8_u64(svzip1_u64(svreinterpret_u64_u8(q4bytes_0), svreinterpret_u64_u8(q4bytes_1)));
+                            l1 = svreinterpret_s8_u64(svzip2_u64(svreinterpret_u64_u8(q4bytes_0), svreinterpret_u64_u8(q4bytes_1)));
+                            l2 = svreinterpret_s8_u64(svzip1_u64(svreinterpret_u64_u8(q4bytes_2), svreinterpret_u64_u8(q4bytes_3)));
+                            l3 = svreinterpret_s8_u64(svzip2_u64(svreinterpret_u64_u8(q4bytes_2), svreinterpret_u64_u8(q4bytes_3)));
+                            svint8_t q8bytes_0 = svld1_s8(pg256_all, q8_0);
+                            svint8_t q8bytes_1 = svld1_s8(pg256_all, q8_1);
+                            svint8_t q8bytes_2 = svld1_s8(pg256_all, q8_0+32);
+                            svint8_t q8bytes_3 = svld1_s8(pg256_all, q8_1+32);
+                            r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                            r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                            r2 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_2), svreinterpret_s64_s8(q8bytes_3)));
+                            r3 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_2), svreinterpret_s64_s8(q8bytes_3)));
+                            sumi1 = svmmla(svmmla(svdup_n_s32(0), r0, l0), r1, l1);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg256_all, svlsl_n_u32_x(pg256_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-1)), 24));
+                            acc_sumif1 = svmla_s32_x(pg256_all, acc_sumif1, svscales, sumi1);
+                            sumi2 = svmmla(svmmla(svdup_n_s32(0), r2, l2), r3, l3);
+                            svscales = svreinterpret_s32_u32(svlsr_n_u32_x(pg256_all, svlsl_n_u32_x(pg256_all, svreinterpret_u32_u64(svdup_n_u64(new_utmp.u64[j/2])), 8*(4-2*(j%2)-2)), 24));
+                            acc_sumif2 = svmla_s32_x(pg256_all, acc_sumif2, svscales, sumi2);
+                            q4_0 += 32; q4_1 += 32; q8_0 += 64; q8_1 += 64;
+                        }
+                        svint32_t acc_sumif = svadd_s32_x(pg256_all, acc_sumif1, acc_sumif2);
+                        svint32_t swap_acc_sumif = svext_s32(acc_sumif, acc_sumif, 4);
+                        acc_sumif = svadd_s32_x(pg32_4, acc_sumif, swap_acc_sumif);
+                        sumf1 = svmla_f32_x(pg32_4,
+                                svmla_f32_x(pg32_4,
+                                    sumf1,
+                                    svcvt_f32_x(pg32_4, acc_sumif),
+                                    svsuper_block_scales),
+                                svdmins,
+                                svsumfs_tmp);
+                    } // end of for nb
+                } // end of case 256-512
+                break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        }
+
+        svst1_f32(pg32_2, s, sumf1);
+        svst1_f32(pg32_2, s + bs, svreinterpret_f32_u8(svext_u8(svreinterpret_u8_f32(sumf1), svdup_n_u8(0), 8)));
+
+        return;
+    }
+#elif defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
         const block_q4_K * GGML_RESTRICT x0 = x;
         const block_q4_K * GGML_RESTRICT x1 = (const block_q4_K *) ((const uint8_t *)vx + bx);
@@ -2235,7 +2467,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT q4 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
-        const int vector_length = ggml_cpu_get_sve_cnt()*8;
         const svuint8_t m4b = svdup_n_u8(0xf);
         const svint32_t mzero = svdup_n_s32(0);
         svint32_t sumi1 = svdup_n_s32(0);
@@ -2480,7 +2711,201 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     const int nb = n / QK_K;
 
-#if defined(__ARM_FEATURE_MATMUL_INT8)
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+#endif
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const svbool_t pg32_2 = svptrue_pat_b32(SV_VL2);
+
+        svfloat32_t sum = svdup_n_f32(0);
+
+        const block_q6_K * GGML_RESTRICT vx0 = vx;
+        const block_q8_K * GGML_RESTRICT vy0 = vy;
+        const block_q6_K * GGML_RESTRICT vx1 = (const block_q6_K *) ((const uint8_t*)vx + bx);
+        const block_q8_K * GGML_RESTRICT vy1 = (const block_q8_K *) ((const uint8_t*)vy + by);
+
+        switch (vector_length) {
+            case 128:
+                {
+                    const svbool_t pg128_all = svptrue_pat_b8(SV_ALL);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT ql0 = vx0[i].ql;
+                        const uint8_t * GGML_RESTRICT qh0 = vx0[i].qh;
+                        const uint8_t * GGML_RESTRICT ql1 = vx1[i].ql;
+                        const uint8_t * GGML_RESTRICT qh1 = vx1[i].qh;
+                        const int8_t  * GGML_RESTRICT q80 = vy0[i].qs;
+                        const int8_t  * GGML_RESTRICT q81 = vy1[i].qs;
+
+                        const int8_t * GGML_RESTRICT scale0 = vx0[i].scales;
+                        const int8_t * GGML_RESTRICT scale1 = vx1[i].scales;
+
+                        svfloat32_t vy_d = svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d));
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg128_all, vy_d, vx_d);
+                        // process q8sum summation 128 bit route
+                        const svint16_t q8sums_01 = svld1_s16(pg128_all, vy0[i].bsums);
+                        const svint16_t q8sums_02 = svld1_s16(pg128_all, vy0[i].bsums + 8);
+                        const svint16_t q8sums_11 = svld1_s16(pg128_all, vy1[i].bsums);
+                        const svint16_t q8sums_12 = svld1_s16(pg128_all, vy1[i].bsums + 8);
+                        const svint64x2_t q6scales_0_tmp = svld2_s64(pg128_all, (const int64_t *)scale0);
+                        const svint16_t q6scales_01 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_0_tmp, 0)));
+                        const svint16_t q6scales_02 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_0_tmp, 1)));
+                        const svint64x2_t q6scales_1_tmp = svld2_s64(pg128_all, (const int64_t *)scale1);
+                        const svint16_t q6scales_11 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_1_tmp, 0)));
+                        const svint16_t q6scales_12 = svunpklo_s16(svreinterpret_s8_s64(svget2_s64(q6scales_1_tmp, 1)));
+                        const svint64_t prod = svdup_n_s64(0);
+
+                        svint32_t isum_tmp1 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_01, q6scales_01), q8sums_02, q6scales_02));
+                        svint32_t isum_tmp2 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_01, q6scales_11), q8sums_02, q6scales_12));
+                        svint32_t isum_tmp3 = svtrn1_s32(isum_tmp1, isum_tmp2);
+                        svint32_t isum_tmp4 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_11, q6scales_01), q8sums_12, q6scales_02));
+                        svint32_t isum_tmp5 = svreinterpret_s32_s64(svdot_s64(svdot_s64(prod, q8sums_11, q6scales_11), q8sums_12, q6scales_12));
+                        svint32_t isum_tmp6 = svtrn1_s32(isum_tmp4, isum_tmp5);
+                        svint32_t isum_tmp7 = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(isum_tmp3), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp8 = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(isum_tmp3), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t svisum_mins = svadd_s32_x(pg128_all, isum_tmp7, isum_tmp8);
+
+                        // process mmla
+                        svint8_t  l0, l1, r0, r1;
+                        svint32_t isum_tmp = svdup_n_s32(0);
+                        for (int j = 0; j < QK_K/128; ++j) {
+                            for (int k = 0; k < 8; ++k) {
+                                svuint8_t qhbits_0 = svld1_u8(pg128_all, qh0+16*(k%2));
+                                svuint8_t qhbits_1 = svld1_u8(pg128_all, qh1+16*(k%2));
+                                svuint8_t q6bits_0 = svld1_u8(pg128_all, ql0+16*(k%4));
+                                svuint8_t q6bits_1 = svld1_u8(pg128_all, ql1+16*(k%4));
+                                const int ql_pos = (k/4)*4;
+                                svuint8_t q6bytes_0_lo = (ql_pos < 4) ? svand_n_u8_x(pg128_all, q6bits_0, 0xf) : svlsr_n_u8_x(pg128_all, q6bits_0, 4);
+                                svuint8_t q6bytes_1_lo = (ql_pos < 4) ? svand_n_u8_x(pg128_all, q6bits_1, 0xf) : svlsr_n_u8_x(pg128_all, q6bits_1, 4);
+                                const int qh_pos = (k/2)*2;
+                                svuint8_t q6bytes_0_hi = svand_n_u8_x(pg128_all, qhbits_0, 0x3 << qh_pos);
+                                svuint8_t q6bytes_1_hi = svand_n_u8_x(pg128_all, qhbits_1, 0x3 << qh_pos);
+                                svint8_t  q6bytes_0, q6bytes_1;
+                                if (qh_pos <= 4) {
+                                    q6bytes_0 = svreinterpret_s8_u8(svmla_n_u8_x(pg128_all, q6bytes_0_lo, q6bytes_0_hi, 1 << (4 - qh_pos)));
+                                    q6bytes_1 = svreinterpret_s8_u8(svmla_n_u8_x(pg128_all, q6bytes_1_lo, q6bytes_1_hi, 1 << (4 - qh_pos)));
+                                } else {
+                                    q6bytes_0 = svreinterpret_s8_u8(svorr_u8_x(pg128_all, q6bytes_0_lo, svlsr_n_u8_x(pg128_all, q6bytes_0_hi, (qh_pos - 4))));
+                                    q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg128_all, q6bytes_1_lo, svlsr_n_u8_x(pg128_all, q6bytes_1_hi, (qh_pos - 4))));
+                                }
+                                svint8_t  q8bytes_0 = svld1_s8(pg128_all, q80+16*(k%8));
+                                svint8_t  q8bytes_1 = svld1_s8(pg128_all, q81+16*(k%8));
+                                l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                svint32_t svscale = svzip1_s32(svdup_n_s32(scale0[k]), svdup_n_s32(scale1[k]));
+                                isum_tmp = svmla_s32_x(pg128_all, isum_tmp, svmmla_s32(svmmla_s32(svdup_n_s32(0), r0, l0), r1, l1), svscale);
+                            }
+                            qh0 += 32;  qh1 += 32;
+                            ql0 += 64;  ql1 += 64;
+                            q80 += 128; q81 += 128;
+                            scale0 += 8; scale1 += 8;
+                        }
+                        sum = svmla_f32_x(pg128_all, sum,
+                                svcvt_f32_x(pg128_all, svmla_s32_x(pg128_all, isum_tmp,
+                                        svisum_mins, svdup_n_s32(-32))),
+                                svsuper_block_scales);
+                    }
+                } // end of case 128
+                break;
+            case 256:
+            case 512:
+                {
+                    const svbool_t pg256_all = svptrue_pat_b8(SV_ALL);
+                    const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4);
+                    for (int i = 0; i < nb; ++i) {
+                        const uint8_t * GGML_RESTRICT ql0 = vx0[i].ql;
+                        const uint8_t * GGML_RESTRICT qh0 = vx0[i].qh;
+                        const uint8_t * GGML_RESTRICT ql1 = vx1[i].ql;
+                        const uint8_t * GGML_RESTRICT qh1 = vx1[i].qh;
+                        const int8_t  * GGML_RESTRICT q80 = vy0[i].qs;
+                        const int8_t  * GGML_RESTRICT q81 = vy1[i].qs;
+
+                        const int8_t * GGML_RESTRICT scale0 = vx0[i].scales;
+                        const int8_t * GGML_RESTRICT scale1 = vx1[i].scales;
+                        svfloat32_t vx_d = svzip1_f32(svdup_n_f32(GGML_FP16_TO_FP32(vx0[i].d)), svdup_n_f32(GGML_FP16_TO_FP32(vx1[i].d)));
+                        svfloat64_t vy_d_tmp = svreinterpret_f64_f32(svuzp1_f32(svdup_n_f32(vy0[i].d), svdup_n_f32(vy1[i].d)));
+                        svfloat32_t vy_d = svreinterpret_f32_f64(svuzp1_f64(vy_d_tmp, vy_d_tmp));
+                        svfloat32_t svsuper_block_scales = svmul_f32_x(pg32_4, vy_d, vx_d);
+                        // process q8sum summation 256 bit route
+                        const svint16_t q8sums_0 = svld1_s16(pg256_all, vy0[i].bsums);
+                        const svint16_t q8sums_1 = svld1_s16(pg256_all, vy1[i].bsums);
+                        const svint16_t q6scales_0 = svunpklo_s16(svld1_s8(pg256_all, scale0));
+                        const svint16_t q6scales_1 = svunpklo_s16(svld1_s8(pg256_all, scale1));
+                        const svint64_t prod = svdup_n_s64(0);
+                        svint32_t isum_tmp1  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_0, q6scales_0));
+                        svint32_t isum_tmp2  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_0, q6scales_1));
+                        svint32_t isum_tmp3  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_1, q6scales_0));
+                        svint32_t isum_tmp4  = svreinterpret_s32_s64(svdot_s64(prod, q8sums_1, q6scales_1));
+                        svint32_t isum_tmp5  = svtrn1_s32(isum_tmp1, isum_tmp2);
+                        svint32_t isum_tmp6  = svtrn1_s32(isum_tmp3, isum_tmp4);
+                        svint32_t isum_tmp7  = svreinterpret_s32_s64(svtrn2_s64(svreinterpret_s64_s32(isum_tmp5), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp8  = svreinterpret_s32_s64(svtrn1_s64(svreinterpret_s64_s32(isum_tmp5), svreinterpret_s64_s32(isum_tmp6)));
+                        svint32_t isum_tmp9  = svadd_s32_x(pg256_all, isum_tmp7, isum_tmp8);
+                        svint32_t isum_tmp10 = svreinterpret_s32_u8(svext_u8(svreinterpret_u8_s32(isum_tmp9), svreinterpret_u8_s32(isum_tmp9), 16));
+                        svint32_t svisum_mins = svadd_s32_z(pg32_4, isum_tmp9, isum_tmp10);
+
+                        // process mmla
+                        svint8_t l0, l1, r0, r1;
+                        svint32_t isum_tmp = svdup_n_s32(0);
+                        for (int j = 0; j < QK_K/128; ++j) {
+                            for (int k = 0; k < 8; k+=2) { // process 2 block
+                                svuint8_t qhbits_0  = svld1_u8(pg256_all, qh0);
+                                svuint8_t qhbits_1  = svld1_u8(pg256_all, qh1);
+                                svuint8_t q6bits_0  = svld1_u8(pg256_all, ql0+32*((k%4)/2));
+                                svuint8_t q6bits_1  = svld1_u8(pg256_all, ql1+32*((k%4)/2));
+                                const int ql_pos = (k/4)*4;
+                                svuint8_t q6bytes_0_lo = (ql_pos < 4) ? svand_n_u8_x(pg256_all, q6bits_0, 0xf) : svlsr_n_u8_x(pg256_all, q6bits_0, 4);
+                                svuint8_t q6bytes_1_lo = (ql_pos < 4) ? svand_n_u8_x(pg256_all, q6bits_1, 0xf) : svlsr_n_u8_x(pg256_all, q6bits_1, 4);
+                                const int qh_pos = (k/2)*2;
+                                svuint8_t q6bytes_0_hi = svand_n_u8_x(pg256_all, qhbits_0, 0x3 << qh_pos);
+                                svuint8_t q6bytes_1_hi = svand_n_u8_x(pg256_all, qhbits_1, 0x3 << qh_pos);
+                                svint8_t  q6bytes_0, q6bytes_1;
+                                if (qh_pos <= 4) {
+                                    q6bytes_0 = svreinterpret_s8_u8(svmla_n_u8_x(pg256_all, q6bytes_0_lo, q6bytes_0_hi, 1 << (4 - qh_pos)));
+                                    q6bytes_1 = svreinterpret_s8_u8(svmla_n_u8_x(pg256_all, q6bytes_1_lo, q6bytes_1_hi, 1 << (4 - qh_pos)));
+                                } else {
+                                    q6bytes_0 = svreinterpret_s8_u8(svorr_u8_x(pg256_all, q6bytes_0_lo, svlsr_n_u8_x(pg256_all, q6bytes_0_hi, (qh_pos - 4))));
+                                    q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg256_all, q6bytes_1_lo, svlsr_n_u8_x(pg256_all, q6bytes_1_hi, (qh_pos - 4))));
+                                }
+                                svint8_t  q8bytes_0 = svld1_s8(pg256_all, q80+32*(k/2));
+                                svint8_t  q8bytes_1 = svld1_s8(pg256_all, q81+32*(k/2));
+                                l0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                l1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q6bytes_0), svreinterpret_s64_s8(q6bytes_1)));
+                                r0 = svreinterpret_s8_s64(svzip1_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                r1 = svreinterpret_s8_s64(svzip2_s64(svreinterpret_s64_s8(q8bytes_0), svreinterpret_s64_s8(q8bytes_1)));
+                                svint32_t svscale0 = svzip1_s32(svdup_n_s32(scale0[k]), svdup_n_s32(scale1[k]));
+                                svint32_t svscale1 = svzip1_s32(svdup_n_s32(scale0[k+1]), svdup_n_s32(scale1[k+1]));
+                                isum_tmp = svmla_s32_x(pg256_all, isum_tmp, svmmla_s32(svdup_n_s32(0), r0, l0), svscale0);
+                                isum_tmp = svmla_s32_x(pg256_all, isum_tmp, svmmla_s32(svdup_n_s32(0), r1, l1), svscale1);
+                            }
+                            qh0 += 32;  qh1 += 32;
+                            ql0 += 64;  ql1 += 64;
+                            q80 += 128; q81 += 128;
+                            scale0 += 8; scale1 += 8;
+                        } // end of for
+                        svint32_t swap_isum_tmp = svext_s32(isum_tmp, isum_tmp, 4);
+                        isum_tmp = svadd_s32_x(pg32_4, isum_tmp, swap_isum_tmp);
+                        sum = svmla_f32_x(pg32_4, sum,
+                                svcvt_f32_x(pg32_4, svmla_s32_x(pg32_4, isum_tmp,
+                                        svisum_mins, svdup_n_s32(-32))),
+                                svsuper_block_scales);
+                    }
+                } // end of case 256
+                break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        } // end of switch
+
+        svst1_f32(pg32_2, s, sum);
+        svst1_f32(pg32_2, s + bs, svreinterpret_f32_u8(svext_u8(svreinterpret_u8_f32(sum), svdup_n_u8(0), 8)));
+
+        return;
+    }
+#elif defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
         const block_q6_K * GGML_RESTRICT x0 = x;
         const block_q6_K * GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
@@ -2594,27 +3019,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             // adjust bias, apply superblock scale
             {
                 int32_t bias[4];
-#ifdef __ARM_FEATURE_SVE
-                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
-                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
-                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
-                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
-                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
-                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
-                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
-                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
-                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
-                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
-                const svint64_t zero = svdup_n_s64(0);
-                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
-                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
-                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
-                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
-#else
                 // NEON doesn't support int16 dot product, fallback to separated mul and add
                 const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
                 const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
@@ -2646,7 +3050,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                            vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
                 bias[3] = vaddvq_s32(prod);
 
-#endif
                 const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
 
                 const float32x4_t superblock_scale = {
@@ -2672,7 +3075,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 
 #ifdef __ARM_FEATURE_SVE
-    const int vector_length = ggml_cpu_get_sve_cnt()*8;
     float sum = 0;
     svuint8_t m4b = svdup_n_u8(0xf);
     svint32_t vzero = svdup_n_s32(0);
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/repack.cpp
index fdd0a513..082bd2bf 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/repack.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/repack.cpp
@@ -24,6 +24,29 @@
 
 #define UNUSED GGML_UNUSED
 
+static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
+                                             int16x8_t *     out_mins,
+                                             int8_t *        out_scales) {
+    constexpr uint32_t kmask1 = 0x3f3f3f3f;
+    constexpr uint32_t kmask2 = 0x0f0f0f0f;
+    constexpr uint32_t kmask3 = 0x03030303;
+    constexpr uint8_t  scales_size = 12;
+
+    uint32_t sm[3];
+    memcpy(sm, scales_in, scales_size);
+
+    const uint32_t   mins_0_3 = sm[1] & kmask1;
+    const uint32_t   mins_4_7 = ((sm[2] >> 4) & kmask2) | (((sm[1] >> 6) & kmask3) << 4);
+    const uint32x2_t mins_u32 = { mins_0_3, mins_4_7 };
+
+    *out_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins_u32)));
+
+    uint32_t scales_u32[2];
+    scales_u32[0] = sm[0] & kmask1;
+    scales_u32[1] = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
+    memcpy(out_scales, scales_u32, 8);
+}
+
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -474,6 +497,295 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_groups = ncols_interleaved / 4; // 0123 and 4567
+    const uint8x16_t m4b        = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[col_groups];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < col_groups; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q4_d_0        = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q4_d_1        = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d          = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0123 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_4567 = vmulq_f32(q4_d_1, q8_d);
+            float32x4_t q4_dmin_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q4_dmin_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0123   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_4567   = vmulq_f32(q4_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            int32x4_t acc_lo[col_groups];
+            int32x4_t acc_hi[col_groups];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_groups; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q4sb_mins[2];
+                int16x8_t q4sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q4sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                }
+
+                int8x16_t q8_qs[64 / 16];
+                for (int i = 0; i < 64 / 16; i++) {
+                    q8_qs[i] = vld1q_s8(q8_ptr[b].qs + sb * 64 + i * 16);
+                }
+
+                for (int c = 0; c < col_groups; c++) {
+                    uint8x16_t q4_cols[8];
+                    for (int i = 0; i < 8; i++) {
+                        q4_cols[i] = vld1q_u8(q4_ptr[b].qs + sb * QK_K + i * 32 + 16 * c);
+                    }
+
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[0], m4b)), q8_qs[0], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[1], m4b)), q8_qs[0], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[2], m4b)), q8_qs[0], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[3], m4b)), q8_qs[0], 3);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[4], m4b)), q8_qs[1], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[5], m4b)), q8_qs[1], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[6], m4b)), q8_qs[1], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[7], m4b)), q8_qs[1], 3);
+
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[0], 4)), q8_qs[2], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[1], 4)), q8_qs[2], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[2], 4)), q8_qs[2], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[3], 4)), q8_qs[2], 3);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[4], 4)), q8_qs[3], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[5], 4)), q8_qs[3], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[6], 4)), q8_qs[3], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[7], 4)), q8_qs[3], 3);
+                }
+
+                // Scales
+                // row c0123 blk0 and blk1
+                const int16x4_t   sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                const int16x4_t   sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                const float32x4_t sumf_0123  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[0]),
+                                                                       vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[0])));
+                acc_f32[0]                   = vfmaq_f32(acc_f32[0], sb_scale_0123, sumf_0123);
+                // row c4567 blk0 and blk1
+                const int16x4_t   sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                const int16x4_t   sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                const float32x4_t sumf_4567  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[1]),
+                                                                       vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[1])));
+                acc_f32[1]                   = vfmaq_f32(acc_f32[1], sb_scale_4567, sumf_4567);
+
+                // Bias Correction
+                const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0123);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_4567);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // #if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q4_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q4_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_pairs = ncols_interleaved / 2;
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[ncols_interleaved / 4];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < ncols_interleaved / 4; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q4_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q4_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q4_d_1, q8_d);
+            float32x4_t q4_dmin_0  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q4_dmin_1  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_1   = vmulq_f32(q4_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            // 2 sb each iteration
+            int32x4_t acc_lo[col_pairs];
+            int32x4_t acc_hi[col_pairs];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_pairs; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                int16x8_t q4sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q4sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                }
+
+                const uint8_t * q4_base = q4_ptr[b].qs + sb * QK_K;
+
+                // Load the 64 quants from q8K duplicated to use vecdots with the interelaved columns
+                // but still need the qs to use the low and hi bits from q4
+                const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
+                int8x16_t      q8_qs[8];
+                for (int i = 0; i < 8; i++) {
+                    q8_qs[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base + i * 8));
+                }
+
+                // Q4s columns iterated in pairs (01, 23, 45, 67)
+                for (int cp = 0; cp < col_pairs; cp++) {
+                    uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_base + 16 * cp);
+                    uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_base + 16 * cp + 64);
+                    uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_base + 16 * cp + 128);
+                    uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_base + 16 * cp + 192);
+
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)), q8_qs[0]);  // 0 .. 7
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)), q8_qs[1]);  // 8 ..15
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)), q8_qs[2]);  // 16..23
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)), q8_qs[3]);  // 24..31
+
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)), q8_qs[4]);  // 32..39
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)), q8_qs[5]);  // 40..47
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)), q8_qs[6]);  // 48..55
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)), q8_qs[7]);  // 56..63
+                }
+
+                // Iterates over a pair of column pairs (4 columns) to use a single 128 register
+                // p = 0 -> 0123  p2 -> 4567
+                for (int i = 0, p = 0; p < col_pairs; i++, p += 2) {
+                    int16x4_t   group_scales_lo = p == 0 ? vget_low_s16(q4sb_scales[0]) : vget_high_s16(q4sb_scales[0]);
+                    int16x4_t   group_scales_hi = p == 0 ? vget_low_s16(q4sb_scales[1]) : vget_high_s16(q4sb_scales[1]);
+                    float32x4_t sb_scale        = p == 0 ? sb_scale_0 : sb_scale_1;
+
+                    // 0123 or 4567
+                    float32x4_t sumf_0 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_lo), vpaddq_s32(acc_lo[p], acc_lo[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_0);
+
+                    float32x4_t sumf_1 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_hi), vpaddq_s32(acc_hi[p], acc_hi[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_1);
+                }
+
+                // Multiply Acc bsum + mins
+                // Each pair of subblocks share the same bsums
+                // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                // cols 0-3 bias
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                // cols 4-7 bias
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_1);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -1889,3 +2201,412 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
     ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
+
+void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    acc_size  = 2 * 4;  // 2 row pairs × 4 col pairs
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[acc_size];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < acc_size; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // d4 0 1 2 3, 4 5 6 7
+                float32x4_t q4_d_0123    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));
+                float32x4_t q4_d_4567    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));
+                // d8 0 1 2 3
+                float32x4_t q8_d_0123    = vld1q_f32(q8_ptr[b].d);
+                // mins
+                float32x4_t q4_dmin_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));
+                float32x4_t q4_dmin_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));
+
+                // Precomputation of scales and mins
+                float32x4_t sbd_scale_0123[q8_k_blocklen];
+                float32x4_t sbd_scale_4567[q8_k_blocklen];
+                float32x4_t sbd_min_0123[q8_k_blocklen];
+                float32x4_t sbd_min_4567[q8_k_blocklen];
+
+                sbd_scale_0123[0] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 0);
+                sbd_scale_4567[0] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 0);
+                sbd_min_0123[0]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 0);
+                sbd_min_4567[0]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 0);
+
+                sbd_scale_0123[1] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 1);
+                sbd_scale_4567[1] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 1);
+                sbd_min_0123[1]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 1);
+                sbd_min_4567[1]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 1);
+
+                sbd_scale_0123[2] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 2);
+                sbd_scale_4567[2] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 2);
+                sbd_min_0123[2]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 2);
+                sbd_min_4567[2]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 2);
+
+                sbd_scale_0123[3] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 3);
+                sbd_scale_4567[3] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 3);
+                sbd_min_0123[3]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 3);
+                sbd_min_4567[3]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 3);
+
+                // Precomputation of bsums, each vpaddq calcs all the bsums for each row
+                const int16x8_t bsums[q8_k_blocklen] = {
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[QK_K / 64][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                // interleaved bias_acc: [0]->r0 0123, [1]->r1 0123, .., [4]->r0 4567, [5]->r1 4567 ..
+                int32x4_t bias_acc[acc_size];
+                for (int i = 0; i < acc_size; i++) {
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Int accumulators for qs vecdot (4 row x 2 col quartets)
+                    int32x4_t acc_lo[acc_size];
+                    int32x4_t acc_hi[acc_size];
+                    for (int i = 0; i < acc_size; i++) {
+                        acc_lo[i] = vdupq_n_s32(0);
+                        acc_hi[i] = vdupq_n_s32(0);
+                    }
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int16x8_t q4sb_scales[2];
+                    int16x8_t q4sb_mins[2];
+                    for (int i = 0; i < 2; i++) {
+                        int8_t    aux_q4sb[8];
+                        const int offset = sb * 24 + i * 12;
+                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                        q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                    }
+
+                    constexpr int reads_per_sb = 8;  // 8 * 16 bytes each => 32 qs * 4 rows
+                    for (int k = 0; k < reads_per_sb; k++) {
+                        const int8x16_t q8_blk0 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k);
+                        const int8x16_t q8_blk1 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k + 128);
+
+                        // 0..3 & 32..35
+                        const uint8x16_t q4_0123 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k);
+                        const uint8x16_t q4_4567 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k + 16);
+
+                        const int8x16_t q4_0123_lo = vreinterpretq_s8_u8(vandq_u8(q4_0123, m4b));
+                        const int8x16_t q4_0123_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_0123, 4));
+
+                        acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q4_0123_lo, q8_blk0, 0);  //  0..3  r0 c0123
+                        acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q4_0123_lo, q8_blk0, 1);  //  0..3  r1 c0123
+                        acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q4_0123_lo, q8_blk0, 2);  //  0..3  r2 c0123
+                        acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q4_0123_lo, q8_blk0, 3);  //  0..3  r3 c0123
+
+                        acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q4_0123_hi, q8_blk1, 0);  // 32..35 r0 c0123
+                        acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q4_0123_hi, q8_blk1, 1);  // 32..35 r1 c0123
+                        acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q4_0123_hi, q8_blk1, 2);  // 32..35 r2 c0123
+                        acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q4_0123_hi, q8_blk1, 3);  // 32..35 r3 c0123
+
+                        const int8x16_t q4_4567_lo = vreinterpretq_s8_u8(vandq_u8(q4_4567, m4b));
+                        const int8x16_t q4_4567_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_4567, 4));
+
+                        acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q4_4567_lo, q8_blk0, 0);  //  0..3  r0 c4567
+                        acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q4_4567_lo, q8_blk0, 1);  //  0..3  r1 c4567
+                        acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q4_4567_lo, q8_blk0, 2);  //  0..3  r2 c4567
+                        acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q4_4567_lo, q8_blk0, 3);  //  0..3  r3 c4567
+
+                        acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q4_4567_hi, q8_blk1, 0);  // 32..35 r0 c4567
+                        acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q4_4567_hi, q8_blk1, 1);  // 32..35 r1 c4567
+                        acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q4_4567_hi, q8_blk1, 2);  // 32..35 r2 c4567
+                        acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q4_4567_hi, q8_blk1, 3);  // 32..35 r3 c4567
+                    }
+
+                    // Scale and bias application
+                    // acc is stored interleaved to match output layout
+                    const int16x4_t sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                    const int16x4_t sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                    const int16x4_t sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                    const int16x4_t sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                    for (int row = 0; row < q8_k_blocklen; row++) {
+                        // Bias correction
+                        // row c0123 blk0 and blk1
+                        const float32x4_t sumf_0123 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[row]),
+                                                    vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[row])));
+                        acc_f32[2 * row] = vfmaq_f32(acc_f32[2 * row], sbd_scale_0123[row], sumf_0123);
+
+                        // row c4567 blk0 and blk1
+                        const float32x4_t sumf_4567 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[row + 4]),
+                                                    vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[row + 4])));
+                        acc_f32[2 * row + 1] = vfmaq_f32(acc_f32[2 * row + 1], sbd_scale_4567[row], sumf_4567);
+
+                        // Bias
+                        const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][row * 2]);
+                        const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][row * 2 + 1]);
+
+                        // row c0123 blk0 and blk1
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                        // row c4567 blk0 and blk1
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                for (int row = 0; row < q8_k_blocklen; row++) {
+                    acc_f32[2 * row] = vmlsq_f32(acc_f32[2 * row], vcvtq_f32_s32(bias_acc[2 * row]), sbd_min_0123[row]);
+                    acc_f32[2 * row + 1] =
+                        vmlsq_f32(acc_f32[2 * row + 1], vcvtq_f32_s32(bias_acc[2 * row + 1]), sbd_min_4567[row]);
+                }
+            }  // for b
+
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemm_q4_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemm_q4_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    constexpr int    q8_k_blocklen = 4;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[blocklen];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < blocklen; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // bsums pairs belongs to the same q8_k subblock
+                const int16x8_t bsums[4]{
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[4][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                int32x4_t sb_acc[4];    // Aux accumulators to store subblock (partial) results
+                int32x4_t acc[8];       // rows 01 stored in [0][1][2][3] rows 23 stored in [4][5][6][7]
+                int32x4_t bias_acc[8];  // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567, [2]->r1 0123 ...
+                for (int i = 0; i < 8; i++) {
+                    acc[i]      = vdupq_n_s32(0);
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int8_t    q4sb_scales[2][8];
+                    int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                    for (int i = 0; i < 2; i++) {
+                        const int offset = sb * 24 + i * 12;
+                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
+                    }
+
+                    // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
+                    const int8_t * q8_base = q8_ptr[b].qs + sb * 256;
+
+                    int8x16_t q8_qs_01[8];
+                    int8x16_t q8_qs_23[8];
+
+                    // Load 32-byte per row pair, 1 subblock each time
+                    for (int i = 0; i < 8; i++) {
+                        const int offset = i * 32;  // 16 for row 01, 16 for row 23
+                        q8_qs_01[i]      = vld1q_s8(q8_base + offset);
+                        q8_qs_23[i]      = vld1q_s8(q8_base + offset + 16);
+                    }
+
+                    const int8x16_t q8s[2][8] = {
+                        { q8_qs_01[0], q8_qs_01[1], q8_qs_01[2], q8_qs_01[3],
+                          q8_qs_01[4], q8_qs_01[5], q8_qs_01[6], q8_qs_01[7] },
+                        { q8_qs_23[0], q8_qs_23[1], q8_qs_23[2], q8_qs_23[3],
+                          q8_qs_23[4], q8_qs_23[5], q8_qs_23[6], q8_qs_23[7] },
+                    };
+
+                    // Q4s columns iterated in pairs (01, 23, 45, 67)
+                    for (int cp = 0; cp < ncols_interleaved / 2; cp++) {
+                        for (int i = 0; i < 4; i++) {
+                            sb_acc[i] = vdupq_n_s32(0);
+                        }
+
+                        uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 0);    // 0 .. 7 & 32..39
+                        uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 64);   // 8 ..15 & 40..47
+                        uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 128);  // 16..23 & 48..55
+                        uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 192);  // 24..31 & 56..63
+                        const int8x16_t q4_nibbles[2][4] = {
+                            {
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)),
+                            },
+                            {
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)),
+                            }
+                        };
+
+                        // Calculates the Qs muladd of every row pair (rp) rows 01 and 23 of q8
+                        // for each of the internal 32 qs subblock (blk)
+                        for (int rp = 0; rp < 2; rp++) {
+                            for (int blk = 0; blk < 2; blk++) {
+                                const int8x16_t * q8  = &q8s[rp][4 * blk];
+                                const int8x16_t * q4  = q4_nibbles[blk];
+                                int32x4_t         acc = sb_acc[2 * rp + blk];
+                                // mul add for each qs in the same subblock
+                                for (int qs_offset = 0; qs_offset < 4; qs_offset++) {
+                                    acc = vmmlaq_s32(acc, q4[qs_offset], q8[qs_offset]);
+                                }
+                                sb_acc[2 * rp + blk] = acc;
+                            }
+                        }
+
+                        // Scales[i] corresponds to column i
+                        const int scale_offset = cp * 2;
+                        for (int blk = 0; blk < 2; blk++) {
+                            const int32x4_t block_scale = {
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                            };
+                            acc[cp]     = vmlaq_s32(acc[cp], sb_acc[blk], block_scale);
+                            acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[blk + 2], block_scale);
+                        }
+                    }
+
+                    // Multiply Acc bsum + mins
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        // Each pair of subblocks share the same bsums
+                        // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                        int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][q8_row * 2]);
+                        int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][q8_row * 2 + 1]);
+
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                // Reorder of i8mm output with bias and output layout
+                for (int i = 0; i < 8; i++) {
+                    int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i]));
+                    acc[i]          = vcombine_s32(aux.val[0], aux.val[1]);
+                }
+                int32x4_t reorder_acc[8] = {
+                    vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])),
+                    vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])),
+                    vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])),
+                    vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])),
+                    vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])),
+                    vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])),
+                    vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])),
+                    vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])),
+                };
+
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    for (int j = 0; j < 2; j++) {
+                        float32x4_t       q8_d    = vdupq_n_f32(q8_ptr[b].d[i]);
+                        float32x4_t       q4_dmin = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].dmin + j * 4)));
+                        const float32x4_t dmins   = vmulq_f32(q4_dmin, q8_d);
+
+                        float32x4_t       q4_d  = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].d + j * 4)));
+                        const float32x4_t scale = vmulq_f32(q4_d, q8_d);
+
+                        acc_f32[2 * i + j] = vmlsq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(bias_acc[2 * i + j]), dmins);
+                        acc_f32[2 * i + j] =
+                            vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale);
+                    }
+                }
+            }  // for b
+
+            // With the previous reorder, the tile is already in the correct memory layout.
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ml/backend/ggml/ggml/src/ggml-cpu/arch/x86/repack.cpp
index fe18225c..7dda9eea 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -646,7 +646,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
     __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
     int64_t xstart = 0;
     int anr = nr - nr%16; // Used to align nr with boundary of 16
-#ifdef __AVX512F__
+#if defined(__AVX512BW__) && defined(__AVX512DQ__)
     int anc = nc - nc%16; // Used to align nc with boundary of 16
                           // Mask to mask out nibbles from packed bytes expanded to 512 bit length
     const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
@@ -1041,7 +1041,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
         xstart = anc/8;
         y = 0;
     }
-#endif // __AVX512F__
+#endif // __AVX512BW__ && __AVX512DQ__
 
     // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
 
@@ -1989,7 +1989,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
     int64_t xstart = 0;
     int anr = nr - nr % 16;; // Used to align nr with boundary of 16
-#ifdef __AVX512F__
+#if defined(__AVX512BW__) && defined(__AVX512DQ__)
     int anc = nc - nc % 16; // Used to align nc with boundary of 16
     // Mask to mask out nibbles from packed bytes expanded to 512 bit length
     const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
@@ -2727,7 +2727,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         xstart = anc/8;
         y = 0;
     }
-#endif //AVX512F
+#endif // __AVX512BW__ && __AVX512DQ__
 
     // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
     for (; y < anr / 4; y += 4) {
@@ -3467,7 +3467,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     __m256i scalesmask2 = _mm256_castsi128_si256(scalesmask2_sse);
     scalesmask2 = _mm256_permute2f128_si256(scalesmask2, scalesmask2, 0);
 
-#ifdef __AVX512F__
+#if defined(__AVX512BW__) && defined(__AVX512DQ__)
 
     int anc = nc - nc % 16; // Used to align nc with boundary of 16
 
@@ -4947,7 +4947,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         y = 0;
     }
 
-#endif //AVX512F
+#endif // __AVX512BW__ && __AVX512DQ__
 
     // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
     for (; y < anr / 4; y += 4) {
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu-impl.h b/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu-impl.h
index 713bf85e..7597377c 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu-impl.h
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu-impl.h
@@ -500,13 +500,15 @@ inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
 
 #endif
 
-#if defined(__loongarch_asx)
+#if defined(__loongarch_sx)
 /* float type data load instructions */
 static __m128 __lsx_vreplfr2vr_s(const float val) {
     v4f32 res = {val, val, val, val};
     return (__m128)res;
 }
+#endif
 
+#if defined(__loongarch_asx)
 static __m256 __lasx_xvreplfr2vr_s(const float val) {
     v8f32 res = {val, val, val, val, val, val, val, val};
     return (__m256)res;
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c b/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c
index 04664628..7a0df30c 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c
@@ -1615,13 +1615,8 @@ static void ggml_compute_forward_mul_mat_id(
             chunk_size = 64;
         }
 
-#if defined(__aarch64__)
-        // disable for ARM
-        const bool disable_chunking = true;
-#else
         // disable for NUMA
         const bool disable_chunking = ggml_is_numa();
-#endif // defined(__aarch64__)
 
         int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
         int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
@@ -1738,6 +1733,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_sum_rows(params, tensor);
             } break;
+        case GGML_OP_CUMSUM:
+            {
+                ggml_compute_forward_cumsum(params, tensor);
+            } break;
         case GGML_OP_MEAN:
             {
                 ggml_compute_forward_mean(params, tensor);
@@ -1814,22 +1813,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_cont(params, tensor);
             } break;
-        case GGML_OP_RESHAPE:
-            {
-                ggml_compute_forward_reshape(params, tensor);
-            } break;
-        case GGML_OP_VIEW:
-            {
-                ggml_compute_forward_view(params, tensor);
-            } break;
-        case GGML_OP_PERMUTE:
-            {
-                ggml_compute_forward_permute(params, tensor);
-            } break;
-        case GGML_OP_TRANSPOSE:
-            {
-                ggml_compute_forward_transpose(params, tensor);
-            } break;
         case GGML_OP_GET_ROWS:
             {
                 ggml_compute_forward_get_rows(params, tensor);
@@ -1946,10 +1929,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_argsort(params, tensor);
             } break;
+        case GGML_OP_TOP_K:
+            {
+                ggml_compute_forward_top_k(params, tensor);
+            } break;
         case GGML_OP_LEAKY_RELU:
             {
                 ggml_compute_forward_leaky_relu(params, tensor);
             } break;
+        case GGML_OP_TRI:
+            {
+                ggml_compute_forward_tri(params, tensor);
+            } break;
+        case GGML_OP_FILL:
+            {
+                ggml_compute_forward_fill(params, tensor);
+            } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_compute_forward_flash_attn_ext(params, tensor);
@@ -2005,6 +2000,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rwkv_wkv7(params, tensor);
             } break;
+        case GGML_OP_SOLVE_TRI:
+            {
+                ggml_compute_forward_solve_tri(params, tensor);
+            } break;
         case GGML_OP_MAP_CUSTOM1:
             {
                 ggml_compute_forward_map_custom1(params, tensor);
@@ -2049,6 +2048,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 // nop
             } break;
+        case GGML_OP_RESHAPE:
+            {
+                // nop
+            } break;
+        case GGML_OP_PERMUTE:
+            {
+                // nop
+            } break;
+        case GGML_OP_VIEW:
+            {
+                // nop
+            } break;
+        case GGML_OP_TRANSPOSE:
+            {
+                // nop
+            } break;
         case GGML_OP_COUNT:
             {
                 GGML_ABORT("fatal error");
@@ -2147,6 +2162,9 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ADD_ID:
         case GGML_OP_ADD1:
         case GGML_OP_ACC:
+        case GGML_OP_CUMSUM:
+        case GGML_OP_TRI:
+        case GGML_OP_FILL:
             {
                 n_tasks = n_threads;
             } break;
@@ -2164,6 +2182,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = 1;
             } break;
         case GGML_OP_COUNT_EQUAL:
+        case GGML_OP_SOLVE_TRI:
             {
                 n_tasks = n_threads;
             } break;
@@ -2186,6 +2205,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_SOFTPLUS:
+                case GGML_UNARY_OP_EXPM1:
                 case GGML_UNARY_OP_FLOOR:
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:
@@ -2296,6 +2317,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
+        case GGML_OP_TOP_K:
         case GGML_OP_FLASH_ATTN_EXT:
         case GGML_OP_FLASH_ATTN_BACK:
         case GGML_OP_SSM_CONV:
@@ -2819,6 +2841,10 @@ struct ggml_cplan ggml_graph_plan(
                         cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
                         cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
                     } break;
+                case GGML_OP_TOP_K:
+                    {
+                        cur += sizeof(int32_t)*node->src[0]->ne[0]*n_tasks;
+                    } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
                         const int64_t ne10 = node->src[1]->ne[0]; // DK
@@ -2891,6 +2917,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
+        if (ggml_op_is_empty(node->op)) {
+            // skip NOPs
+            continue;
+        }
+
         ggml_compute_forward(&params, node);
 
 #ifdef OLLAMA_DEBUG
@@ -3280,6 +3311,13 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
         __m128 y_vec = _mm_cvtph_ps(x_vec);
         _mm_storeu_ps(y + i, y_vec);
     }
+#elif defined(__riscv_zvfh)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m1(n - i);
+        vfloat16m1_t vx = __riscv_vle16_v_f16m1((_Float16 *)&x[i], vl);
+        vfloat32m2_t vy = __riscv_vfwcvt_f_f_v_f32m2(vx, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
 #endif
 
     for (; i < n; ++i) {
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/ops.cpp b/ml/backend/ggml/ggml/src/ggml-cpu/ops.cpp
index 70955347..3155cb4b 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/ops.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/ops.cpp
@@ -7,8 +7,10 @@
 #include "unary-ops.h"
 #include "vec.h"
 
-#include <float.h>
+#include <cfloat>
 #include <algorithm>
+#include <cmath>
+#include <functional>
 
 // ggml_compute_forward_dup
 
@@ -1394,6 +1396,56 @@ void ggml_compute_forward_sum(
     }
 }
 
+// ggml_compute_forward_cumsum
+
+static void ggml_compute_forward_cumsum_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(dst->nb[0] == sizeof(float));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(ne0 == ne00);
+    GGML_ASSERT(ne1 == ne01);
+    GGML_ASSERT(ne2 == ne02);
+    GGML_ASSERT(ne3 == ne03);
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        float * src_row = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+        float * dst_row = (float *) ((char *) dst->data  + i01*nb1  + i02*nb2  + i03*nb3);
+
+        ggml_vec_cumsum_f32(ne00, dst_row, src_row);
+    }
+}
+
+void ggml_compute_forward_cumsum(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_cumsum_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_sum_rows
 
 static void ggml_compute_forward_sum_rows_f32(
@@ -2140,6 +2192,83 @@ static void ggml_compute_forward_gelu(
     }
 }
 
+// ggml_compute_fill
+
+static void ggml_compute_forward_fill_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const float c = ggml_get_op_params_f32(dst, 0);
+
+    GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
+    GGML_TENSOR_LOCALS(size_t,  nb, dst, nb);
+
+    const auto [ir0, ir1] = get_thread_range(params, dst);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne2*ne1);
+        const int64_t i02 = (ir - i03*ne2*ne1)/ne1;
+        const int64_t i01 = (ir - i03*ne2*ne1 - i02*ne1);
+
+        float * dst_ptr  = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1);
+
+        ggml_vec_set_f32(ne0, dst_ptr, c);
+    }
+}
+
+void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor * dst) {
+    ggml_compute_forward_fill_f32(params, dst);
+}
+
+// ggml_compute_tri
+
+static void ggml_compute_forward_tri_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    const ggml_tri_type ttype = (ggml_tri_type) ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    const auto [ir0, ir1] = get_thread_range(params, src0);
+
+    bool (*bipred)(int, int);
+
+    switch (ttype) {
+        case GGML_TRI_TYPE_LOWER:      bipred = [](int i, int r) { return i <  r; }; break;
+        case GGML_TRI_TYPE_LOWER_DIAG: bipred = [](int i, int r) { return i <= r; }; break;
+        case GGML_TRI_TYPE_UPPER:      bipred = [](int i, int r) { return i >  r; }; break;
+        case GGML_TRI_TYPE_UPPER_DIAG: bipred = [](int i, int r) { return i >= r; }; break;
+        default: GGML_ABORT("invalid tri type");
+    }
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+        const float * src_ptr = (const float  *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+              float * dst_ptr = (      float  *) ((      char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1);
+
+        for (int i0 = 0; i0 < ne0; ++i0) {
+            dst_ptr[i0] = bipred(i0, i01) ? src_ptr[i0] : 0.0f;
+        }
+    }
+}
+
+void ggml_compute_forward_tri(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_tri_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_gelu_erf
 
 static void ggml_compute_forward_gelu_erf_f32(
@@ -4455,46 +4584,6 @@ void ggml_compute_forward_cont(
     ggml_compute_forward_dup(params, dst);
 }
 
-// ggml_compute_forward_reshape
-
-void ggml_compute_forward_reshape(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_view
-
-void ggml_compute_forward_view(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_permute
-
-void ggml_compute_forward_permute(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
-// ggml_compute_forward_transpose
-
-void ggml_compute_forward_transpose(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-    // NOP
-    GGML_UNUSED(params);
-    GGML_UNUSED(dst);
-}
-
 // ggml_compute_forward_get_rows
 
 static void ggml_compute_forward_get_rows_q(
@@ -5474,7 +5563,7 @@ static void ggml_rope_cache_init(
 }
 
 static void ggml_mrope_cache_init(
-     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool is_imrope, bool indep_sects,
      float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
      float * cache, float sin_sign, float theta_scale) {
     // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
@@ -5509,11 +5598,26 @@ static void ggml_mrope_cache_init(
         }
 
         float theta = theta_t;
-        if (sector % 3 == 1 && sector < 1 + 3 * sections[1]) {
-            theta = theta_h;
-        }
-        else if (sector % 3 == 2 && sector < 2 + 3 * sections[2]) {
-            theta = theta_w;
+        if (is_imrope) { // qwen3vl apply interleaved mrope
+            if (sector % 3 == 1 && sector < 1 + 3 * sections[1]) {
+                theta = theta_h;
+            } else if (sector % 3 == 2 && sector < 2 + 3 * sections[2]) {
+                theta = theta_w;
+            } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                theta = theta_t;
+            // } else {
+            //     theta = theta_e;
+            }
+        } else {
+            if (sector >= sections[0] && sector < sec_w) {
+                theta = theta_h;
+            }
+            else if (sector >= sec_w && sector < sec_w + sections[2]) {
+                theta = theta_w;
+            }
+            else if (sector >= sec_w + sections[2]) {
+                theta = theta_e;
+            }
         }
 
         rope_yarn(
@@ -5528,7 +5632,28 @@ static void ggml_mrope_cache_init(
     }
 }
 
-static void ggml_compute_forward_rope_f32(
+
+template<typename T>
+static void rotate_pairs(const int64_t n, const int64_t n_offset, const float * cache, const T * src_data, T * dst_data, const int scale = 2) {
+  for (int64_t i0 = 0; i0 < n; i0 += 2) {
+    const int64_t ic = i0/scale; // hack for GGML_ROPE_TYPE_NORMAL, where we need ic = i0; for all other cases, ic = i0/2
+
+    const float cos_theta = cache[i0 + 0];
+    const float sin_theta = cache[i0 + 1];
+
+    const T * const src = src_data + ic;
+    T * dst             = dst_data + ic;
+
+    const float x0 = type_conversion_table<T>::to_f32(src[0]);
+    const float x1 = type_conversion_table<T>::to_f32(src[n_offset]);
+
+    dst[0]        = type_conversion_table<T>::from_f32(x0*cos_theta - x1*sin_theta);
+    dst[n_offset] = type_conversion_table<T>::from_f32(x0*sin_theta + x1*cos_theta);
+  }
+}
+
+template<typename T> //float or ggml_fp16_t
+static void ggml_compute_forward_rope_flt(
         const ggml_compute_params * params,
         ggml_tensor * dst,
         const bool forward) {
@@ -5537,6 +5662,9 @@ static void ggml_compute_forward_rope_f32(
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * src2 = dst->src[2];
 
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
     int sections[4];
 
@@ -5559,7 +5687,8 @@ static void ggml_compute_forward_rope_f32(
     //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
     //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
 
-    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb0 == nb00);
+    GGML_ASSERT(nb0 == sizeof(T));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -5584,11 +5713,11 @@ static void ggml_compute_forward_rope_f32(
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
+    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
-    if (is_mrope) {
+    if (mrope_used) {
         GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
     }
 
@@ -5614,7 +5743,7 @@ static void ggml_compute_forward_rope_f32(
         for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len
 
             float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            if (!is_mrope) {
+            if (!mrope_used) {
                 const int64_t p = pos[i2];
                 ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
@@ -5624,7 +5753,7 @@ static void ggml_compute_forward_rope_f32(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -5632,268 +5761,36 @@ static void ggml_compute_forward_rope_f32(
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                if (is_neox || is_mrope) {
-                    if (is_vision){
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
+                T * src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+                T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1);
 
-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = src[0];
-                            const float x1 = src[n_dims];
-
-                            dst_data[0]      = x0*cos_theta - x1*sin_theta;
-                            dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
-                        }
-                    } else {
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
-
-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = src[0];
-                            const float x1 = src[n_dims/2];
-
-                            dst_data[0]        = x0*cos_theta - x1*sin_theta;
-                            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
-                        }
-                    }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        const float x0 = src[0];
-                        const float x1 = src[1];
-
-                        dst_data[0] = x0*cos_theta - x1*sin_theta;
-                        dst_data[1] = x0*sin_theta + x1*cos_theta;
-                    }
+                switch (mode) {
+                    case GGML_ROPE_TYPE_NORMAL:
+                        rotate_pairs<T>(n_dims, 1, cache, src, dst_data, 1);
+                        break;
+                    case GGML_ROPE_TYPE_NEOX:
+                    case GGML_ROPE_TYPE_MROPE:
+                    case GGML_ROPE_TYPE_IMROPE:
+                        rotate_pairs<T>(n_dims, n_dims/2, cache, src, dst_data);
+                        break;
+                    case GGML_ROPE_TYPE_VISION:
+                        rotate_pairs<T>(ne0, n_dims, cache, src, dst_data);
+                        break;
+                    default:
+                        GGML_ABORT("rope type not supported");
                 }
 
-                if (is_vision) {
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const int64_t ic = i0/2;
-
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                        const float x0 = src[0];
-                        const float x1 = src[n_dims];
-
-                        dst_data[0]      = x0*cos_theta - x1*sin_theta;
-                        dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
-                    }
-                } else {
+                if (!is_vision) {
                     // fill the remain channels with data from src tensor
                     for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                        const T * const src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                        T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
                         dst_data[0] = src[0];
                         dst_data[1] = src[1];
                     }
                 }
-            }
-        }
-    }
-}
-
-// TODO: deduplicate f16/f32 code
-static void ggml_compute_forward_rope_f16(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const bool forward) {
-
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-    const ggml_tensor * src2 = dst->src[2];
-
-    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-    int sections[4];
-
-    //const int n_past     = ((int32_t *) dst->op_params)[0];
-    const int n_dims     = ((int32_t *) dst->op_params)[1];
-    const int mode       = ((int32_t *) dst->op_params)[2];
-    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
-    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
-    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
-
-
-    GGML_TENSOR_UNARY_OP_LOCALS
-
-    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
-    //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
-
-    GGML_ASSERT(nb0 == sizeof(ggml_fp16_t));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nr = ggml_nrows(dst);
-
-    GGML_ASSERT(n_dims <= ne0);
-    GGML_ASSERT(n_dims % 2 == 0);
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    // row index used to determine which thread to use
-    int ir = 0;
-
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
-
-    float corr_dims[2];
-    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
-
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
-    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
-
-    if (is_mrope) {
-        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
-    }
-
-    if (is_vision) {
-        GGML_ASSERT(n_dims == ne0/2);
-    }
-
-    const float * freq_factors = NULL;
-    if (src2 != NULL) {
-        GGML_ASSERT(src2->type == GGML_TYPE_F32);
-        GGML_ASSERT(src2->ne[0] >= n_dims / 2);
-        freq_factors = (const float *) src2->data;
-    }
-
-    // backward process uses inverse rotation by cos and sin.
-    // cos and sin build a rotation matrix, where the inverse is the transpose.
-    // this essentially just switches the sign of sin.
-    const float sin_sign = forward ? 1.0f : -1.0f;
-
-    const int32_t * pos = (const int32_t *) src1->data;
-
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
-
-            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            if (!is_mrope) {
-                const int64_t p = pos[i2];
-                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-            else {
-                const int64_t p_t = pos[i2];
-                const int64_t p_h = pos[i2 + ne2];
-                const int64_t p_w = pos[i2 + ne2 * 2];
-                const int64_t p_e = pos[i2 + ne2 * 3];
-                ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
-                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
-                if (ir++ < ir0) continue;
-                if (ir   > ir1) break;
-
-                if (is_neox || is_mrope) {
-                    if (is_vision) {
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
-
-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                            const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
-
-                            dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                            dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                        }
-                    } else {
-                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                            const int64_t ic = i0/2;
-
-                            const float cos_theta = cache[i0 + 0];
-                            const float sin_theta = cache[i0 + 1];
-
-                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                            const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims/2]);
-
-                            dst_data[0]        = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                            dst_data[n_dims/2] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                        }
-                    }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_CPU_FP16_TO_FP32(src[1]);
-
-                        dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[1] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                    }
-                }
-
-                if (is_vision) {
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const int64_t ic = i0/2;
-
-                        const float cos_theta = cache[i0 + 0];
-                        const float sin_theta = cache[i0 + 1];
-
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
-
-                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
-
-                        dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                    }
-                } else {
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        dst_data[0] = src[0];
-                        dst_data[1] = src[1];
-                    }
-                }
-            }
+            } //attn-heads
         }
     }
 }
@@ -5907,11 +5804,11 @@ void ggml_compute_forward_rope(
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_f16(params, dst, true);
+                ggml_compute_forward_rope_flt<ggml_fp16_t>(params, dst, true);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_f32(params, dst, true);
+                ggml_compute_forward_rope_flt<float>(params, dst, true);
             } break;
         default:
             {
@@ -5931,11 +5828,11 @@ void ggml_compute_forward_rope_back(
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_f16(params, dst, false);
+                ggml_compute_forward_rope_flt<ggml_fp16_t>(params, dst, false);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_f32(params, dst, false);
+                ggml_compute_forward_rope_flt<float>(params, dst, false);
             } break;
         default:
             {
@@ -7067,7 +6964,11 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
     const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
 
 #ifdef GGML_SIMD
-    const int64_t pkg_size = GGML_F32_EPR;
+    #if defined(__ARM_FEATURE_SVE)
+        const int64_t pkg_size = svcntw();
+    #else
+        const int64_t pkg_size = GGML_F32_EPR;
+    #endif
     const int64_t pkg_count = c / pkg_size;
     const int64_t c_pkg_end = pkg_count * pkg_size;
 #else
@@ -7490,10 +7391,17 @@ static void ggml_compute_forward_upscale_f32(
     float sf1 = (float)ne1/src0->ne[1];
     float sf2 = (float)ne2/src0->ne[2];
     float sf3 = (float)ne3/src0->ne[3];
+    float pixel_offset = 0.5f;
 
     const int32_t mode_flags = ggml_get_op_params_i32(dst, 0);
     const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
 
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        pixel_offset = 0.0f;
+        sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+        sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+    }
+
     if (mode == GGML_SCALE_MODE_NEAREST) {
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
@@ -7513,13 +7421,6 @@ static void ggml_compute_forward_upscale_f32(
             }
         }
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
-        float pixel_offset = 0.5f;
-        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-            pixel_offset = 0.0f;
-            sf0 = (float)(ne0 - 1) / (src0->ne[0] - 1);
-            sf1 = (float)(ne1 - 1) / (src0->ne[1] - 1);
-        }
-
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
             for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
@@ -7554,6 +7455,51 @@ static void ggml_compute_forward_upscale_f32(
 
                         const float val = a*(1 - dx)*(1 - dy) + b*dx*(1 - dy) + c*(1 - dx)*dy + d*dx*dy;
 
+                        float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *y_dst = val;
+                    }
+                }
+            }
+        }
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+        const float a = -0.75f; // use alpha = -0.75 (same as PyTorch)
+        auto weight1 = [a](float x) { return ((a + 2) * x - (a + 3)) * x * x + 1; };
+        auto weight2 = [a](float x) { return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a; };
+        auto bicubic = [=](float p0, float p1, float p2, float p3, float x) {
+            const float w0 = weight2(x + 1);
+            const float w1 = weight1(x + 0);
+            const float w2 = weight1(1 - x);
+            const float w3 = weight2(2 - x);
+            return p0*w0 + p1*w1 + p2*w2 + p3*w3;
+        };
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float)i1 + pixel_offset) / sf1 - pixel_offset;
+                    const int64_t y0 = (int64_t)floorf(y);
+                    const float dy = y - (float)y0;
+
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float)i0 + pixel_offset) / sf0 - pixel_offset;
+                        const int64_t x0 = (int64_t)floorf(x);
+                        const float dx = x - (float)x0;
+
+                        auto p = [=](int64_t x_off, int64_t y_off) -> float {
+                            int64_t i00 = std::max(int64_t(0), std::min(x0 + x_off, ne00 - 1));
+                            int64_t i01 = std::max(int64_t(0), std::min(y0 + y_off, ne01 - 1));
+                            return *(const float *)((const char *)src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                        };
+
+                        const float val = bicubic(
+                            bicubic(p(-1,-1), p(0,-1), p(1,-1), p(2,-1), dx),
+                            bicubic(p(-1, 0), p(0, 0), p(1, 0), p(2, 0), dx),
+                            bicubic(p(-1, 1), p(0, 1), p(1, 1), p(2, 1), dx),
+                            bicubic(p(-1, 2), p(0, 2), p(1, 2), p(2, 2), dx), dy);
+
                         float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
                         *y_dst = val;
                     }
@@ -7847,6 +7793,18 @@ void ggml_compute_forward_timestep_embedding(
 
 // ggml_compute_forward_argsort
 
+template<enum ggml_sort_order order>
+struct cmp_argsort {
+    const float * data;
+    bool operator()(int32_t a, int32_t b) const {
+        if constexpr (order == GGML_SORT_ORDER_ASC) {
+            return data[a] < data[b];
+        } else {
+            return data[a] > data[b];
+        }
+    }
+};
+
 static void ggml_compute_forward_argsort_f32(
     const ggml_compute_params * params,
     ggml_tensor * dst) {
@@ -7865,23 +7823,25 @@ static void ggml_compute_forward_argsort_f32(
     ggml_sort_order order = (ggml_sort_order) ggml_get_op_params_i32(dst, 0);
 
     for (int64_t i = ith; i < nr; i += nth) {
-        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
         const float * src_data = (float *)((char *) src0->data + i*nb01);
 
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+
         for (int64_t j = 0; j < ne0; j++) {
             dst_data[j] = j;
         }
 
-        // C doesn't have a functional sort, so we do a bubble sort instead
-        for (int64_t j = 0; j < ne0; j++) {
-            for (int64_t k = j + 1; k < ne0; k++) {
-                if ((order == GGML_SORT_ORDER_ASC  && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
-                    (order == GGML_SORT_ORDER_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
-                    int32_t tmp = dst_data[j];
-                    dst_data[j] = dst_data[k];
-                    dst_data[k] = tmp;
-                }
-            }
+        switch (order) {
+            case GGML_SORT_ORDER_ASC:
+                std::sort(dst_data, dst_data + ne0, cmp_argsort<GGML_SORT_ORDER_ASC>{src_data});
+                break;
+
+            case GGML_SORT_ORDER_DESC:
+                std::sort(dst_data, dst_data + ne0, cmp_argsort<GGML_SORT_ORDER_DESC>{src_data});
+                break;
+
+            default:
+                GGML_ABORT("invalid sort order");
         }
     }
 }
@@ -7947,12 +7907,78 @@ void ggml_compute_forward_argsort(
     }
 }
 
+// ggml_compute_forward_top_k
+
+struct cmp_top_k {
+    const float * data;
+    bool operator()(int32_t a, int32_t b) const {
+        return data[a] > data[b];
+    }
+};
+
+static void ggml_compute_forward_top_k_f32(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
+
+    const int top_k = ne0;
+
+    int32_t * tmp = (int32_t *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
+
+    for (int64_t i = ith; i < nr; i += nth) {
+        const float * src_data = (float *)((char *) src0->data + i*nb01);
+
+        for (int64_t j = 0; j < ne00; j++) {
+            tmp[j] = j;
+        }
+
+        std::partial_sort(tmp, tmp + top_k, tmp + ne00, cmp_top_k{src_data});
+
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+
+        std::copy(tmp, tmp + top_k, dst_data);
+
+        // emphasize that the order is not important
+        if (top_k > 1) {
+            std::swap(dst_data[0], dst_data[1]);
+        }
+    }
+}
+
+void ggml_compute_forward_top_k(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_top_k_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_flash_attn_ext
 
-static void ggml_compute_forward_flash_attn_ext_f16(
+static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         const ggml_compute_params * params,
-        ggml_tensor * dst) {
-
+        ggml_tensor * dst,
+        int ir0, int ir1) {
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -7968,9 +7994,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
-    const int ith = params->ith;
-    const int nth = params->nth;
-
     const int64_t DK = nek0;
     const int64_t DV = nev0;
     const int64_t N  = neq1;
@@ -8004,16 +8027,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
     // parallelize by q rows using ggml_vec_dot_f32
 
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
     float scale         = 1.0f;
     float max_bias      = 0.0f;
     float logit_softcap = 0.0f;
@@ -8040,6 +8053,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
     GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
+    int ith = params->ith;
+
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
@@ -8187,6 +8202,91 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     }
 }
 
+static void ggml_compute_forward_flash_attn_ext_f16(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
+
+    GGML_ASSERT(ne0 == DV);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
+
+    GGML_ASSERT(neq1 == N);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // parallelize by q rows using ggml_vec_dot_f32
+
+    // total rows in q
+    const int64_t nr = neq1*neq2*neq3;
+
+    // rows per thread
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // disable for NUMA
+    const bool disable_chunking = ggml_is_numa();
+
+    // 4x chunks per thread
+    int nth_scaled = nth * 4;
+    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+    if (nth == 1 || nchunk < nth || disable_chunking) {
+        nchunk = nth;
+    }
+
+    if (ith == 0) {
+        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+        ggml_threadpool_chunk_set(params->threadpool, nth);
+    }
+
+    ggml_barrier(params->threadpool);
+
+    // The number of elements in each chunk
+    const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+    // The first chunk comes from our thread_id, the rest will get auto-assigned.
+    int current_chunk = ith;
+
+    while (current_chunk < nchunk) {
+        const int64_t ir0 = dr * current_chunk;
+        const int64_t ir1 = MIN(ir0 + dr, nr);
+
+        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+
+        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+    }
+}
+
 void ggml_compute_forward_flash_attn_ext(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
@@ -8673,7 +8773,7 @@ static void ggml_compute_forward_ssm_scan_f32(
                 // n_head
                 for (int h = ih0; h < ih1; ++h) {
                     // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
-                    const float dt_soft_plus = ggml_softplus(dt[h]);
+                    const float dt_soft_plus = ggml_compute_softplus_f32(dt[h]);
                     const float dA = expf(dt_soft_plus * A[h]);
                     const int g = h / (nh / ng); // repeat_interleave
 
@@ -8770,7 +8870,7 @@ static void ggml_compute_forward_ssm_scan_f32(
                 // n_head
                 for (int h = ih0; h < ih1; ++h) {
                     // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
-                    const float dt_soft_plus = ggml_softplus(dt[h]);
+                    const float dt_soft_plus = ggml_compute_softplus_f32(dt[h]);
                     const int g = h / (nh / ng); // repeat_interleave
 
                     // dim
@@ -9053,6 +9153,14 @@ void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_xielu(params, dst);
             } break;
+        case GGML_UNARY_OP_EXPM1:
+            {
+                ggml_compute_forward_expm1(params, dst);
+            } break;
+        case GGML_UNARY_OP_SOFTPLUS:
+            {
+                ggml_compute_forward_softplus(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -9649,6 +9757,76 @@ void ggml_compute_forward_gla(
     }
 }
 
+static void ggml_compute_forward_solve_tri_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];  // A (lower triangular)
+    const struct ggml_tensor * src1 = dst->src[1];  // B (RHS)
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ne00 == ne01); // A must be square
+    GGML_ASSERT(ne0  == ne10); // solution cols == B cols
+    GGML_ASSERT(ne1  == ne11); // solution rows == B rows
+
+    GGML_ASSERT(ne02 == ne12 && ne12 == ne2);
+    GGML_ASSERT(ne03 == ne13 && ne13 == ne3);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t k = ne10;   // number of RHS columns
+    const int64_t n = ne11;   // A is n×n
+    const int64_t nr = ne02 * ne03 * k; // we're parallelizing on columns here, so seq x token x column will be the unit
+
+    // chunks per thread
+    const int64_t dr = (nr + nth - 1)/nth;
+
+    // chunk range for this thread
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
+
+    const float * A = (const float *) src0->data;  // [n, n, B1, B2]
+    const float * B = (const float *) src1->data;  // [n, k, B1, B2]
+          float * X = (      float *) dst->data;   // [n, k, B1, B2]
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*k);
+        const int64_t i02 = (ir - i03*ne02*k)/k;
+        const int64_t i01 = (ir - i03*ne02*k - i02*k);
+
+        const float * A_batch = A + i02 * nb02 / sizeof(float) + i03 * nb03 / sizeof(float);
+        const float * B_batch = B + i02 * nb12 / sizeof(float) + i03 * nb13 / sizeof(float);
+
+        float * X_batch = X + i02 * nb2 / sizeof(float) + i03 * nb3 / sizeof(float);
+
+        for (int64_t i00 = 0; i00 < n; ++i00) {
+            float sum = 0.0f;
+            for (int64_t t = 0; t < i00; ++t) {
+                sum += A_batch[i00 * n + t] * X_batch[t * k + i01];
+            }
+
+            const float diag = A_batch[i00 * n + i00];
+            assert(diag != 0.0f && "Zero diagonal in triangular matrix");
+
+            X_batch[i00 * k + i01] = (B_batch[i00 * k + i01] - sum) / diag;
+        }
+    }
+}
+
+void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        ggml_compute_forward_solve_tri_f32(params, dst);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
 // ggml_compute_forward_rwkv_wkv7
 
 static void ggml_compute_forward_rwkv_wkv7_f32(
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/ops.h b/ml/backend/ggml/ggml/src/ggml-cpu/ops.h
index 9824a03b..0fdfee79 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/ops.h
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/ops.h
@@ -34,6 +34,7 @@ void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_sum_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_cumsum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_mean(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argmax(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_count_equal(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -51,10 +52,6 @@ void ggml_compute_forward_scale(const struct ggml_compute_params * params, struc
 void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -84,7 +81,10 @@ void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_top_k(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_fill(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_flash_attn_back(
         const struct ggml_compute_params * params,
@@ -100,6 +100,7 @@ void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params,
 void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/repack.cpp b/ml/backend/ggml/ggml/src/ggml-cpu/repack.cpp
index f531d21e..9f0d449b 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/repack.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/repack.cpp
@@ -124,6 +124,58 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
     }
 }
 
+
+void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK_K == 256);
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK_K];
+    float iscale[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+            float max = 0;
+
+            for (int j = 0; j < QK_K; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK_K + j];
+                // Update the maximum value of the corresponding super block
+                if(amax < fabsf(srcv[row_iter][j])) {
+                    amax = fabsf(srcv[row_iter][j]);
+                    max = srcv[row_iter][j];
+                }
+            }
+
+            iscale[row_iter] = amax ? -127.f/max : 0;
+
+            y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0;
+        }
+
+        for (int j = 0; j < QK_K / 4; j++) {
+            y[i].bsums[j] = 0;
+        }
+
+        // Quants values are interleaved in sequence of four bytes from corresponding super blocks
+        // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving
+        // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
+        for (int j = 0; j < QK_K * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id     = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+            int index = (((j & 15) >> 2) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3);
+
+            float x0 = srcv[src_id][src_offset] * iscale[src_id];
+            y[i].qs[j] = nearest_int(x0);
+            y[i].bsums[index] += y[i].qs[j];
+        }
+    }
+}
+
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK_K == 256);
     assert(k % QK_K == 0);
@@ -192,6 +244,12 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_0>(const float * GGML_RESTR
     ggml_quantize_mat_q8_0_4x8(x, vy, n_per_row);
 }
 
+template <> void ggml_quantize_mat_t<4, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    ggml_quantize_mat_q8_K_4x4(x, vy, n_per_row);
+}
+
 template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) {
     assert(nrow == 4);
     UNUSED(nrow);
@@ -333,6 +391,77 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    float sum_minf[8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2] = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                        sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;
@@ -727,6 +856,89 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 4;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    float sum_minf[4][8];
+    uint32_t utmp[32];
+    int sumi1;
+    int sumi2;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2] = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32;
+                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
+                                sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;
@@ -1228,9 +1440,10 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
 
     GGML_UNUSED(data_size);
 }
+
 static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_K);
-    GGML_ASSERT(interleave_block == 8);
+    GGML_ASSERT(interleave_block == 8 || interleave_block == 4);
     constexpr int nrows_interleaved = 8;
 
     block_q4_Kx8 * dst = (block_q4_Kx8*)t->data;
@@ -1468,6 +1681,10 @@ template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q4_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_K_to_q4_K_8_bl(t, 4, data, data_size);
+}
+
 template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }
@@ -1501,6 +1718,10 @@ template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -1529,6 +1750,10 @@ template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -1600,6 +1825,55 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         return false;
     }
 
+    void forward_mul_mat_one_chunk(ggml_compute_params * params,
+                                   ggml_tensor *         op,
+                                   int64_t               src0_start,
+                                   int64_t               src0_end,
+                                   int64_t               src1_start,
+                                   int64_t               src1_end) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
+
+        GGML_ASSERT(ne03 == 1 && ne13 == 1);
+        GGML_ASSERT(ne12 % ne02 == 0);
+        const int64_t r2 = ne12 / ne02;
+
+        const int64_t i12 = src1_start / ne1;
+        const int64_t i11 = src1_start - i12 * ne1;
+
+        // Determine batch index
+        const int64_t i02 = i12 / r2;
+
+        const int64_t i1 = i11;
+        const int64_t i2 = i12;
+
+        const char * src0_ptr = (const char *) src0->data + i02 * nb02;
+        const char * src1_ptr = (const char *) params->wdata + (i11 + i12 * ne11) * src1_col_stride;
+        char *       dst_ptr  = ((char *) dst->data + (i1 * nb1 + i2 * nb2));
+
+        const int64_t nrows = src1_end - src1_start;
+        const int64_t ncols = src0_end - src0_start;
+
+        GGML_ASSERT(src1_ptr + src1_col_stride * nrows <= (const char *) params->wdata + params->wsize);
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (nrows > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00, (float *) (dst_ptr) + src0_start, nb1 / nb0,
+                                                             src0_ptr + src0_start * nb01, src1_ptr,
+                                                             nrows - (nrows % 4), ncols);
+        }
+        for (int iter = nrows - (nrows % 4); iter < nrows; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00, (float *) (dst_ptr + (iter * nb1)) + src0_start,
+                                                             ne01, src0_ptr + src0_start * nb01,
+                                                             src1_ptr + (src1_col_stride * iter), 1 /* nrows */, ncols);
+        }
+    }
+
     void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
         const ggml_tensor * src0 = op->src[0];
         const ggml_tensor * src1 = op->src[1];
@@ -1621,6 +1895,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         GGML_ASSERT(nb1 <= nb2);
         GGML_ASSERT(nb2 <= nb3);
 
+        // TODO: General batched mul mat for 4D tensors
+        // Currently only supports 3D tensors
+        GGML_ASSERT(ne03 == 1);
+        GGML_ASSERT(ne13 == 1);
+        GGML_ASSERT(ne3 == 1);
+
         GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
         GGML_ASSERT(ggml_n_dims(op->src[0]) == 2);
@@ -1628,46 +1908,102 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 
         char *       wdata = static_cast<char *>(params->wdata);
         const size_t nbw1  = ggml_row_size(PARAM_TYPE, ne10);
+        const size_t nbw2  = nbw1 * ne11;
 
-        assert(params->wsize >= nbw1 * ne11);
+        assert(params->wsize >= nbw2 * ne12);
 
         const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float;
 
-        int64_t i11_processed = 0;
-        for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
-            ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10);
+        // INFO: Quantization is done in planes to avoid extra complexity in chunking.
+        // Flattening dimensions not multiple of INTER_SIZE would require extra handling depending on how
+        // the planes are broadcast.
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            char * data_ptr  = (char *) src1->data + i12 * nb12;
+            char * wdata_ptr = wdata + i12 * nbw2;
+
+            for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
+                ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((float *) (data_ptr + i11 * nb11),
+                                                            (void *) (wdata_ptr + i11 * nbw1), 4, ne10);
+            }
+
+            const int64_t i11_processed = ne11 - ne11 % 4;
+            for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
+                from_float((float *) (data_ptr + i11 * nb11), (void *) (wdata_ptr + i11 * nbw1), ne10);
+            }
         }
 
-        i11_processed = ne11 - ne11 % 4;
-        for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
-            from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+
+        // 4x chunks per thread
+        const int64_t nr0 = ggml_nrows(op->src[0]);
+
+        int     nth_scaled  = nth * 4;
+        int64_t chunk_size0 = (nr0 + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk0     = (nr0 + chunk_size0 - 1) / chunk_size0;
+
+        // src1 is chunked only by full planes.
+        // When we flatten we need to address dimensions not multiple of the q8 INTER_SIZE
+        // to route them thorugh GEMV.
+        // nchunk1 = ne12 also avoids messing the chunking for models with no 3d tensors
+        // to avoid affecting their performance
+        int64_t nchunk1 = ne12;
+
+        // Ensure minimum chunk size to avoid alignment issues with high thread counts
+        // Minimum chunk size should be at least NB_COLS to prevent overlapping chunks after alignment
+        const int64_t min_chunk_size = NB_COLS;
+        if (nchunk0 > 0 && (nr0 / nchunk0) < min_chunk_size && nr0 >= min_chunk_size) {
+            nchunk0 = (nr0 + min_chunk_size - 1) / min_chunk_size;
+        }
+
+        int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
+        // Only increase nchunk0 to nth if it won't make chunks too small
+        if (nth == 1 || ((nchunk0 < nth || disable_chunking) && (nr0 + nth - 1) / nth >= min_chunk_size)) {
+            nchunk0 = nth;
+            dr0 = (nr0 + nchunk0 - 1) / nchunk0;
+        }
+
+        // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size
+        // This prevents creating too many tiny chunks that could overlap after alignment
+        const int64_t max_nchunk = (nr0 + min_chunk_size - 1) / min_chunk_size;
+        nchunk0                  = MIN(nchunk0, max_nchunk);
+
+        if (ith == 0) {
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, nth);
         }
 
         ggml_barrier(params->threadpool);
 
-        const void * src1_wdata      = params->wdata;
-        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
-        int64_t      src0_start      = (ith * ne01) / nth;
-        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
-        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
-        src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
-        if (src0_start >= src0_end) {
-            return;
-        }
+        // The first chunk comes from our thread_id, the rest will get auto-assigned.
+        int current_chunk = ith;
 
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (ne11 > 3) {
-            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
-        }
-        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
-            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                    src0_end - src0_start);
+        while (current_chunk < nchunk0 * nchunk1) {
+            const int64_t ith0 = current_chunk % nchunk0;
+            const int64_t ith1 = current_chunk / nchunk0;
+
+            int64_t src0_start = dr0 * ith0;
+            int64_t src0_end   = MIN(src0_start + dr0, nr0);
+
+            // full-plane range for src1
+            int64_t src1_start = ith1 * ne11;
+            int64_t src1_end = (ith1 + 1) * ne11;
+
+            // Align boundaries to NB_COLS - round up to ensure all data is included
+            // The chunk size limiting above ensures chunks are large enough to prevent overlaps
+            src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+            src0_end   = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end;
+            src0_end   = MIN(src0_end, ne01);
+
+            // Make sure current plane is the last one before exiting
+            if (src0_start >= src0_end) {
+                current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+                continue;
+            }
+
+            forward_mul_mat_one_chunk(params, dst, src0_start, src0_end, src1_start, src1_end);
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
     }
 
@@ -1772,8 +2108,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int64_t src0_cur_start = (ith * ne01) / nth;
             int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
 
+            // Align boundaries to NB_COLS - round up to ensure all data is included
             src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
             src0_cur_end   = (src0_cur_end   % NB_COLS) ? src0_cur_end   + NB_COLS - (src0_cur_end   % NB_COLS) : src0_cur_end;
+            if (src0_cur_end > ne01) {
+                src0_cur_end = ne01;
+            }
 
             if (src0_cur_start >= src0_cur_end) {
                 return;
@@ -1816,6 +2156,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+
+    // instance for Q4_K
+    static const ggml::cpu::repack::tensor_traits<block_q4_K, 4, 8, GGML_TYPE_Q8_K> q4_K_8x4_q8_K;
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
 
     // instance for Q2
@@ -1847,6 +2190,16 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q4_K_8x8_q8_K;
             }
         }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x8_q8_K;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_K_8x4_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_Q2_K) {
         if (ggml_cpu_has_avx512()) {
             if (cur->ne[1] % 8 == 0) {
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/repack.h b/ml/backend/ggml/ggml/src/ggml-cpu/repack.h
index cb32b503..c4d928cd 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/repack.h
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/repack.h
@@ -80,10 +80,12 @@ extern "C" {
 
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -91,6 +93,7 @@ void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -99,10 +102,12 @@ void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -110,6 +115,7 @@ void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/simd-mappings.h b/ml/backend/ggml/ggml/src/ggml-cpu/simd-mappings.h
index 8daec663..101a9c08 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/simd-mappings.h
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/simd-mappings.h
@@ -160,18 +160,18 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F32xt                        svfloat32_t
 #define GGML_F32xt_ZERO                   svdup_n_f32(0.0f)
 #define GGML_F32xt_SET1(x)                svdup_n_f32(x)
-#define GGML_F32xt_LOAD_IMPL(pg, a, ...)  svld1_f32(pg, a)
-#define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
-#define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_LOAD_IMPL(pg, a)       svld1_f32(pg, a)
+#define GGML_F32xt_LOAD(a)                GGML_F32xt_LOAD_IMPL(DEFAULT_PG, a)
+#define GGML_F32xt_STORE_IMPL(pg, a, b)   svst1_f32(pg, a, b)
+#define GGML_F32xt_STORE(a, b)            GGML_F32xt_STORE_IMPL(DEFAULT_PG, a, b)
 #define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, b, c, a)
-#define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_FMA(a, b, c)           GGML_F32xt_FMA_IMPL(DEFAULT_PG, a, b, c)
 #define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
-#define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_ADD(a, b)              GGML_F32xt_ADD_IMPL(DEFAULT_PG, a, b)
 #define GGML_F32xt_MUL_IMPL(pg, a, b)     svmul_f32_m(pg, a, b)
-#define GGML_F32xt_MUL(...)               GGML_F32xt_MUL_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_MUL(a, b)              GGML_F32xt_MUL_IMPL(DEFAULT_PG, a, b)
 #define GGML_F32xt_REDUCE_ONE_IMPL(pg, a) svaddv(pg, a)
-#define GGML_F32xt_REDUCE_ONE(...)        GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE_ONE(a)          GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, a)
 #define GGML_F32xt_REDUCE_IMPL(pg, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)  \
 {                                                      \
     sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum2);        \
@@ -183,7 +183,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum5);        \
     (res) = (ggml_float) GGML_F32xt_REDUCE_ONE(sum1);  \
 }
-#define GGML_F32xt_REDUCE(...) GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE(res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)  \
+        GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)
 
 #define GGML_F32_VEC        GGML_F32xt
 #define GGML_F32_VEC_ZERO   GGML_F32xt_ZERO
@@ -206,11 +207,11 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F32Cxt_STORE(dst_ptr, src_vec) svst1_f16(DEFAULT_PG16, (__fp16 *)(dst_ptr), (src_vec))
 
 #define GGML_F32Cxt_FMA_IMPL(pg, a, b, c)   svmad_f16_x(pg, b, c, a)
-#define GGML_F32Cxt_FMA(...)                GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_FMA(a, b, c)            GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, a, b, c)
 #define GGML_F32Cxt_ADD_IMPL(pg, a, b)      svadd_f16_x(pg, a, b)
-#define GGML_F32Cxt_ADD(...)                GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_ADD(a, b)               GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, a, b)
 #define GGML_F32Cxt_MUL_IMPL(pg, a, b)      svmul_f16_x(pg, a, b)
-#define GGML_F32Cxt_MUL(...)                GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_MUL(a, b)               GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, a, b)
 #define GGML_F32Cxt_REDUCE                  GGML_F16xt_REDUCE_MIXED
 
 #define GGML_F16x_VEC                GGML_F32Cxt
@@ -224,7 +225,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F16x_VEC_REDUCE         GGML_F32Cxt_REDUCE
 
 #define GGML_F16xt_REDUCE_ONE_IMPL(pg, a) svaddv_f16(pg, a)
-#define GGML_F16xt_REDUCE_ONE(...)        GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F16xt_REDUCE_ONE(a)          GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, a)
 
 #define GGML_F16xt_REDUCE_MIXED_IMPL(pg16, res, sum1, sum2, sum3, sum4)  \
 {                                                      \
@@ -234,7 +235,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     __fp16 sum_f16 = svaddv_f16(pg16, sum1);           \
     (res) = (ggml_float) sum_f16;                      \
 }
-#define GGML_F16xt_REDUCE_MIXED(...) GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F16xt_REDUCE_MIXED(res, sum1, sum2, sum3, sum4)  \
+        GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, res, sum1, sum2, sum3, sum4)
 
 // F16 NEON
 
@@ -956,7 +958,7 @@ do {                                                              \
 
 #define GGML_F32Cx8          __m256
 #define GGML_F32Cx8_ZERO    (__m256)__lasx_xvldi(0)
-#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
+#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
 
 static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
     __m256i a;
@@ -999,34 +1001,34 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 
 #define GGML_F32x4         __m128
 #define GGML_F32x4_ZERO    (__m128)__lsx_vldi(0)
-#define GGML_F32x4_SET1(x) (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32x4_SET1(x) (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32x4_LOAD(x) (__m128)__lsx_vld((x), 0)
 #define GGML_F32x4_STORE(x, y)   __lsx_vst(y, x, 0)
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
-#define GGML_F32x4_REDUCE(res, x)                                                     \
-{                                                                                     \
-    int offset = GGML_F32_ARR >> 1;                                                   \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    __m128i tmp     = __lsx_vsrli_d((__m128i) x[0], 32);                              \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]);                    \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    const __m128 t0 = (__m128)__lsx_vshuf4i_w(tmp, 0x88);                                     \
-    tmp             = __lsx_vsrli_d((__m128i) t0, 32);                                \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, t0);                      \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    res             = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \
+
+#define GGML_F32x4_REDUCE(res, x)                               \
+{                                                               \
+    int offset = GGML_F32_ARR >> 1;                             \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    __m128i t0 = __lsx_vpickev_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128i t1 = __lsx_vpickod_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128 t2 = __lsx_vfadd_s((__m128)t0, (__m128)t1);          \
+    __m128i t3 = __lsx_vpickev_w((__m128i)t2, (__m128i)t2);     \
+    __m128i t4 = __lsx_vpickod_w((__m128i)t2, (__m128i)t2);     \
+    __m128 t5 = __lsx_vfadd_s((__m128)t3, (__m128)t4);          \
+    res = (ggml_float) ((v4f32)t5)[0];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -1068,7 +1070,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
 #define GGML_F32Cx4             __m128
 #define GGML_F32Cx4_ZERO        (__m128)__lsx_vldi(0)
-#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32Cx4_LOAD(x)     (__m128)__lsx_f16x4_load(x)
 #define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
 #define GGML_F32Cx4_FMA         GGML_F32x4_FMA
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.cpp b/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.cpp
index a047537b..1d9873ad 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.cpp
@@ -73,6 +73,14 @@ static inline float op_log(float x) {
     return logf(x);
 }
 
+static inline float op_expm1(float x) {
+    return expf(x) - 1.0f;
+}
+
+static inline float op_softplus(float x) {
+    return (x > 20.0f) ? x : logf(1.0f + expf(x));
+}
+
 static inline float op_floor(float x) {
     return floorf(x);
 }
@@ -290,6 +298,14 @@ void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor *
     unary_op<op_log>(params, dst);
 }
 
+void ggml_compute_forward_expm1(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_expm1>(params, dst);
+}
+
+void ggml_compute_forward_softplus(const ggml_compute_params * params, ggml_tensor * dst) {
+    unary_op<op_softplus>(params, dst);
+}
+
 void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
     unary_op<op_floor>(params, dst);
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.h b/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.h
index fa45d9f0..bcad5a3a 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.h
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/unary-ops.h
@@ -22,6 +22,8 @@ void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct
 void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_expm1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_softplus(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/vec.cpp b/ml/backend/ggml/ggml/src/ggml-cpu/vec.cpp
index 43dc7537..ac8633e2 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/vec.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/vec.cpp
@@ -360,6 +360,13 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
     for (; i + 3 < n; i += 4) {
         vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
     }
+#elif defined(__riscv_v_intrinsic)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+        vfloat32m2_t vy = ggml_v_silu_m2(vx, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
 #endif
     for (; i < n; ++i) {
         y[i] = ggml_silu_f32(x[i]);
@@ -460,6 +467,16 @@ ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const floa
         val = vec_mul(val, val);
         sum += (ggml_float)vec_hsum_f32x4(val);
     }
+#elif defined(__riscv_v_intrinsic)
+    vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t val = __riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], vl), mean, vl);
+        __riscv_vse32_v_f32m2(&y[i], val, vl);
+        val = __riscv_vfmul_vv_f32m2(val, val, vl);
+        vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, vl);
+    }
+    sum = (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
 #endif
     for (; i < n; ++i) {
         float val = x[i] - mean;
diff --git a/ml/backend/ggml/ggml/src/ggml-cpu/vec.h b/ml/backend/ggml/ggml/src/ggml-cpu/vec.h
index 65c7dfb6..bd80805f 100644
--- a/ml/backend/ggml/ggml/src/ggml-cpu/vec.h
+++ b/ml/backend/ggml/ggml/src/ggml-cpu/vec.h
@@ -397,119 +397,118 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 }
 
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16;
-        const int ggml_f16_step = 8 * ggml_f16_epr;
+#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
+    const int sve_register_length = svcntb() * 8;
+    const int ggml_f16_epr = sve_register_length / 16;
+    const int ggml_f16_step = 8 * ggml_f16_epr;
 
-        GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
+    GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
 
-        const int np= (n & ~(ggml_f16_step - 1));
+    int np = (n & ~(ggml_f16_step - 1));
 
-        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
-        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
+    svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+    svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+    for (int i = 0; i < np; i += ggml_f16_step) {
+        ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
+        GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
 
-            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
+        ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
+        GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
 
-            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
-            ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
+        ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
+        ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
+        ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
+        GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
 
-            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
-            ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
+        ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
+        ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
+        ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
+        GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
 
-            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
-            ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
+        ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
+        ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
+        ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
+        GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
 
-            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
-            ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
+        ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
+        ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
+        ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
+        GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
 
-            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
-            ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
+        ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
+        ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
+        ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
+        GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
 
-            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
-            ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
+        ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
+        ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
+        ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
 
-            GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+        GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+    }
+    const int np2 = (n & ~(ggml_f16_epr - 1));
+    for (int k = np; k < np2; k += ggml_f16_epr) {
+        svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
+        svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
+        ry = GGML_F16x_VEC_FMA(ry, rx, vx);
+
+        GGML_F16x_VEC_STORE(y + k, ry, 0);
+    }
+
+    if (np2 < n) {
+        svbool_t pg = svwhilelt_b16(np2, n);
+        svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
+        svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
+        hy = svmad_f16_x(pg, hx, vx, hy);
+        svst1_f16(pg, (__fp16 *)(y + np2), hy);
+    }
+    np = n;
+#elif defined(__riscv_zvfh) // implies __riscv_v_intrinsic
+    const int np = n;
+    _Float16 hv = (_Float16)v;
+    for (int i = 0, avl; i < n; i += avl) {
+        avl = __riscv_vsetvl_e16m8(n - i);
+        vfloat16m8_t ax = __riscv_vle16_v_f16m8((const _Float16 *)&x[i], avl);
+        vfloat16m8_t ay = __riscv_vle16_v_f16m8((_Float16 *)&y[i], avl);
+        vfloat16m8_t ny = __riscv_vfmadd_vf_f16m8(ax, hv, ay, avl);
+        __riscv_vse16_v_f16m8((_Float16 *)&y[i], ny, avl);
+    }
+#elif defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
-        const int np2 = (n & ~(ggml_f16_epr - 1));
-        for (int k = np; k < np2; k += ggml_f16_epr) {
-            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
-            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
-            ry = GGML_F16x_VEC_FMA(ry, rx, vx);
-
-            GGML_F16x_VEC_STORE(y + k, ry, 0);
-        }
-
-        if (np2 < n) {
-            svbool_t pg = svwhilelt_b16(np2, n);
-            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
-            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
-            hy = svmad_f16_x(pg, hx, vx, hy);
-            svst1_f16(pg, (__fp16 *)(y + np2), hy);
-        }
-
-    #elif defined(__riscv_v_intrinsic)
-        // todo: RVV impl
-        // scalar
-        for (int i = 0; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
-
-        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-        GGML_F16_VEC ax[GGML_F16_ARR];
-        GGML_F16_VEC ay[GGML_F16_ARR];
-
-        for (int i = 0; i < np; i += GGML_F16_STEP) {
-            for (int j = 0; j < GGML_F16_ARR; j++) {
-                ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
-
-                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
-            }
-        }
-
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-        }
-    #endif
+    }
 #else
-    // scalar
-    for (int i = 0; i < n; ++i) {
+    const int np = 0;
+#endif
+
+    // leftovers
+    for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
-#endif
 }
 
 // xs and vs are byte strides of x and v
@@ -698,60 +697,61 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 }
 
 inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
-#if defined(GGML_SIMD)
-    #if defined(__ARM_FEATURE_SVE)
-        const int sve_register_length = svcntb() * 8;
-        const int ggml_f16_epr = sve_register_length / 16;
-        const int ggml_f16_step = 2 * ggml_f16_epr;
+#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
+    const int sve_register_length = svcntb() * 8;
+    const int ggml_f16_epr = sve_register_length / 16;
+    const int ggml_f16_step = 2 * ggml_f16_epr;
 
-        GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
-        const int np = (n & ~(ggml_f16_step - 1));
-        svfloat16_t ay1, ay2;
+    GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
+    const int np = (n & ~(ggml_f16_step - 1));
+    svfloat16_t ay1, ay2;
 
-        for (int i = 0; i < np; i += ggml_f16_step) {
-            ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
-            ay1 = GGML_F16x_VEC_MUL(ay1, vx);
-            GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
+    for (int i = 0; i < np; i += ggml_f16_step) {
+        ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
+        ay1 = GGML_F16x_VEC_MUL(ay1, vx);
+        GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
 
-            ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
-            ay2 = GGML_F16x_VEC_MUL(ay2, vx);
-            GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
+        ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
+        ay2 = GGML_F16x_VEC_MUL(ay2, vx);
+        GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
+    }
+    // leftovers
+    // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
+    if (np < n) {
+        svbool_t pg = svwhilelt_b16(np, n);
+        svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
+        svfloat16_t out = svmul_f16_m(pg, hy, vx);
+        svst1_f16(pg, (__fp16 *)(y + np), out);
+    }
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
+    for (int i = 0, vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m2(n - i);
+        vfloat16m2_t vy = __riscv_vle16_v_f16m2((_Float16 *)&y[i], vl);
+        vfloat32m4_t vy32 = __riscv_vfwcvt_f_f_v_f32m4(vy, vl);
+        vy32 = __riscv_vfmul_vf_f32m4(vy32, v, vl);
+        vy = __riscv_vfncvt_f_f_w_f16m2(vy32, vl);
+        __riscv_vse16_v_f16m2((_Float16 *)&y[i], vy, vl);
+    }
+#elif defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
-        // leftovers
-        // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
-        if (np < n) {
-            svbool_t pg = svwhilelt_b16(np, n);
-            svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
-            svfloat16_t out = svmul_f16_m(pg, hy, vx);
-            svst1_f16(pg, (__fp16 *)(y + np), out);
-        }
-    #elif defined(__riscv_v_intrinsic)
-        // todo: RVV impl
-        // scalar
-        for (int i = 0; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-        }
-    #else
-        const int np = (n & ~(GGML_F16_STEP - 1));
+    }
 
-        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
-
-        GGML_F16_VEC ay[GGML_F16_ARR];
-
-        for (int i = 0; i < np; i += GGML_F16_STEP) {
-            for (int j = 0; j < GGML_F16_ARR; j++) {
-                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-                ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
-
-                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
-            }
-        }
-
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-        }
-    #endif
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+    }
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -1416,6 +1416,16 @@ inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #endif
 }
 
+inline static void ggml_vec_cumsum_f32(const int n, float * y, const float * x) {
+    for (int i = 0; i < n; ++i) {
+        if (i == 0) {
+            y[i] = x[i];
+        } else {
+            y[i] = y[i - 1] + x[i];
+        }
+    }
+}
+
 inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float * x) {
     ggml_float sum = 0.0;
     for (int i = 0; i < n; ++i) {
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt b/ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt
index 30247751..67af1d8c 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt
@@ -124,6 +124,7 @@ if (CUDAToolkit_FOUND)
 
     if (GGML_CUDA_DEBUG)
         list(APPEND CUDA_FLAGS -lineinfo)
+        add_compile_definitions(GGML_CUDA_DEBUG)
     endif()
 
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/argsort.cu b/ml/backend/ggml/ggml/src/ggml-cuda/argsort.cu
index 08dd3052..b82be371 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/argsort.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/argsort.cu
@@ -44,7 +44,7 @@ static void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
     const dim3 offset_grid((nrows + block_size - 1) / block_size);
     init_offsets<<<offset_grid, block_size, 0, stream>>>(d_offsets, ncols, nrows);
 
-    cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream);
+    CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));
 
     size_t temp_storage_bytes = 0;
 
@@ -87,7 +87,7 @@ template<ggml_sort_order order>
 static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
     // bitonic sort
     int col = threadIdx.x;
-    int row = blockIdx.y;
+    int row = blockIdx.x;
 
     if (col >= ncols_pad) {
         return;
@@ -151,7 +151,7 @@ static void argsort_f32_i32_cuda_bitonic(const float *   x,
     const int ncols_pad = next_power_of_2(ncols);
 
     const dim3 block_dims(ncols_pad, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
+    const dim3 block_nums(nrows, 1, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     // FIXME: this limit could be raised by ~2-4x on Ampere or newer
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh
index 2931c15c..c3f8ca91 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/common.cuh
@@ -21,10 +21,12 @@
 #include "ggml-common.h"
 
 #include <array>
+#include <algorithm>
 #include <cassert>
 #include <cfloat>
 #include <cstdio>
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 #if defined(GGML_USE_HIP)
@@ -119,12 +121,12 @@ static cudaError_t cudaMemsetAsyncReserve ( void* devPtr, int value, size_t coun
 
 #define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
 #define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_PH1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // MTT S5000
 
 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
-#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
-#define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_PH1)
+#define GGML_CUDA_CC_IS_PH1(cc)      (cc >= GGML_CUDA_CC_PH1)
 
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
 #    define GGML_CUDA_USE_CUB
@@ -247,9 +249,9 @@ static const char * cu_get_error_str(CUresult err) {
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 
-#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
@@ -259,6 +261,15 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
+#if defined(GGML_USE_HIP) && defined(RDNA4)
+#define AMD_WMMA_AVAILABLE
+#endif // defined(GGML_USE_HIP) && defined(RDNA4)
+
+// The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#define VOLTA_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define TURING_MMA_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@@ -276,12 +287,14 @@ static const char * cu_get_error_str(CUresult err) {
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
 static bool fp16_available(const int cc) {
-    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
+    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
 }
 
 static bool fast_fp16_available(const int cc) {
     return GGML_CUDA_CC_IS_AMD(cc) ||
-        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
+        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && fp16_available(cc));
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -298,7 +311,9 @@ static bool fp16_mma_hardware_available(const int cc) {
 }
 
 static bool bf16_mma_hardware_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3 ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
 }
 
 static bool fp32_mma_hardware_available(const int cc) {
@@ -313,7 +328,14 @@ static bool amd_mfma_available(const int cc) {
 #endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
-// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
+static bool amd_wmma_available(const int cc) {
+    return GGML_CUDA_CC_IS_RDNA4(cc);
+}
+
+static bool volta_mma_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
+}
+
 static bool turing_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }
@@ -577,8 +599,12 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
     acc += v.y*u.y;
 }
 
-static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
 #if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+#define V_DOT2_F32_F16_AVAILABLE
+#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
+#ifdef V_DOT2_F32_F16_AVAILABLE
     asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
 #else
 #ifdef FAST_FP16_AVAILABLE
@@ -590,7 +616,7 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v,
     acc += tmpv.x * tmpu.x;
     acc += tmpv.y * tmpu.y;
 #endif // FAST_FP16_AVAILABLE
-#endif // defined(GGML_USE_HIP) && (defined(RDNA2)  || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
+#endif // V_DOT2_F32_F16_AVAILABLE
 }
 
 static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {
@@ -613,6 +639,12 @@ static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v,
 //     If dst and src point at different address spaces then they are guaranteed to not be aliased.
 template <int nbytes, int alignment = 0>
 static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
+    static_assert(
+        nbytes <= ggml_cuda_get_max_cpy_bytes() || alignment == 0,
+        "You are misusing the alignment parameter for ggml_cuda_memcpy_1. "
+        "The intent is for the parameter is only as a workaround if either one of the pointers is not properly aligned. "
+        "If you use it to do more bytes per copy than ggml_cuda_max_cpy_bytes() the reads and writes may not be coalesced. "
+        "Call ggml_cuda_memcpy_1 in a loop instead.");
     if constexpr (alignment != 0) {
         static_assert(nbytes % alignment == 0, "bad alignment");
     }
@@ -660,8 +692,11 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 // and a shift:
 //
 // n/d = (mulhi(n, mp) + n) >> L;
-static const uint3 init_fastdiv_values(uint32_t d) {
-    GGML_ASSERT(d != 0);
+static const uint3 init_fastdiv_values(uint64_t d_64) {
+    GGML_ASSERT(d_64 != 0);
+    GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());
+
+    uint32_t d = (uint32_t)d_64;
 
     // compute L = ceil(log2(d));
     uint32_t L = 0;
@@ -985,6 +1020,154 @@ struct ggml_cuda_graph {
 #endif
 };
 
+struct ggml_cuda_concurrent_event {
+    std::vector<cudaEvent_t> join_events;
+    cudaEvent_t              fork_event = nullptr;
+
+    int                                          n_streams = 0;
+    std::unordered_map<const ggml_tensor *, int> stream_mapping;
+
+    const ggml_tensor * join_node;
+
+    ggml_cuda_concurrent_event() = default;
+
+    ggml_cuda_concurrent_event(const ggml_cuda_concurrent_event &) = delete;
+    ggml_cuda_concurrent_event & operator=(const ggml_cuda_concurrent_event &) = delete;
+
+    explicit ggml_cuda_concurrent_event(int n_streams) : n_streams(n_streams) {
+        join_events.resize(n_streams);
+
+        for (size_t i = 0; i < join_events.size(); ++i) {
+            CUDA_CHECK(cudaEventCreateWithFlags(&join_events[i], cudaEventDisableTiming));
+        }
+
+        CUDA_CHECK(cudaEventCreateWithFlags(&fork_event, cudaEventDisableTiming));
+    }
+
+    ggml_cuda_concurrent_event(ggml_cuda_concurrent_event && other) noexcept
+    : join_events(std::move(other.join_events))
+    , fork_event(other.fork_event)
+    , n_streams(other.n_streams)
+    , stream_mapping(std::move(other.stream_mapping))
+    , join_node(other.join_node) {
+        other.fork_event = nullptr;
+    }
+
+    // 1. check if any branches write to overlapping memory ranges (except the join node)
+    // 2. check whether all srcs are either within the branch or outside the nodes covered by ggml_cuda_concurrent_event
+    // we assume all nodes have the same buffer
+    bool is_valid() const {
+        std::vector<std::vector<std::pair<int64_t, int64_t>>> write_ranges;
+        write_ranges.resize(n_streams);
+
+        // get join_node's memory range to exclude from overlap checking.
+        // multiple nodes can use join_node's buffer; we synchronize on the join node.
+        const ggml_tensor * join_t     = join_node->view_src ? join_node->view_src : join_node;
+        const int64_t       join_start = (int64_t) join_t->data;
+        const int64_t       join_end   = join_start + ggml_nbytes(join_t);
+
+        for (const auto & [tensor, stream] : stream_mapping) {
+            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
+            const int64_t       t_start = (int64_t) t->data;
+            const int64_t       t_end   = t_start + ggml_nbytes(t);
+
+            // skip tensors that overlap with join_node's buffer.
+            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
+                continue;
+            }
+
+            // concurrent streams begin from 1
+            write_ranges[stream - 1].emplace_back(t_start, t_end);
+        }
+
+        for (int i = 0; i < n_streams; ++i) {
+            // sorts first by start then by end of write range
+            std::sort(write_ranges[i].begin(), write_ranges[i].end());
+        }
+
+        bool writes_overlap = false;
+        bool dependent_srcs = false;
+        for (const auto & [tensor, stream] : stream_mapping) {
+            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
+            const int64_t       t_start = (int64_t) t->data;
+            const int64_t       t_end   = t_start + ggml_nbytes(t);
+
+            // skip tensors that overlap with join_node's buffer
+            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
+                continue;
+            }
+
+            // check if this buffer's write data overlaps with another stream's
+            std::pair<int64_t, int64_t> data_range = std::make_pair(t_start, t_end);
+            for (int i = 0; i < n_streams; ++i) {
+                if (i == stream - 1) {
+                    continue;
+                }
+                auto it = std::lower_bound(write_ranges[i].begin(), write_ranges[i].end(), data_range);
+
+                if (it != write_ranges[i].end()) {
+                    const std::pair<int64_t, int64_t> & other = *it;
+
+                    // std::lower_bound returns the first element where other >= data_range (lexicographically).
+                    // This guarantees other.first >= data_range.first.
+                    // Therefore, overlap occurs iff other.first < data_range.second
+                    // (i.e., the other range starts before this range ends).
+                    if (other.first < data_range.second) {
+                        GGML_LOG_DEBUG("Writes overlap for %s", tensor->name);
+                        writes_overlap = true;
+                        break;
+                    }
+                }
+            }
+
+            //check if all srcs are either in branch or don't have a branch
+            for (int i = 0; i < GGML_MAX_SRC; ++i) {
+                if (!tensor->src[i]) {
+                    continue;
+                }
+
+                auto it = stream_mapping.find(tensor->src[i]);
+
+                if (it == stream_mapping.end()) {
+                    continue;
+                }
+
+                if (it->second != stream) {
+                    dependent_srcs = true;
+                    break;
+                }
+            }
+
+            if (dependent_srcs || writes_overlap) {
+                break;
+            }
+        }
+
+        return !writes_overlap && !dependent_srcs;
+    }
+
+    ~ggml_cuda_concurrent_event() {
+        if (fork_event != nullptr) {
+            CUDA_CHECK(cudaEventDestroy(fork_event));
+        }
+        for (cudaEvent_t e : join_events) {
+            if (e != nullptr) {
+                CUDA_CHECK(cudaEventDestroy(e));
+            }
+        }
+    }
+};
+
+struct ggml_cuda_stream_context {
+    std::vector<const ggml_tensor *>                                    original_nodes;
+    std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
+
+    void reset() {
+        original_nodes.clear();
+        concurrent_events.clear();
+    }
+};
+
 struct ggml_backend_cuda_context {
     int device;
     std::string name;
@@ -995,11 +1178,15 @@ struct ggml_backend_cuda_context {
 
     std::unique_ptr<ggml_cuda_graph> cuda_graph;
 
+    int curr_stream_no = 0;
+
     explicit ggml_backend_cuda_context(int device) :
         device(device),
         name(GGML_CUDA_NAME + std::to_string(device)) {
     }
 
+    ggml_cuda_stream_context concurrent_stream_context;
+
     ~ggml_backend_cuda_context();
 
     cudaStream_t stream(int device, int stream) {
@@ -1010,9 +1197,9 @@ struct ggml_backend_cuda_context {
         return streams[device][stream];
     }
 
-    cudaStream_t stream() {
-        return stream(device, 0);
-    }
+    cudaStream_t stream() { return stream(device, curr_stream_no); }
+
+    ggml_cuda_stream_context & stream_context() { return concurrent_stream_context; }
 
     cublasHandle_t cublas_handle(int device) {
         if (cublas_handles[device] == nullptr) {
@@ -1028,15 +1215,15 @@ struct ggml_backend_cuda_context {
     }
 
     // pool
-    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
+    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];
 
-    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, bool alloc);
+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, int stream_no, bool alloc);
 
     ggml_cuda_pool & pool(int device) {
-        if (pools[device] == nullptr) {
-            pools[device] = new_pool_for_device(device, true);
+        if (pools[device][curr_stream_no] == nullptr) {
+            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no, true);
         }
-        return *pools[device];
+        return *pools[device][curr_stream_no];
     }
 
     ggml_cuda_pool & pool() {
@@ -1044,18 +1231,31 @@ struct ggml_backend_cuda_context {
     }
 
     void pool_set_alloc(bool alloc) {
-        GGML_ASSERT(pools[device] == nullptr || pools[device]->alloc_memory() == alloc);
+        GGML_ASSERT(pools[device][curr_stream_no] == nullptr || pools[device][curr_stream_no]->alloc_memory() == alloc);
 
-        if (pools[device] == nullptr) {
-            pools[device] = new_pool_for_device(device, alloc);
+        if (pools[device][curr_stream_no] == nullptr) {
+            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no, alloc);
         }
     }
 
     size_t pool_get_alloc_size() {
-        if (pools[device] == nullptr) {
+        if (pools[device][curr_stream_no] == nullptr) {
             return 0;
         }
 
-        return pools[device]->alloc_size();
+        return pools[device][curr_stream_no]->alloc_size();
     }
 };
+
+struct ggml_cuda_mm_fusion_args_host {
+    const ggml_tensor * x_bias = nullptr;
+    const ggml_tensor * gate = nullptr;
+    const ggml_tensor * gate_bias = nullptr;
+    ggml_glu_op glu_op;
+};
+struct ggml_cuda_mm_fusion_args_device {
+    const void * x_bias = nullptr;
+    const void * gate = nullptr;
+    const void * gate_bias = nullptr;
+    ggml_glu_op glu_op;
+};
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/convert.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/convert.cuh
index ef9e1299..09f9a33f 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/convert.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/convert.cuh
@@ -1,3 +1,4 @@
+#pragma once
 #include "common.cuh"
 
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
@@ -38,6 +39,15 @@ template<typename dst_t, typename src_t>
         return __float2bfloat16(float(x));
     } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
         return __bfloat162float(x);
+    } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, half2>) {
+        return __float22half2_rn(x);
+    } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, nv_bfloat162>) {
+        // bypass compile error on cuda 12.0.1
+#ifdef GGML_USE_HIP
+        return __float22bfloat162_rn(x);
+#else
+        return {x.x, x.y};
+#endif // GGML_USE_HIP
     } else if constexpr(std::is_same_v<dst_t, int32_t>) {
         return int32_t(x);
     } else {
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/cpy-utils.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/cpy-utils.cuh
index 597c0c8b..00d773dd 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/cpy-utils.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/cpy-utils.cuh
@@ -212,7 +212,7 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
 }
 
 template<typename src_t, typename dst_t>
-static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
+static __device__ void cpy_1_scalar(const char * cxi, char * cdsti) {
     *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
 }
 
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/cpy.cu b/ml/backend/ggml/ggml/src/ggml-cuda/cpy.cu
index a0e34030..0e53ecc3 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/cpy.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/cpy.cu
@@ -7,11 +7,15 @@
 
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
+const int CUDA_CPY_TILE_DIM_2D = 32; // 2D tile dimension for transposed blocks
+const int CUDA_CPY_BLOCK_NM = 8;     // block size of 3rd dimension if available
+const int CUDA_CPY_BLOCK_ROWS = 8;   // block dimension for marching through rows
+
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
-                               const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-                               const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                               const int nb12, const int nb13) {
+static __global__ void cpy_scalar(const char * cx, char * cdst, const int ne,
+                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -35,6 +39,58 @@ static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
+template <typename T>
+static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const int ne,
+                               const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                               const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                               const int nb12, const int nb13) {
+
+    const T* src = reinterpret_cast<const T*>(cx);
+    T* dst = reinterpret_cast<T*>(cdst);
+
+    const int64_t nmat = ne / (ne00 * ne01);
+    const int64_t n = ne00 * ne01;
+
+    const int x = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.x;
+    const int y = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+    const int tx = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
+    const int ty = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+
+    __shared__ float tile[CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
+
+#pragma unroll
+    for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
+
+        const unsigned int imat = blockIdx.z * CUDA_CPY_BLOCK_NM + i;
+        if (imat >= nmat)
+            break;
+
+#pragma unroll
+        for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
+            if(x < ne01 && y + j < ne00){
+                const int row = threadIdx.y+j;
+                const int col = threadIdx.x * sizeof(float)/sizeof(T);
+                T *tile2 = reinterpret_cast<T*>(tile[row]);
+                tile2[col] = src[imat*n + (y+j)*ne01 + x];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
+            if (ty + j < ne01 && tx < ne00) {
+                const int col = (threadIdx.y+j)*sizeof(float)/sizeof(T);
+                const T *tile2 = reinterpret_cast<const T*>(tile[threadIdx.x]);
+                dst[imat*n + (ty+j)*ne00 + tx] = tile2[col];
+            }
+        }
+    }
+
+    GGML_UNUSED_VARS(ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+        nb12, nb13);
+}
+
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
 
@@ -113,14 +169,59 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
 }
 
 template<typename src_t, typename dst_t>
-static void ggml_cpy_flt_cuda(
+static __global__ void cpy_scalar_contiguous(const char * cx, char * cdst, const int64_t ne) {
+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const src_t * x = (const src_t *) cx;
+    dst_t *     dst = (dst_t *) cdst;
+
+    dst[i] = ggml_cuda_cast<dst_t>(x[i]);
+}
+
+template<typename src_t, typename dst_t>
+static void ggml_cpy_scalar_contiguous_cuda(
+    const char * cx, char * cdst, const int64_t ne,
+cudaStream_t stream) {
+
+    const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_scalar_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne);
+}
+
+template<typename src_t, typename dst_t, bool transposed = false>
+static void ggml_cpy_scalar_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    if (transposed) {
+        GGML_ASSERT(ne == ne00*ne01*ne02);  // ne[3] is 1 assumed
+        int ne00n, ne01n, ne02n;
+        if (nb00 <= nb02) { // most likely safe to handle nb00 = nb02 case here
+            ne00n = ne00;
+            ne01n = ne01;
+            ne02n = ne02;
+        } else {
+            ne00n = ne00;
+            ne01n = ne01*ne02;
+            ne02n = 1;
+        }
+
+        dim3 dimGrid( (ne01n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D,
+                      (ne00n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D,
+                      (ne/(ne01n*ne00n) + CUDA_CPY_BLOCK_NM - 1) / CUDA_CPY_BLOCK_NM);
+        dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
+        cpy_scalar_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
+            (cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    } else {
+        const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+        cpy_scalar<cpy_1_scalar<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+            (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    }
 }
 
 static void ggml_cpy_f32_q8_0_cuda(
@@ -322,7 +423,11 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
 
-    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+    const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
+    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) &&
+        src0->ne[3] == 1 && nb02 == ne00 * ne01 * (int64_t)ggml_element_size(src0);
+
+    if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
 #if defined(GGML_USE_MUSA) && defined(GGML_MUSA_MUDNN_COPY)
         if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) {
@@ -333,55 +438,137 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
             CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (can_be_transposed) {
+            ggml_cpy_scalar_cuda<float, float, true>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<float, float>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<float, nv_bfloat16>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<float, nv_bfloat16>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<float, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<float, half>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<float, half>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q8_0_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q8_0_f32_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_0_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q4_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q4_0_f32_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_1_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q4_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q4_1_f32_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_0_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q5_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q5_0_f32_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_iq4_nl_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_1_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q5_1_f32_cuda
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (can_be_transposed) {
+            ggml_cpy_scalar_cuda<half, half, true>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<half, half>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<half, nv_bfloat16>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<half, nv_bfloat16>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<half, float>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<half, float>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
         // TODO consider converting to template
         ggml_cpy_i32_i32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (can_be_transposed) {
+            ggml_cpy_scalar_cuda<nv_bfloat16, nv_bfloat16, true>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<nv_bfloat16, nv_bfloat16>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<nv_bfloat16, half>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<nv_bfloat16, half>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<nv_bfloat16, float>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<nv_bfloat16, float>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
+        if (can_be_transposed) {
+            ggml_cpy_scalar_cuda<int32_t, int32_t, true>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<int32_t, int32_t>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
-        ggml_cpy_flt_cuda<float, int32_t> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<float, int32_t>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<float, int32_t>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<int32_t, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<int32_t, float>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<int32_t, float>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-common.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-common.cuh
index 218ccff1..5cdd4bb2 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-common.cuh
@@ -55,11 +55,11 @@ static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_f16(
         ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, K_h2 + k_KQ_0 + (threadIdx.x % nthreads)*cpy_ne);
 #pragma unroll
         for (int k_KQ_1 = 0; k_KQ_1 < cpy_ne; ++k_KQ_1) {
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             ggml_cuda_mad(sum,                tmp[k_KQ_1] , ((const half2  *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
 #else
             ggml_cuda_mad(sum, __half22float2(tmp[k_KQ_1]), ((const float2 *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
-#endif // FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cu b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cu
index 3a5806d9..3fcb09b7 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cu
@@ -14,6 +14,10 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case< 64,  64>(ctx, dst);
         } break;
+        case  72: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 72,  72>(ctx, dst);
+        } break;
         case  80: {
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case< 80,  80>(ctx, dst);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cuh
index 2b60b3bb..3e58d64f 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cuh
@@ -6,7 +6,7 @@
 // nbatch_K == number of K columns to load in parallel for KQ calculation
 
 // TODO optimize kernel parameters for FP16 NVIDIA (P100)
-// TODO optimize kernel parameters for head sizes 40, 80, 96, 112
+// TODO optimize kernel parameters for head sizes 40, 72, 80, 96, 112
 
 // The ROCm compiler cannot handle templating in __launch_bounds__.
 // As a workaround, define a macro to package the kernel parameters as uint32_t:
@@ -32,6 +32,12 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  64,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  64,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  64,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  64,  40)
@@ -80,6 +86,12 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 128, 3,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -130,6 +142,13 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -185,6 +204,13 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 128, 4,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 128, 5,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -583,7 +609,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
             float KQ_sum_add = 0.0f;
 #pragma unroll
             for (int i0 = 0; i0 < nbatch_fa; i0 += np*warp_size) {
-                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < k_VKQ_sup ?
+                const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < static_cast<uint32_t>(k_VKQ_sup) ?
                     expf(KQ_acc[(i0/(np*warp_size))*cpw + jc] - KQ_max[jc]) : 0.0f;
                 KQ_sum_add += val;
                 tmp[i0/(np*warp_size)][jc1] = val;
@@ -723,7 +749,7 @@ static __global__ void flash_attn_tile(
 
     if (
 #ifdef GGML_USE_WMMA_FATTN
-            (ncols2 != 1 && DV != 40 && DV != 512) ||
+            (ncols2 != 1 && DV != 40 && DV != 72 && DV != 512) ||
 #endif // GGML_USE_WMMA_FATTN
             (use_logit_softcap && !(DV == 128 || DV == 256))
     ) {
@@ -1198,6 +1224,7 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
 
 extern DECL_FATTN_TILE_CASE( 40,  40);
 extern DECL_FATTN_TILE_CASE( 64,  64);
+extern DECL_FATTN_TILE_CASE( 72,  72);
 extern DECL_FATTN_TILE_CASE( 80,  80);
 extern DECL_FATTN_TILE_CASE( 96,  96);
 extern DECL_FATTN_TILE_CASE(112, 112);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-vec.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-vec.cuh
index e1838fdd..67aa67ec 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/fattn-vec.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/fattn-vec.cuh
@@ -86,11 +86,11 @@ static __global__ void flash_attn_ext_vec(
 
     constexpr vec_dot_KQ_t vec_dot_KQ = get_vec_dot_KQ<type_K, D, nthreads_KQ>();
     constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
     constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, half,  V_rows_per_thread>();
 #else
     constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, float, V_rows_per_thread>();
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
@@ -112,13 +112,13 @@ static __global__ void flash_attn_ext_vec(
 
     constexpr int ne_KQ      = ncols*D;
     constexpr int ne_combine = nwarps*V_cols_per_iter*D;
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
     half2            VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
     __shared__ half   KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
 #else
     float2           VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
     __shared__ float  KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
 
     float KQ_max[ncols];
     float KQ_sum[ncols];
@@ -129,11 +129,11 @@ static __global__ void flash_attn_ext_vec(
     }
 
     // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
     half2  Q_reg[ncols][(D/2)/nthreads_KQ]; // Will be initialized completely.
 #else
     float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
     int    Q_i32[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
     float2  Q_ds[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
     if constexpr (Q_q8_1) {
@@ -155,7 +155,7 @@ static __global__ void flash_attn_ext_vec(
                 for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
                     const int i = i0 + threadIdx.x;
 
-                    if (i0 + WARP_SIZE <= D/sizeof(int) || i < D/sizeof(int)) {
+                    if (i0 + WARP_SIZE <= int(D/sizeof(int)) || i < int(D/sizeof(int))) {
                         tmp_q_i32[i] = 0;
                     }
                 }
@@ -191,7 +191,7 @@ static __global__ void flash_attn_ext_vec(
 
         __syncthreads();
     } else {
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
         const half2 scale_h2 = make_half2(scale, scale);
 #pragma unroll
         for (int j = 0; j < ncols; ++j) {
@@ -233,7 +233,7 @@ static __global__ void flash_attn_ext_vec(
                 Q_reg[j][k].y *= scale;
             }
         }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
     }
 
     const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
@@ -272,7 +272,7 @@ static __global__ void flash_attn_ext_vec(
 
                 KQ_max_new[j] = fmaxf(KQ_max_new[j], sum);
 
-                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == i_KQ_0) {
+                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == uint32_t(i_KQ_0)) {
                     KQ_reg[j] = sum;
                 }
             }
@@ -291,7 +291,7 @@ static __global__ void flash_attn_ext_vec(
             KQ_sum[j] = KQ_sum[j]*KQ_max_scale + KQ_reg[j];
             KQ[j*nthreads + tid] = KQ_reg[j];
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
 #pragma unroll
             for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
@@ -303,7 +303,7 @@ static __global__ void flash_attn_ext_vec(
                 VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
                 VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
             }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
 
 #ifndef GGML_USE_HIP
@@ -314,7 +314,7 @@ static __global__ void flash_attn_ext_vec(
         for (int k0 = 0; k0 < WARP_SIZE; k0 += V_cols_per_iter) {
             const int k = threadIdx.y*WARP_SIZE + k0 + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V);
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             half2 KQ_k[ncols];
 #pragma unroll
             for (int j = 0; j < ncols; ++j) {
@@ -353,7 +353,7 @@ static __global__ void flash_attn_ext_vec(
                     }
                 }
             }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 
@@ -374,7 +374,7 @@ static __global__ void flash_attn_ext_vec(
 
             KQ_sum[j] = KQ_sum[j]*KQ_max_scale + (threadIdx.x == 0 ? expf(sink - KQ_max[j]) : 0.0f);
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
             const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
 #pragma unroll
             for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
@@ -386,7 +386,7 @@ static __global__ void flash_attn_ext_vec(
                 VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
                 VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
             }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
         }
     }
 
@@ -421,7 +421,7 @@ static __global__ void flash_attn_ext_vec(
         const float kqmax_scale = expf(KQ_max[j_VKQ] - kqmax_new);
         KQ_max[j_VKQ] = kqmax_new;
 
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
         half2 * VKQ_tmp = (half2 *) KQ + threadIdx.y*(V_cols_per_iter*D/2)
             + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V)*(D/2);
 
@@ -452,7 +452,7 @@ static __global__ void flash_attn_ext_vec(
             ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ,                       &VKQ[j_VKQ][i_VKQ_0/nthreads_V]);
             ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ + V_rows_per_thread/4, &VKQ[j_VKQ][i_VKQ_0/nthreads_V + V_rows_per_thread/4]);
         }
-#endif // FAST_FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
 
         KQ_sum[j_VKQ] *= kqmax_scale;
         KQ_sum[j_VKQ] = warp_reduce_sum(KQ_sum[j_VKQ]);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu b/ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu
index 7dee032c..82405991 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu
@@ -223,6 +223,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     switch (K->ne[0]) {
         case  40:
         case  64:
+        case  72:
         case  80:
         case  96:
         case 128:
@@ -275,7 +276,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
     // If Turing tensor cores available, use them:
-    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40) {
+    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72) {
         if (can_use_vector_kernel) {
             if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
@@ -301,7 +302,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     }
 
     // Use the WMMA kernel if possible:
-    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 576) {
+    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 576) {
         if (can_use_vector_kernel && Q->ne[1] <= 2) {
             return BEST_FATTN_KERNEL_VEC;
         }
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu b/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu
index 1a71e07c..7350f675 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -50,8 +50,10 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
+#include "ggml-cuda/solve_tri.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -636,10 +638,13 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     size_t alloc_size() override {
         return pool_size + last_alloc;
     }
+
 };
 #endif // defined(GGML_USE_VMM)
 
-std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device, bool alloc) {
+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int                  device,
+                                                                               [[maybe_unused]] int stream_no,
+                                                                               bool alloc) {
 #if defined(GGML_USE_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device, alloc));
@@ -2089,8 +2094,15 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
 
         size_t src1_stride_size = sizeof(cuda_t);
 
-        dim3 block_dims(ne13, ne12);
-        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
+        const int threads_x = 16;
+        const int threads_y = 16;
+        dim3 block_dims(threads_x, threads_y);
+
+        dim3 grid_dims(
+            (ne13 + threads_x - 1) / threads_x,
+            (ne12 + threads_y - 1) / threads_y
+        );
+        k_compute_batched_ptrs<<<grid_dims, block_dims, 0, main_stream>>>(
                 src0_ptr, src1_ptr, dst_t,
                 ptrs_src.get(), ptrs_dst.get(),
                 ne12, ne13,
@@ -2139,6 +2151,164 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     }
 }
 
+static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
+                                          const ggml_tensor * ffn_gate,
+                                          const ggml_tensor * glu,
+                                          const ggml_tensor * ffn_up_bias = nullptr,
+                                          const ggml_tensor * ffn_gate_bias = nullptr) {
+    const bool has_bias = ffn_up_bias != nullptr || ffn_gate_bias != nullptr;
+
+    if (has_bias && (!ffn_up_bias || !ffn_gate_bias)) {
+        return false;
+    }
+
+    const bool is_mul_mat     = ffn_up->op == GGML_OP_MUL_MAT     && ffn_gate->op == GGML_OP_MUL_MAT     && glu->op == GGML_OP_GLU;
+    const bool is_mul_mat_id  = ffn_up->op == GGML_OP_MUL_MAT_ID  && ffn_gate->op == GGML_OP_MUL_MAT_ID  && glu->op == GGML_OP_GLU;
+
+    GGML_ASSERT(ffn_up && ffn_gate && glu);
+
+    if (!is_mul_mat && !is_mul_mat_id) {
+        return false;
+    }
+
+    const ggml_op expected_bias_op = is_mul_mat ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+    if (has_bias) {
+        if (ffn_up_bias->op != expected_bias_op || ffn_gate_bias->op != expected_bias_op) {
+            return false;
+        }
+
+        if (glu->src[0] != ffn_gate_bias || glu->src[1] != ffn_up_bias) {
+            return false;
+        }
+
+        if (expected_bias_op == GGML_OP_ADD) {
+            const bool up_has_mul   = ffn_up_bias->src[0] == ffn_up || ffn_up_bias->src[1] == ffn_up;
+            const bool gate_has_mul = ffn_gate_bias->src[0] == ffn_gate || ffn_gate_bias->src[1] == ffn_gate;
+            if (!up_has_mul || !gate_has_mul) {
+                return false;
+            }
+        } else { // GGML_OP_ADD_ID
+            if (ffn_up_bias->src[0] != ffn_up || ffn_gate_bias->src[0] != ffn_gate) {
+                return false;
+            }
+            if (ffn_up_bias->src[2] != ffn_up->src[2] || ffn_gate_bias->src[2] != ffn_gate->src[2]) {
+                return false;
+            }
+        }
+    } else {
+        if (glu->src[0] != ffn_gate && glu->src[1] != ffn_up) {
+            return false;
+        }
+    }
+
+    if (ffn_up->src[0]->type != ffn_gate->src[0]->type || !ggml_are_same_shape(ffn_up->src[0], ffn_gate->src[0]) ||
+        !ggml_are_same_stride(ffn_up->src[0], ffn_gate->src[0])) {
+        return false;
+    }
+
+    if (ffn_up->src[1] != ffn_gate->src[1]) {
+        return false;
+    }
+
+    if (ffn_up->src[2] && (ffn_up->src[2] != ffn_gate->src[2])) {
+        return false;
+    }
+
+    static constexpr std::array<ggml_glu_op, 3> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI };
+
+    if (std::find(valid_glu_ops.begin(), valid_glu_ops.end(), ggml_get_glu_op(glu)) == valid_glu_ops.end()) {
+        return false;
+    }
+
+    if (const bool swapped = ggml_get_op_params_i32(glu, 1); swapped) {
+        return false;
+    }
+
+    const bool split = ggml_backend_buft_is_cuda_split(ffn_up->src[0]->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(ffn_gate->src[0]->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    return true;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool is_mul_mat_id = tensor->op == GGML_OP_MUL_MAT_ID;
+
+    bool use_mul_mat_vec_f =
+        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16) &&
+        src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    const int cc      = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+
+    return use_mul_mat_vec_f;
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
+    ggml_tensor *       src0 = tensor->src[0];
+    ggml_tensor *       src1 = tensor->src[1];
+    const ggml_tensor * dst  = tensor;
+
+    const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE &&
+                                   ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) &&
+                                   src0->view_src;
+
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 &&
+                             dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+
+    // fusion is not universally faster on Pascal
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    if (cc <= GGML_CUDA_CC_PASCAL) {
+        return false;
+    }
+    //we only support fusion for ncols_dst = 1
+    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
+        return false;
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
+        return false;
+    }
+
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
+    return use_mul_mat_vec_q;
+}
+
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
@@ -2172,16 +2342,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
             const int cc            = ggml_cuda_info().devices[id].cc;
             const int warp_size     = ggml_cuda_info().devices[id].warp_size;
             use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
             any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
         const int cc            = ggml_cuda_info().devices[ctx.device].cc;
         const int warp_size     = ggml_cuda_info().devices[ctx.device].warp_size;
         use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
         any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
@@ -2252,7 +2422,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             return;
         }
 
-        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2], /*mul_mat_id=*/true)) {
+        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
             ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
             return;
         }
@@ -2400,6 +2570,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SET_ROWS:
             ggml_cuda_op_set_rows(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cuda_op_set(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2478,6 +2651,24 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_XIELU:
                     ggml_cuda_op_xielu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_FLOOR:
+                    ggml_cuda_op_floor(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_CEIL:
+                    ggml_cuda_op_ceil(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ROUND:
+                    ggml_cuda_op_round(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TRUNC:
+                    ggml_cuda_op_trunc(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_EXPM1:
+                    ggml_cuda_op_expm1(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    ggml_cuda_op_softplus(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2662,6 +2853,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_SGD:
             ggml_cuda_opt_step_sgd(ctx, dst);
             break;
+        case GGML_OP_SOLVE_TRI:
+            ggml_cuda_op_solve_tri(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2887,7 +3081,7 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
         }
     }
 
-    if (node->op == GGML_OP_SCALE &&
+    if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
         memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
         return false;
     }
@@ -2953,6 +3147,40 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 }
 #endif
 
+static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
+                                                const ggml_tensor * view,
+                                                const ggml_tensor * set_rows) {
+
+    if (rope->op != GGML_OP_ROPE || view->op != GGML_OP_VIEW || set_rows->op != GGML_OP_SET_ROWS) {
+        return false;
+    }
+    // ne3 not tested
+    if (rope->src[0]->ne[3] != 1) {
+        return false;
+    }
+
+    if (set_rows->type != GGML_TYPE_F32 && set_rows->type != GGML_TYPE_F16) {
+        return false;
+    }
+
+    if (set_rows->src[1]->type != GGML_TYPE_I64) {
+        return false;
+    }
+
+    // The view should flatten two dims of rope into one dim
+    if (!ggml_is_contiguous(view) || view->ne[0] != rope->ne[0] * rope->ne[1]) {
+        return false;
+    }
+
+    // Only norm/neox shaders have the fusion code
+    const int mode = ((const int32_t *) rope->op_params)[2];
+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops, std::initializer_list<enum ggml_unary_op> unary_ops) {
 #ifndef NDEBUG
     const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY);
@@ -2967,27 +3195,31 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     std::initializer_list<enum ggml_op> topk_moe_ops_delayed_softmax =
         ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
 
-    if (ops.size() == topk_moe_ops_with_norm.size() &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 8 })) {
+    const auto is_equal = [](const std::initializer_list<enum ggml_op> & list1,
+                             const std::initializer_list<enum ggml_op> & list2) {
+        return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end());
+    };
+
+    if (is_equal(topk_moe_ops_with_norm, ops) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx+8];
+        ggml_tensor * weights = cgraph->nodes[node_idx + 9];
 
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
             return true;
         }
     }
 
-    if (ops.size() == topk_moe_ops.size() &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
+    if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx+4];
+        ggml_tensor * weights = cgraph->nodes[node_idx + 4];
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
             return true;
         }
     }
 
-    if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2, node_idx + 5 })) {
+    if (is_equal(topk_moe_ops_delayed_softmax, ops) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
         ggml_tensor * weights = cgraph->nodes[node_idx + 5];
 
@@ -2996,6 +3228,48 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
+    std::initializer_list<enum ggml_op> mul_mat_bias_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_GLU };
+    std::initializer_list<enum ggml_op> mul_mat_id_bias_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU };
+
+    std::initializer_list<enum ggml_op> mul_mat_id_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_MUL_MAT_ID, GGML_OP_GLU };
+    std::initializer_list<enum ggml_op> mul_mat_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_MUL_MAT,    GGML_OP_GLU };
+
+    if ((is_equal(mul_mat_bias_glu_ops, ops) || is_equal(mul_mat_id_bias_glu_ops, ops)) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 4 })) {
+        const ggml_tensor * ffn_gate      = cgraph->nodes[node_idx];
+        const ggml_tensor * ffn_gate_bias = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * ffn_up        = cgraph->nodes[node_idx + 2];
+        const ggml_tensor * ffn_up_bias   = cgraph->nodes[node_idx + 3];
+        const ggml_tensor * glu           = cgraph->nodes[node_idx + 4];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu, ffn_up_bias, ffn_gate_bias)) {
+            return true;
+        }
+    }
+
+    if ((is_equal(mul_mat_id_glu_ops, ops) || is_equal(mul_mat_glu_ops, ops)) &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+        const ggml_tensor * ffn_gate = cgraph->nodes[node_idx];
+        const ggml_tensor * ffn_up   = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * glu      = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_mul_mat(ffn_up, ffn_gate, glu)) {
+            return true;
+        }
+    }
+
+    std::initializer_list<enum ggml_op> rope_set_rows_ops = { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS };
+
+    if (is_equal(rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+        const ggml_tensor * rope     = cgraph->nodes[node_idx];
+        const ggml_tensor * view     = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2];
+
+        if (ggml_cuda_should_fuse_rope_set_rows(rope, view, set_rows)) {
+            return true;
+        }
+    }
+
     if (!ggml_can_fuse(cgraph, node_idx, ops)) {
         return false;
     }
@@ -3072,13 +3346,87 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
     // flag used to determine whether it is an integrated_gpu
     const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
 
+    ggml_cuda_stream_context & stream_ctx = cuda_ctx->stream_context();
+    bool                         is_concurrent_event_active = false;
+    ggml_cuda_concurrent_event * concurrent_event           = nullptr;
+    bool                         should_launch_concurrent_events = false;
+
+    const auto try_launch_concurrent_event = [&](const ggml_tensor * node) {
+        if (stream_ctx.concurrent_events.find(node) != stream_ctx.concurrent_events.end()) {
+            concurrent_event = &stream_ctx.concurrent_events[node];
+
+            is_concurrent_event_active = true;
+
+            GGML_LOG_DEBUG("Launching %d streams at %s\n", concurrent_event->n_streams, node->name);
+
+            cudaStream_t main_stream = cuda_ctx->stream();  // this should be stream 0
+            GGML_ASSERT(cuda_ctx->curr_stream_no == 0);
+            CUDA_CHECK(cudaEventRecord(concurrent_event->fork_event, main_stream));
+
+            for (int i = 1; i <= concurrent_event->n_streams; ++i) {
+                cudaStream_t stream = cuda_ctx->stream(cuda_ctx->device, i);
+                CUDA_CHECK(cudaStreamWaitEvent(stream, concurrent_event->fork_event));
+            }
+        }
+    };
+
     while (!graph_evaluated_or_captured) {
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
+            [[maybe_unused]] int prev_i = 0;
+
+            if (stream_ctx.concurrent_events.size() > 0) {
+                should_launch_concurrent_events = true;
+                for (const auto & [tensor, event] : stream_ctx.concurrent_events) {
+                    should_launch_concurrent_events = should_launch_concurrent_events && event.is_valid();
+                }
+            }
+            if (should_launch_concurrent_events) {
+                //Restore the original graph to enable fusion within the streams
+                cgraph->nodes   = const_cast<ggml_tensor **>(stream_ctx.original_nodes.data());
+                cgraph->n_nodes = (int) stream_ctx.original_nodes.size();
+            }
 
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
+                if (is_concurrent_event_active) {
+                    GGML_ASSERT(concurrent_event);
+
+                    if (node == concurrent_event->join_node) {
+                        cuda_ctx->curr_stream_no = 0;
+                        for (int i = 1; i <= concurrent_event->n_streams; ++i) {
+                            // Wait on join events of forked streams in the main stream
+                            CUDA_CHECK(cudaEventRecord(concurrent_event->join_events[i - 1],
+                                                       cuda_ctx->stream(cuda_ctx->device, i)));
+                            CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), concurrent_event->join_events[i - 1]));
+                        }
+
+                        is_concurrent_event_active = false;
+                        concurrent_event           = nullptr;
+                    } else {
+                        GGML_ASSERT (concurrent_event->stream_mapping.find(node) != concurrent_event->stream_mapping.end());
+                        cuda_ctx->curr_stream_no = concurrent_event->stream_mapping[node];
+                        GGML_LOG_DEBUG("Setting stream no to %d for node %s\n", cuda_ctx->curr_stream_no, node->name);
+                    }
+                } else if (i - prev_i > 1) {
+                    //the previous node was fused
+                    const ggml_tensor * prev_node = cgraph->nodes[i - 1];
+                    try_launch_concurrent_event(prev_node);
+
+                    if (is_concurrent_event_active) {
+                        cuda_ctx->curr_stream_no = concurrent_event->stream_mapping[node];
+                        GGML_LOG_DEBUG("Setting stream no to %d for node %s\n", cuda_ctx->curr_stream_no, node->name);
+                    }
+                }
+                prev_i = i;
+
+#ifdef GGML_CUDA_DEBUG
+                const int nodes_fused = i - prev_i - 1;
+                if (nodes_fused > 0) {
+                    GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
+                }
+#endif
 
                 if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                     continue;
@@ -3089,21 +3437,23 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     continue;
                 }
 
+                // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                 if (!disable_fusion) {
 
                     if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i+8];
-                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
+                        ggml_tensor * weights          = cgraph->nodes[i + 9];
+                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
+                        ggml_tensor * clamp            = cgraph->nodes[i + 7];
                         ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ true,
-                                              /*delayed softmax*/ false);
-                        i += 8;
+                                              /*delayed softmax*/ false, clamp);
+                        i += 9;
                         continue;
                     }
 
                     if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ false), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i+4];
-                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
+                        ggml_tensor * weights          = cgraph->nodes[i + 4];
+                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
                         ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ false,
                                               /*delayed softmax*/ false);
                         i += 4;
@@ -3121,6 +3471,15 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         continue;
                     }
 
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
+                        ggml_tensor * rope = cgraph->nodes[i];
+                        ggml_tensor * set_rows = cgraph->nodes[i + 2];
+
+                        ggml_cuda_op_rope_fused(*cuda_ctx, rope, set_rows);
+                        i += 2;
+                        continue;
+                    }
+
                     if (node->op == GGML_OP_ADD) {
                         int n_fuse = 0;
                         ggml_op ops[8];
@@ -3152,6 +3511,195 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         }
                     }
 
+                    bool fused_mul_mat_vec = false;
+                    int fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (ggml_cuda_can_fuse(cgraph, i, { op, bias_op, op, bias_op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu         = cgraph->nodes[i + 4];
+                            ggml_tensor * gate_bias_n = glu->src[0];
+                            ggml_tensor * up_bias_n   = glu->src[1];
+
+                            //we don't assume the order for {gate, up}. Instead infer it from the bias tensor
+                            ggml_tensor * gate_n      = nullptr;
+                            ggml_tensor * up_n        = nullptr;
+
+                            if (gate_bias_n->src[0] == cgraph->nodes[i] || gate_bias_n->src[1] == cgraph->nodes[i]) {
+                                gate_n = cgraph->nodes[i];
+                                up_n   = cgraph->nodes[i + 2];
+                            } else if (gate_bias_n->src[0] == cgraph->nodes[i + 2] || gate_bias_n->src[1] == cgraph->nodes[i + 2]) {
+                                gate_n = cgraph->nodes[i + 2];
+                                up_n   = cgraph->nodes[i];
+                            } else {
+                                continue;
+                            }
+
+                            auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) {
+                                if (op_bias == GGML_OP_ADD) {
+                                    if (bias_node->src[0] == mul_node) {
+                                        return bias_node->src[1];
+                                    }
+                                    if (bias_node->src[1] == mul_node) {
+                                        return bias_node->src[0];
+                                    }
+                                    return (ggml_tensor *) nullptr;
+                                }
+                                GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
+                                GGML_ASSERT(bias_node->src[0] == mul_node);
+                                return bias_node->src[1];
+                            };
+
+                            ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
+                            ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);
+
+                            if (!up_bias_tensor || !gate_bias_tensor) {
+                                continue;
+                            }
+
+                            // we don't support repeating adds
+                            if (bias_op == GGML_OP_ADD &&
+                                (!ggml_are_same_shape(gate_bias_n->src[0], gate_bias_n->src[1]) ||
+                                 !ggml_are_same_shape(up_bias_n->src[0], up_bias_n->src[1]))) {
+                                continue;
+                            }
+
+                            const ggml_tensor * src0 = up_n->src[0];
+                            const ggml_tensor * src1 = up_n->src[1];
+                            const ggml_tensor * ids  = up_n->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate      = gate_n->src[0];
+                                fusion_data.x_bias    = up_bias_tensor;
+                                fusion_data.gate_bias = gate_bias_tensor;
+                                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 5;
+                                break;
+                            }
+                        } else if (ggml_cuda_can_fuse(cgraph, i, { op, op, GGML_OP_GLU }, {})) {
+                            ggml_tensor * glu  = cgraph->nodes[i + 2];
+                            ggml_tensor * gate = glu->src[0];
+                            ggml_tensor * up   = glu->src[1];
+
+                            bool ok = (gate == cgraph->nodes[i] && up == cgraph->nodes[i + 1])
+                                || (gate == cgraph->nodes[i + 1] && up == cgraph->nodes[i]);
+
+                            if (!ok) continue;
+
+                            const ggml_tensor * src0 = up->src[0];
+                            const ggml_tensor * src1 = up->src[1];
+                            const ggml_tensor * ids  = up->src[2];
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+
+                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
+                                ggml_cuda_mm_fusion_args_host fusion_data{};
+                                fusion_data.gate   = gate->src[0];
+                                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                                fused_mul_mat_vec = true;
+                                fused_node_count = 3;
+                                break;
+                            }
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
+
+                    fused_mul_mat_vec = false;
+                    fused_node_count = 0;
+
+                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+                        if (!ggml_can_fuse(cgraph, i, { op, bias_op })) {
+                            continue;
+                        }
+
+                        ggml_tensor * mm_node   = cgraph->nodes[i];
+                        ggml_tensor * bias_node = cgraph->nodes[i + 1];
+
+                        ggml_tensor * bias_tensor = nullptr;
+                        if (bias_op == GGML_OP_ADD) {
+                            if (bias_node->src[0] == mm_node) {
+                                bias_tensor = bias_node->src[1];
+                            } else if (bias_node->src[1] == mm_node) {
+                                bias_tensor = bias_node->src[0];
+                            } else {
+                                continue;
+                            }
+                        } else {
+                            if (bias_node->src[0] != mm_node) {
+                                continue;
+                            }
+                            bias_tensor = bias_node->src[1];
+                        }
+
+                        const ggml_tensor * src0 = mm_node->src[0];
+                        const ggml_tensor * src1 = mm_node->src[1];
+                        const ggml_tensor * ids  = mm_node->src[2];
+
+                        if (bias_op == GGML_OP_ADD_ID && bias_node->src[2] != ids) {
+                            continue;
+                        }
+
+                        if (bias_op == GGML_OP_ADD && !ggml_are_same_shape(bias_node->src[0], bias_node->src[1])) {
+                            continue;
+                        }
+
+                        ggml_cuda_mm_fusion_args_host fusion_data{};
+                        fusion_data.x_bias = bias_tensor;
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_f(mm_node)) {
+                            ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+
+                        if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
+                            ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+                            fused_mul_mat_vec = true;
+                            fused_node_count = 2;
+                            break;
+                        }
+                    }
+
+                    if (fused_mul_mat_vec) {
+                        i += fused_node_count - 1;
+                        continue;
+                    }
 
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD}, {})) {
                         ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
@@ -3182,13 +3730,17 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 }
 #else
                 GGML_UNUSED(integrated);
-#endif // NDEBUG
+#endif  // NDEBUG
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
                     GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
                 }
                 GGML_ASSERT(ok);
+
+                if (!is_concurrent_event_active) {
+                    try_launch_concurrent_event(node);
+               }
             }
         }
 
@@ -3366,7 +3918,7 @@ static size_t ggml_backend_cuda_buffer_size(ggml_backend_t backend) {
 
 static void ggml_backend_cuda_reset(ggml_backend_t backend) {
     ggml_backend_cuda_context * ctx = (ggml_backend_cuda_context *)backend->context;
-    ctx->pools[ctx->device] = NULL;
+    ctx->pools[ctx->device][ctx->curr_stream_no] = NULL;
 }
 
 static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
@@ -3394,6 +3946,235 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
     }
 }
 
+static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
+
+    static bool enable_graph_optimization = [] {
+        const char * env = getenv("GGML_CUDA_GRAPH_OPT");
+        return env != nullptr && atoi(env) == 1;
+    }();
+
+    if (!enable_graph_optimization) {
+        return;
+    }
+
+    GGML_ASSERT(ggml_backend_cuda_get_device_count() == 1 && "compute graph optimization is only supported on single GPU in the CUDA backend");
+    GGML_LOG_DEBUG("Optimizing CUDA graph %p with %d nodes\n", cgraph->nodes, cgraph->n_nodes);
+
+    ggml_cuda_stream_context & stream_context = cuda_ctx->stream_context();
+    stream_context.reset();
+
+    // number of out-degrees for a particular node
+    std::unordered_map<const ggml_tensor *, int> fan_out;
+    // reverse mapping of node to index in the cgraph
+    std::unordered_map<const ggml_tensor *, int> node_indices;
+
+    const auto & is_noop = [](const ggml_tensor * node) -> bool {
+        return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE ||
+               node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
+    };
+
+    const auto & depends_on = [](const ggml_tensor * dst, const ggml_tensor * src) -> bool {
+        for (uint32_t s = 0; s < GGML_MAX_SRC; ++s) {
+            if (dst->src[s] == src) {
+                return true;
+            }
+        }
+        // implicit dependency if they view the same tensor
+        const ggml_tensor * dst2 = dst->view_src ? dst->view_src : dst;
+        const ggml_tensor * src2 = src->view_src ? src->view_src : src;
+        if (dst2 == src2) {
+            return true;
+        }
+        return false;
+    };
+
+    for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
+        const ggml_tensor * node = cgraph->nodes[node_idx];
+        node_indices[node]       = node_idx;
+
+        if (is_noop(node)) {
+            continue;
+        }
+        for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+            const ggml_tensor * src = cgraph->nodes[node_idx]->src[src_idx];
+            //TODO: check why nrows > 1 fails
+            if (node && !is_noop(node) && ggml_nrows(node) <= 1) {
+                fan_out[src] += 1;
+            }
+        }
+    }
+
+    // Target Q, K, V for concurrency
+    // this is a more general way to find nodes which can be candidates for concurrency (although it has not been tested for anything else):
+    // 1. find fan-out (fork) nodes where the same input is used at least N times (in QKV, it would be "attn-norm")
+    // 2. find the join node, where 2 or more of the outputs are required (in QKV, this would "KQ" or "flash-attn")
+    // 3. account for all branches from the fork to the join
+    // 4. To extend lifetimes of the tensors, we interleave the branches (see below for more details)
+    // 5. save the original cgraph and restore it in graph_compute, to enable fusion within streams
+    // See discussion: https://github.com/ggml-org/llama.cpp/pull/16991#issuecomment-3522620030
+
+    const int min_fan_out = 3;
+    const int max_fan_out = 3;
+
+    // store {fork_idx, join_idx}
+    std::vector<std::pair<int, int>> concurrent_node_ranges;
+
+    // save the original nodes
+    std::vector<const ggml_tensor *> original_nodes;
+    original_nodes.reserve(cgraph->n_nodes);
+    for (int i = 0; i < cgraph->n_nodes; ++i) {
+        original_nodes.push_back(cgraph->nodes[i]);
+    }
+    cuda_ctx->stream_context().original_nodes = std::move(original_nodes);
+
+    for (const auto & [root_node, count] : fan_out) {
+        if (count >= min_fan_out && count <= max_fan_out) {
+            const int root_node_idx = node_indices[root_node];
+
+            bool is_part_of_event = false;
+            for (const auto & [start, end] : concurrent_node_ranges) {
+                if (root_node_idx >= start && root_node_idx <= end) {
+                    is_part_of_event = true;
+                }
+            }
+
+            if (is_part_of_event) {
+                continue;
+            }
+
+            std::vector<std::vector<const ggml_tensor *>> nodes_per_branch;
+            for (int i = root_node_idx + 1; i < cgraph->n_nodes; ++i) {
+                const ggml_tensor * node = cgraph->nodes[i];
+                if (!is_noop(node) && depends_on(node, root_node)) {
+                    nodes_per_branch.push_back({ node });
+                }
+            }
+
+            GGML_ASSERT(nodes_per_branch.size() == (size_t) count);
+
+            //find the join point
+            const ggml_tensor * join_node = nullptr;
+
+            const auto & belongs_to_branch = [&](const ggml_tensor *                      node,
+                                                 const std::vector<const ggml_tensor *> & branch) -> bool {
+                for (const ggml_tensor * n : branch) {
+                    if (depends_on(node, n)) {
+                        return true;
+                    }
+                }
+                return false;
+            };
+
+            for (int i = root_node_idx + 1; i < cgraph->n_nodes; ++i) {
+                const ggml_tensor * curr_node = cgraph->nodes[i];
+
+                int num_joins = 0;
+                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
+                    if (belongs_to_branch(curr_node, nodes_per_branch[branch_idx])) {
+                        num_joins++;
+                    }
+                }
+
+                if (num_joins >= 2) {
+                    join_node = curr_node;
+                    break;
+                }
+
+                bool found_branch = false;
+                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
+                    std::vector<const ggml_tensor *> & branch_vec = nodes_per_branch[branch_idx];
+                    if (belongs_to_branch(curr_node, branch_vec)) {
+                        //continue accumulating
+                        if (std::find(branch_vec.begin(), branch_vec.end(), curr_node) == branch_vec.end()) {
+                            branch_vec.push_back(curr_node);
+                        }
+                        found_branch = true;
+                    }
+                }
+
+                if (!found_branch && is_noop(curr_node)) {
+                    // we can put it in any branch because it will be ignored
+                    nodes_per_branch[0].push_back({ curr_node });
+                }
+            }
+
+            if (join_node) {
+                //Create ggml_cuda_concurrent_event
+                ggml_cuda_concurrent_event concurrent_event(nodes_per_branch.size());
+                concurrent_event.join_node = join_node;
+
+                for (size_t branch_idx = 0; branch_idx < nodes_per_branch.size(); branch_idx++) {
+                    for (const ggml_tensor * n : nodes_per_branch[branch_idx]) {
+                        concurrent_event.stream_mapping[n] = branch_idx + 1;
+                    }
+                }
+
+                int fork_node_idx = node_indices[root_node];
+                int join_node_idx = node_indices[join_node];
+
+                int       current_branch_idx = 0;
+                int       current_node_idx   = fork_node_idx + 1;
+                const int n_branches         = nodes_per_branch.size();
+
+                int total_branch_nodes = 0;
+                for (std::vector<const ggml_tensor *> branch_nodes : nodes_per_branch) {
+                    total_branch_nodes += branch_nodes.size();
+                }
+
+                // there are other nodes in the middle which are unaccounted for
+                // usually (cpy) nodes, then ignore this fork
+                if (join_node_idx - fork_node_idx - 1 != total_branch_nodes) {
+                    GGML_LOG_DEBUG(
+                        "Skipping %s because the number of nodes in the middle is not equal to the total number of "
+                        "branch nodes %d != %d\n",
+                        root_node->name, join_node_idx - fork_node_idx - 1, total_branch_nodes);
+                    continue;
+                }
+
+                std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> & concurrent_events = cuda_ctx->stream_context().concurrent_events;
+                GGML_ASSERT(concurrent_events.find(root_node) == concurrent_events.end());
+                concurrent_events.emplace(root_node, std::move(concurrent_event));
+                GGML_LOG_DEBUG("Adding stream at node %s %p\n", root_node->name, root_node);
+                concurrent_node_ranges.emplace_back(fork_node_idx, join_node_idx);
+
+                // interleave tensors to extend lifetimes so that ggml graph doesn't recycle them
+                // example transformation:
+                // [attn-norm, QMul, QNorm, QRope, KMul, KNorm, KRope, VMul, attn] ->
+                // [attn-norm, QMul, KMul, VMul, QNorm, VNorm, QRope, KRope, attn]
+                while (current_node_idx < join_node_idx) {
+                    std::vector<const ggml_tensor *> & branch_nodes = nodes_per_branch[current_branch_idx];
+
+                    bool has_node = false;
+                    for (std::vector<const ggml_tensor *> branch_node : nodes_per_branch) {
+                        has_node |= branch_node.size() > 0;
+                    }
+
+                    GGML_ASSERT(has_node);
+
+                    if (branch_nodes.empty()) {
+                        current_branch_idx = (current_branch_idx + 1) % n_branches;
+                        continue;
+                    }
+
+                    cgraph->nodes[current_node_idx] = const_cast<ggml_tensor *>(branch_nodes.front());
+                    current_node_idx++;
+                    branch_nodes.erase(branch_nodes.begin());
+
+                    // append all empty nodes
+                    while (!branch_nodes.empty() && is_noop(branch_nodes.front())) {
+                        cgraph->nodes[current_node_idx] = const_cast<ggml_tensor *>(branch_nodes.front());
+                        current_node_idx++;
+                        branch_nodes.erase(branch_nodes.begin());
+                    }
+
+                    current_branch_idx = (current_branch_idx + 1) % n_branches;
+                }
+            }
+        }
+    }
+}
+
 static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .get_name                = */ ggml_backend_cuda_get_name,
     /* .free                    = */ ggml_backend_cuda_free,
@@ -3408,7 +4189,7 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .event_record            = */ ggml_backend_cuda_event_record,
     /* .event_wait              = */ ggml_backend_cuda_event_wait,
-    /* .graph_optimize          = */ NULL,
+    /* .graph_optimize          = */ ggml_backend_cuda_graph_optimize,
     /* .graph_reserve           = */ ggml_backend_cuda_graph_reserve,
     /* .buffer_size             = */ ggml_backend_cuda_buffer_size,
     /* .reset                   = */ ggml_backend_cuda_reset,
@@ -3500,6 +4281,82 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
     return ctx->description.c_str();
 }
 
+#if defined(__linux__)
+// Helper function to get available memory from /proc/meminfo for UMA systems
+static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_kb, long * free_swap_kb) {
+    FILE * meminfo_file = nullptr;
+    // 2KB buffer for reading /proc/meminfo since it does not report size info, should be enough
+    const size_t BUFFER_SIZE = 2048;
+    auto file_buffer = std::make_unique<char[]>(BUFFER_SIZE);
+    size_t bytes_read = 0;
+    long huge_tlb_total_pages = -1;
+    long huge_tlb_free_pages = -1;
+    long huge_tlb_page_size = -1;
+
+    if (available_memory_kb == nullptr || free_swap_kb == nullptr) {
+        return false;
+    }
+
+    meminfo_file = fopen("/proc/meminfo", "r");
+    if (meminfo_file == nullptr) {
+        GGML_LOG_ERROR("%s: failed to open /proc/meminfo\n", __func__);
+        return false;
+    }
+
+    // Read file into buffer
+    bytes_read = fread(file_buffer.get(), 1, BUFFER_SIZE - 1, meminfo_file);
+    fclose(meminfo_file);
+
+    if (bytes_read == 0) {
+        GGML_LOG_ERROR("%s: failed to read from /proc/meminfo\n", __func__);
+        return false;
+    }
+    file_buffer[bytes_read] = '\0';
+
+    *available_memory_kb = -1;
+    *free_swap_kb = -1;
+
+    // Parse the file buffer line by line
+    char * line = file_buffer.get();
+    char * line_next;
+    while (line < file_buffer.get() + bytes_read) {
+        // Find the end of the current line
+        line_next = strchr(line, '\n');
+        if (line_next != nullptr) {
+            *line_next = '\0';
+            line_next++;
+        } else {
+            line_next = file_buffer.get() + bytes_read;
+        }
+
+        long value;
+        if (sscanf(line, "MemAvailable: %ld kB", &value) == 1) {
+            *available_memory_kb = value;
+        } else if (sscanf(line, "SwapFree: %ld kB", &value) == 1) {
+            *free_swap_kb = value;
+        } else if (sscanf(line, "HugePages_Total: %ld", &value) == 1) {
+            huge_tlb_total_pages = value;
+        } else if (sscanf(line, "HugePages_Free: %ld", &value) == 1) {
+            huge_tlb_free_pages = value;
+        } else if (sscanf(line, "Hugepagesize: %ld kB", &value) == 1) {
+            huge_tlb_page_size = value;
+        }
+
+        line = line_next;
+    }
+
+    if (huge_tlb_total_pages != 0 && huge_tlb_total_pages != -1) {
+        *available_memory_kb = huge_tlb_free_pages * huge_tlb_page_size;
+
+        // Hugetlbfs pages are not swappable.
+        *free_swap_kb = 0;
+    }
+
+    GGML_LOG_DEBUG("%s: final available_memory_kb: %ld\n", __func__, *available_memory_kb);
+    return true;
+}
+#endif // defined(__linux__)
+
 static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     return ctx->id.c_str();
@@ -3531,6 +4388,30 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
     }
 #endif
     CUDA_CHECK(cudaMemGetInfo(free, total));
+
+// ref: https://github.com/ggml-org/llama.cpp/pull/17368
+#if defined(__linux__)
+    // Check if this is a UMA (Unified Memory Architecture) system
+    cudaDeviceProp prop;
+    CUDA_CHECK(cudaGetDeviceProperties(&prop, ctx->device));
+
+    // Check if UMA is explicitly enabled via environment variable
+    bool uma_env = getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr;
+    bool is_uma = prop.integrated > 0 || uma_env;
+
+    if (is_uma) {
+        // For UMA systems (like DGX Spark), use system memory info
+        long available_memory_kb = 0;
+        long free_swap_kb = 0;
+
+        if (ggml_backend_cuda_get_available_uma_memory(&available_memory_kb, &free_swap_kb) && available_memory_kb > 0) {
+            *free = (size_t)available_memory_kb * 1024;
+        } else {
+            GGML_LOG_ERROR("%s: /proc/meminfo reading failed, using cudaMemGetInfo\n", __func__);
+        }
+    }
+#endif // defined(__linux__)
+
 }
 
 static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
@@ -3636,7 +4517,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_EXPM1:
+                case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -3754,6 +4641,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                        op->src[0]->type == GGML_TYPE_F32 &&
                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
             } break;
+        case GGML_OP_SET:
+            {
+                const ggml_type t = op->type;
+                return (t == GGML_TYPE_F32 || t == GGML_TYPE_I32) &&
+                    t == op->src[0]->type &&
+                    t == op->src[1]->type;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -3802,6 +4696,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_I32 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
+                if (src0_type == GGML_TYPE_I32 && src1_type == GGML_TYPE_I32) {
+                    return true;
+                }
                 if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
                     return true;
                 }
@@ -3939,6 +4836,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return true;
+        case GGML_OP_SOLVE_TRI:
+            return op->src[0]->ne[0] <= 64 && op->src[1]->ne[0] <= 32;
         default:
             return false;
     }
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mma.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/mma.cuh
index c1f24243..0ed42e87 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mma.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mma.cuh
@@ -18,6 +18,10 @@
 
 #include "common.cuh"
 
+// On Volta each warp is doing 4 8x8 mma operations in parallel.
+// The basic memory layout for a 32x8 output tile is to stack 4 input tiles in I direction and to mirror the B tile.
+// However, the i indices in this file are by default permuted to simplify the index calculations.
+// #define GGML_CUDA_MMA_NO_VOLTA_PERM
 
 #if CUDART_VERSION >= 11080
 
@@ -69,10 +73,19 @@ namespace ggml_cuda_mma {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
 
-#if defined(GGML_USE_HIP)
+#if defined(AMD_MFMA_AVAILABLE)
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I == 64 && J ==  2) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 32 && J ==  4) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J == 32) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
                 return threadIdx.x % 16;
@@ -85,7 +98,8 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return 4 * (threadIdx.x / 32) + 8 * (l / 4) + (l % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -101,22 +115,95 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return threadIdx.x % 32;
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
+#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
+#else
+                return (l & 2) | (threadIdx.x & ~2);
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 32 && J == 8) {
+                return (threadIdx.x & 2) | (l & (4 + 1));
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#elif defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA4)
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 16) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 16) {
+                return 8 * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 16) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#endif
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 8 && (J == 4 || J == 8)) {
+            if constexpr (I == 8 && J == 4) {
+                return threadIdx.x / 4;
+            } else if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 8 + threadIdx.x / 4;
+                return ((l / 2) * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 16) {
-                return ((l / 2) % 2) * 8 + threadIdx.x / 4;
+                return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -124,13 +211,16 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 8 && J == 8) {
-                return 4 * l + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 8) {
-                return 2 * (threadIdx.x % 4) + l % 2;
+                return ((threadIdx.x % 4) * 2) | (l % 2);
             } else if constexpr (I == 16 && J == 16) {
-                return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2;
+                return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #endif // defined(GGML_USE_HIP)
@@ -140,64 +230,179 @@ namespace ggml_cuda_mma {
     struct tile<I_, J_, half2> {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
+
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 8) {
+                return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+#else
+                return threadIdx.x;
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr ((I == 8 || I == 32) && J == 8) {
+                return l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#elif defined(AMD_WMMA_AVAILABLE)
+        static constexpr int ne = I * J / 32;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return 4 * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#else
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l & 2) * 2) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
     };
 
     template <int I_, int J_>
     struct tile<I_, J_, nv_bfloat162> {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
+
+#if defined(AMD_WMMA_AVAILABLE)
+        static constexpr int ne = I * J / 32;
+        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return 4 * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#else
         static constexpr int ne = I * J / WARP_SIZE;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  4) return true;
+            if (I == 16 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
+#endif  // defined(AMD_WMMA_AVAILABLE)
     };
 
     template <int I, int J>
@@ -231,6 +436,30 @@ namespace ggml_cuda_mma {
             const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
             xi[0] = xs[0];
         }
+#elif defined(AMD_WMMA_AVAILABLE)
+        if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
+
+        } else if constexpr (std::is_same_v<T, int>) {
+            if constexpr (I == 16 && J == 4) {
+                int64_t * xi = (int64_t *) t.x;
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
+                xi[0] = xs[0];
+
+            }else if constexpr (I == 16 && J == 8) {
+                int64_t * xi = (int64_t *) t.x;
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I));
+                xi[0] = xs[0];
+
+                const int64_t * xs1 = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I) + 2);
+                xi[1] = xs1[0];
+
+            }else{
+                NO_DEVICE_CODE;
+            }
+        } else {
+            NO_DEVICE_CODE;
+        }
 #else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
@@ -263,8 +492,12 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(xs0, stride);
-        GGML_UNUSED(t);
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
+#else
+        load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -277,11 +510,35 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
+#else
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #else
         load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
+    template <typename T>
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if 1
+        // TODO: more generic handling
+        static_assert(sizeof(T) == 4, "bad type size");
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#else
+        load_generic(t, xs0, stride);
+#endif // 1
+#else
+        tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
+        load_ldmatrix(t16[0], xs0 +  0*stride, stride);
+        load_ldmatrix(t16[1], xs0 + 16*stride, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+    }
+
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
@@ -489,12 +746,34 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#elif defined(AMD_WMMA_AVAILABLE)
+        using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
+        using floatx8_t = __attribute__((ext_vector_type(8))) float;
+        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
+        const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
+        const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
+    static __device__ __forceinline__ void mma(
+            tile<16, 16, float> & D, const tile<16, 8, nv_bfloat162> & A, const tile<16, 8, nv_bfloat162> & B) {
+#if defined(AMD_WMMA_AVAILABLE)
+        using bf16x8_t = __attribute__((ext_vector_type(8))) __bf16;
+        using floatx8_t = __attribute__((ext_vector_type(8))) float;
+        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
+        const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
+        const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // AMPERE_MMA_AVAILABLE
+    }
+
     static __device__ __forceinline__ void mma(
             tile<16, 16, int> & D, const tile<16, 8, int> & A, const tile<16, 8, int> & B) {
 #if defined(AMD_MFMA_AVAILABLE)
@@ -515,6 +794,36 @@ namespace ggml_cuda_mma {
                                                       acc[0],
                                                       0, 0, 0);
 #endif // defined(CDNA3)
+
+#elif defined(AMD_WMMA_AVAILABLE)
+        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+        int32x2_t * a_vec = (int32x2_t *) A.x;
+        int32x2_t * b_vec = (int32x2_t *) B.x;
+
+        using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
+        int32x8_t * acc = (int32x8_t *) D.x;
+
+#if defined(RDNA4)
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+            true,
+            a_vec[0],
+            true,
+            b_vec[0],
+            acc[0],
+            true
+        );
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+            true,
+            a_vec[1],
+            true,
+            b_vec[1],
+            acc[0],
+            true
+        );
+#endif // defined(RDNA4)
+
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -541,9 +850,76 @@ namespace ggml_cuda_mma {
                                                      acc[0],
                                                      0, 0, 0);
 #endif // defined(CDNA3)
+
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
+
+    template <typename T1, typename T2, int J, int K>
+    static __device__ __forceinline__ void mma(
+            tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
+        tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
+        tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
+#else
+        tile      <16, 8, float> * D16 = reinterpret_cast<tile      <16, 8, float> *>(&D);
+        const tile<16, 8, half2> * A16 = reinterpret_cast<const tile<16, 8, half2> *>(&A);
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+    }
+
+static __device__ __forceinline__ void mma(
+            tile<16, 16, int> & D, const tile<16, 4, int> & A, const tile<16, 4, int> & B) {
+#if defined(AMD_WMMA_AVAILABLE)
+    using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+    int32x2_t * a_vec = (int32x2_t *) A.x;
+    int32x2_t * b_vec = (int32x2_t *) B.x;
+
+    using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
+    int32x8_t * acc = (int32x8_t *) D.x;
+
+    acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
+        true,
+        a_vec[0],
+        true,
+        b_vec[0],
+        acc[0],
+        false
+    );
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif
+    }
 }
+
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cu b/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cu
index 9e2aaf52..be2ad1c6 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cu
@@ -119,15 +119,27 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     }
 }
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols, bool mul_mat_id) {
-
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne,
+        const size_t * src0_nb, const int src1_ncols, bool mul_mat_id) {
     if (ggml_is_quantized(type)) {
         return false;
     }
 
-    if (src0_ne[0] % (warp_size * (4/ggml_type_size(type))) != 0) {
+    const size_t ts = ggml_type_size(type);
+    if (src0_ne[0] % (warp_size * (4/ts)) != 0) {
         return false;
     }
+
+    if (src0_nb[0] != ts) {
+        return false;
+    }
+
+    // Pointers not aligned to the size of half2/nv_bfloat162/float2 would result in a crash:
+    for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+        if (src0_nb[i] % (2*ts) != 0) {
+            return false;
+        }
+    }
     if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
         return false;
     }
@@ -148,9 +160,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return turing_mma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc);
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc);
+            return ampere_mma_available(cc) || amd_wmma_available(cc);
         default:
             return false;
     }
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cuh
index 49d5295b..c2a0a2e4 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmf.cuh
@@ -2,6 +2,7 @@
 
 #include "mma.cuh"
 #include "common.cuh"
+#include "convert.cuh"
 
 using namespace ggml_cuda_mma;
 
@@ -17,7 +18,7 @@ struct mmf_ids_data {
 
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols, bool mul_mat_id);
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const size_t * src0_nb, const int src1_ncols, bool mul_mat_id);
 
 template <typename T, int rows_per_block, int cols_per_block, int nwarps, bool has_ids>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
@@ -27,10 +28,35 @@ static __global__ void mul_mat_f(
         const int stride_col_id, const int stride_row_id,
         const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+// TODO: handle this in a consistent and simpler way after AMD MFMA support has been added
+#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE)
+    // Special case for tf32, just dummy mma layout as wmma doesn't support it.
+    constexpr int tile_B_I = std::is_same_v<T, float> ? 8 : 16;
+    constexpr int tile_C_J = std::is_same_v<T, float> ? 8 : 16;
+    typedef tile<16,       8, T>     tile_A;
+    typedef tile<tile_B_I, 8, T>     tile_B;
+    typedef tile<16,       tile_C_J, float> tile_C;
+
+    constexpr bool a_supported = tile_A::supported();
+    constexpr bool b_supported = tile_B::supported();
+    constexpr bool c_supported = tile_C::supported();
+    constexpr bool supported = a_supported && b_supported && c_supported;
+#else
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+    constexpr bool supported = I_16_supported || I_32_supported;
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
+#endif // defined(AMD_WMMA_AVAILABLE)
+    if constexpr (!supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -151,11 +177,11 @@ static __global__ void mul_mat_f(
 
                     if constexpr (!has_ids) {
                         const float2 tmp = j < cols_per_block ? y2[j*stride_col_y + col] : make_float2(0.0f, 0.0f);
-                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = ggml_cuda_cast<T>(tmp);
                     } else {
                         const bool valid = j < cols_per_block && (col_base + j) < ncols_dst_total && slot_map[j] >= 0;
                         float2 tmp = valid ? *(const float2*) &y[slot_map[j]*stride_channel_y + 2*(j*stride_col_y + col)] : make_float2(0.0f, 0.0f);
-                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = ggml_cuda_cast<T>(tmp);
                     }
                 }
             } else {
@@ -229,10 +255,9 @@ static __global__ void mul_mat_f(
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     NO_DEVICE_CODE;
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
 }
 
-
 //This kernel is for larger batch sizes of mul_mat_id
 template <typename T, int rows_per_block, int cols_per_block, int nwarps>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
@@ -244,10 +269,35 @@ static __global__ void mul_mat_f_ids(
         const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const uint3 sis1_fd, const uint3 nch_fd) {
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+// TODO: handle this in a consistent and simpler way after AMD MFMA support has been added
+#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_WMMA_AVAILABLE)
+    // Special case for tf32, just dummy mma layout as wmma doesn't support it.
+    constexpr int tile_B_I = std::is_same_v<T, float> ? 8 : 16;
+    constexpr int tile_C_J = std::is_same_v<T, float> ? 8 : 16;
+    typedef tile<16,       8, T>     tile_A;
+    typedef tile<tile_B_I, 8, T>     tile_B;
+    typedef tile<16,       tile_C_J, float> tile_C;
+
+    constexpr bool a_supported = tile_A::supported();
+    constexpr bool b_supported = tile_B::supported();
+    constexpr bool c_supported = tile_C::supported();
+    constexpr bool supported = a_supported && b_supported && c_supported;
+#else
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+    constexpr bool supported = I_16_supported || I_32_supported;
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
+#endif // defined(AMD_WMMA_AVAILABLE)
+    if constexpr (!supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -389,7 +439,7 @@ static __global__ void mul_mat_f_ids(
 #pragma unroll
                 for (int j0 = 0; j0 < tile_B::I; ++j0) {
                     const float2 tmp = vals_buf[curr_buf][j0];
-                    tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = ggml_cuda_cast<T>(tmp);
                 }
 
                 if (itB + 1 < ntB) {
@@ -473,7 +523,7 @@ static __global__ void mul_mat_f_ids(
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, sis1_fd, nch_fd);
     NO_DEVICE_CODE;
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
 }
 
 template<typename T, int cols_per_block, int nwarps>
@@ -533,8 +583,10 @@ void mul_mat_f_cuda(
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         cudaStream_t stream, const mmf_ids_data * ids_data) {
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
+    typedef tile<16, 8, T>     tile_A_16;
+    typedef tile<32, 8, T>     tile_A_32;
+    typedef tile<16, 8, T>     tile_B_16;
+    typedef tile< 8, 8, T>     tile_B_8;
 
     GGML_ASSERT(ncols_x      % 2 == 0);
     GGML_ASSERT(stride_row   % 2 == 0);
@@ -544,7 +596,8 @@ void mul_mat_f_cuda(
     const int64_t channel_ratio = nchannels_dst / nchannels_x;
     const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
 
-    const int device = ggml_cuda_get_device();
+    const int device    = ggml_cuda_get_device();
+    const int cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
 
     int64_t nwarps_best     = 1;
@@ -559,8 +612,9 @@ void mul_mat_f_cuda(
     }
 
     constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
-    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
-    const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
+    const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + 4) * 4;
+    const int nbytes_cols_per_block_pad = amd_wmma_available(cc) ? tile_B_16::I : tile_B_8::I;
+    const int nbytes_shared_combine = GGML_PAD(cols_per_block, nbytes_cols_per_block_pad) * (nwarps_best*rows_per_block + 4) * 4;
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;
     const int nbytes_shared_total = nbytes_shared + nbytes_slotmap;
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu
index a2c8760a..03ceba87 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu
@@ -306,5 +306,11 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return false;
     }
 
-    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    if (amd_wmma_available(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+            return true;
+        }
+    }
+
+    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh
index c9a07e82..82468b38 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh
@@ -92,7 +92,7 @@ struct tile_x_sizes {
 };
 
 static int get_mmq_x_max_host(const int cc) {
-    return (amd_mfma_available(cc) || turing_mma_available(cc)) ? 128 :
+    return (amd_mfma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 :
         GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ?
 #ifdef GGML_CUDA_FORCE_MMQ
             128                     : 64;
@@ -102,7 +102,7 @@ static int get_mmq_x_max_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_x_max_device() {
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     return 128;
 #else // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
@@ -121,7 +121,7 @@ static constexpr __device__ int get_mmq_x_max_device() {
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
 #endif // defined(GGML_USE_HIP)
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }
 
 static int get_mmq_y_host(const int cc) {
@@ -231,7 +231,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
 #define MMQ_TILE_Y_K (MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI8_1)
 
 static int mmq_get_granularity_host(const int mmq_x, const int cc) {
-    if (amd_mfma_available(cc)) {
+    if (amd_mfma_available(cc) || amd_wmma_available(cc)) {
         return mmq_x >= 128 ? 32 : 16;
     } else if (turing_mma_available(cc) && mmq_x >= 48) {
         return 16;
@@ -240,7 +240,7 @@ static int mmq_get_granularity_host(const int mmq_x, const int cc) {
     }
 }
 
-#if defined(AMD_MFMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) {
     return mmq_x >= 128 ? 32 : 16;
 }
@@ -265,7 +265,7 @@ static int mmq_get_nwarps_host(const int /*cc*/, const int warp_size) {
 #endif // (GGML_USE_HIP)
 
 static constexpr __device__ int mmq_get_nwarps_device() {
-#if defined(AMD_MFMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     return 8;
 #else
     return 256/ggml_cuda_get_physical_warp_size();
@@ -279,14 +279,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0);
     constexpr int nrows = warp_size / threads_per_row;
@@ -305,7 +305,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx;
         const int qs0 = get_int_b2(bxi->qs, kqsx);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + 0]     = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808);
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808);
 #else
@@ -327,11 +327,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0           + kbxd] = bxi->d;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -382,14 +382,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)  || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1);
     constexpr int nrows = warp_size / threads_per_row;
@@ -408,12 +408,12 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx;
         const int qs0 = get_int_b4(bxi->qs, kqsx);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + 0]     = (qs0 >> 0) & 0x0F0F0F0F;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F;
 #else
         x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1;
@@ -430,11 +430,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_dm[i*MMQ_MMA_TILE_X_K_Q8_1           + kbxd] = bxi->dm;
 #else
         x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -485,14 +485,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0);
     constexpr int nrows = warp_size / threads_per_row;
@@ -527,13 +527,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         qs1    |= (qh <<  9)   & 0x10000000;  // 19 -> 28
         qs1     = __vsubss4(qs1, 0x10101010); // subtract 16
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0;
@@ -550,11 +550,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0           + kbxd] = bxi->d;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)  || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -563,14 +563,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1);
     constexpr int nrows = warp_size / threads_per_row;
@@ -603,13 +603,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         qs1    |= (qh <<  2) & 0x00100000; // 18 -> 20
         qs1    |= (qh <<  9) & 0x10000000; // 19 -> 28
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1;
@@ -626,11 +626,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_dm[i*MMQ_MMA_TILE_X_K_Q8_1           + kbxd] = bxi->dm;
 #else
         x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -639,14 +639,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp
     constexpr int threads_per_row = 32;
@@ -665,13 +665,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 0             + txi] = get_int_b2(bxi[0].qs,                   kqsx);
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx);
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0             + txi] = get_int_b2(bxi[0].qs,                   kqsx);
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0;
@@ -688,11 +688,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0                 + kbxd] = bxi->d;
 #else
         x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -701,14 +701,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4);
     constexpr int nrows = warp_size / threads_per_row;
@@ -730,13 +730,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4);
         const int k0 = kbx * (2 * QI_MXFP4) + kqsx;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + 0]        = v.x;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + QI_MXFP4] = v.y;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0]        = v.x;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)  || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4;
@@ -753,11 +753,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_1                 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -796,7 +796,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
 template <int mmq_x, int mmq_y, mmq_q8_1_ds_layout ds_layout>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     typedef tile<16,  8, int> tile_A;
     typedef tile<16,  8, int> tile_B;
     typedef tile<16, 16, int> tile_C;
@@ -927,7 +927,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
             }
         }
     }
-#endif // defined(AMD_MFMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }
 
 template <int mmq_x, int mmq_y>
@@ -965,7 +965,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     typedef tile<16,  8, int> tile_A;
     typedef tile<16,  8, int> tile_B;
     typedef tile<16, 16, int> tile_C;
@@ -1087,7 +1087,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
             }
         }
     }
-#endif // defined(AMD_MFMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }
 
 // Used for Q3_K, IQ2_S, and IQ2_XS
@@ -1170,6 +1170,54 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
                 tile_C C;
                 mma(C, A[n], B[0]);
 
+#pragma unroll
+                for (int l = 0; l < tile_C::ne; ++l) {
+                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
+                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4] * dB;
+                }
+            }
+        }
+    }
+#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+    typedef tile<16,  4, int> tile_A;
+    typedef tile<16,  4, int> tile_B;
+    typedef tile<16, 16, int> tile_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = granularity;
+    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
+
+    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
+        const int k0 = k00 + k01;
+
+        tile_A A[ntx];
+#pragma unroll
+        for (int n = 0; n < ntx; ++n) {
+            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
+        }
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
+            tile_B B;
+            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+
+            const int j = j0 + tile_C::get_j(0);
+            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                tile_C C;
+                mma(C, A[n], B);
+
 #pragma unroll
                 for (int l = 0; l < tile_C::ne; ++l) {
                     const int i = i0 + n*tile_C::I + tile_C::get_i(l);
@@ -1257,21 +1305,21 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
 #else
     GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
-#endif // AMD_MFMA_AVAILABLE
+#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE
 }
 
 template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
     const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
     constexpr int nwarps = mmq_get_nwarps_device();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K);
     constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row;
@@ -1295,11 +1343,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
             const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
         const int sc_m = bxi->scales[kqsx];
@@ -1310,11 +1358,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4));
 #endif // FAST_FP16_AVAILABLE
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + kqsx] = x_dm_ik;
 #else
         x_dm[i*(MMQ_TILE_NE_K + 1)   + kqsx] = x_dm_ik;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -1438,6 +1486,72 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
                 tile_C Cd;
                 mma(Cd, A[n], B[0]);
 
+#pragma unroll
+                for (int l = 0; l < tile_C::ne; ++l) {
+                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
+                    const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/4]);
+                    float tmp = Cd.x[l]*dm.x;
+                    if (k01 >= MMQ_TILE_NE_K * 3/4) {
+                        tmp -= Cm.x[l]*dm.y;
+                    }
+                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB;
+                    sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB;
+                }
+            }
+        }
+    }
+#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+
+    typedef tile<16,  4, int> tile_A;
+    typedef tile<16,  4, int> tile_B;
+    typedef tile<16, 16, int> tile_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = granularity;
+    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
+
+    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
+        const int k0 = k00 + k01;
+
+        tile_A A[ntx];
+#pragma unroll
+        for (int n = 0; n < ntx; ++n) {
+            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
+        }
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
+            tile_B B;
+            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+
+            const int j = j0 + tile_C::get_j(0);
+            const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y;
+            const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0
+                                              : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y
+                                                             : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x);
+
+            tile_C Cm;
+            if (k01 >= MMQ_TILE_NE_K * 3/4) {
+                tile_A A1;
+                A1.x[0] = 0x01010101;
+                A1.x[1] = 0x01010101;
+                mma(Cm, A1, B);
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                tile_C Cd;
+                mma(Cd, A[n], B);
+
 #pragma unroll
                 for (int l = 0; l < tile_C::ne; ++l) {
                     const int i = i0 + n*tile_C::I + tile_C::get_i(l);
@@ -1574,7 +1688,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
 #else
     GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
-#endif // AMD_MFMA_AVAILABLE
+#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE
 }
 
 template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
@@ -1582,7 +1696,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
@@ -1618,11 +1732,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
             const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k] = x_qs_k;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
     }
 
@@ -1649,7 +1763,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         const int8_t * sc8 = (const int8_t *) &sc;
         const float d = bxi->d;
 
@@ -1659,10 +1773,10 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         }
 #else
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
-#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
+#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE))
 #pragma unroll
     for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) {
         int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y;
@@ -1675,7 +1789,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         x_df[i] = bxi->d;
     }
-#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
+#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) || defined(AMD_WMMA_AVAILABLE)
 }
 
 template <int mmq_x, int mmq_y>
@@ -1728,7 +1842,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
@@ -1736,7 +1850,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
     int   * x_sc = (int   *) (x_dm + txs.dm);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K);
     constexpr int nrows = warp_size / threads_per_row;
@@ -1753,19 +1867,19 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
         const int qs0 = get_int_b4(bxi->qs, txi);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F;
 #else
         x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr int rows_per_warp = warp_size / 2;
 #pragma unroll
     for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) {
-#if defined(AMD_MFMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         // Need if on AMD instead of % because warp_size == 64
         // This causes double work and throughput loss (MI300X)
         // H100 loses about 100 t/s with 'if' condition over '%'
@@ -1774,7 +1888,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #else
         int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y;
         {
-#endif // defined(AMD_MFMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             if (need_check) {
                 i = min(i, i_max);
             }
@@ -1829,7 +1943,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8;
     }
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }
 
 template <int mmq_x, int mmq_y>
@@ -1872,7 +1986,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2);
 #else
@@ -1908,16 +2022,16 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0;
         const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq0] = ql0 | qh0;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq1] = ql1 | qh1;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr int rows_per_warp = warp_size / 2;
 #pragma unroll
     for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) {
@@ -1930,7 +2044,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #else
         int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y;
         {
-#endif // defined(AMD_MFMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             if (need_check) {
                 i = min(i, i_max);
             }
@@ -1986,7 +2100,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8;
     }
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }
 
 template <int mmq_x, int mmq_y>
@@ -2029,7 +2143,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
     int   * x_sc = (int   *) (x_df + MMQ_TILE_NE_K/QI6_K);
@@ -2038,7 +2152,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
     int   * x_sc = (int   *) (x_df + txs.dm);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2065,13 +2179,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int kq0 = 2*txi - txi % (QI6_K/2) + 0;
         const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
         x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
 #pragma unroll
@@ -2084,11 +2198,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q6_K]           = bxi->d;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int rows_per_warp = warp_size / 4;
@@ -2102,11 +2216,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8));
 #else
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8));
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2190,6 +2304,56 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
                 tile_C C;
                 mma(C, A[n], B[0]);
 
+#pragma unroll
+                for (int l = 0; l < tile_C::ne; ++l) {
+                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
+                    const int8_t * sc = (const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q6_K + k00/16);
+                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*MMQ_MMA_TILE_X_K_Q6_K] * dB;
+                }
+            }
+        }
+    }
+#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+    typedef tile<16,  4, int> tile_A;
+    typedef tile<16,  4, int> tile_B;
+    typedef tile<16, 16, int> tile_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = granularity;
+    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
+    const int   * x_sc = (const int   *) x_df + MMQ_TILE_NE_K/QI6_K;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
+
+    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
+        const int k0 = k00 + k01;
+
+        tile_A A[ntx];
+#pragma unroll
+        for (int n = 0; n < ntx; ++n) {
+            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
+        }
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
+            tile_B B;
+            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+
+            const int j = j0 + tile_C::get_j(0);
+            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                tile_C C;
+                mma(C, A[n], B);
+
 #pragma unroll
                 for (int l = 0; l < tile_C::ne; ++l) {
                     const int i = i0 + n*tile_C::I + tile_C::get_i(l);
@@ -2303,7 +2467,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
 #else
     GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
-#endif // AMD_MFMA_AVAILABLE
+#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE
 }
 
 template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_nl(
@@ -2311,14 +2475,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2340,13 +2504,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
         const int k0 = kbx * (2 * QI4_NL) + kqsx;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0]      = v.x;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + QI4_NL] = v.y;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0]      = v.x;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL;
@@ -2363,11 +2527,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0             + kbxd] = __half2float(bxi->d);
 #else
         x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2376,14 +2540,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2414,22 +2578,22 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
             const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid1;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
         const int ls = aux32 >> 28;
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = (ls*d + d/2)/4;
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/4;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)  || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2438,14 +2602,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2472,24 +2636,24 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
             const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
         const int ls = bxi->scales[kqsx];
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
 #else
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2498,15 +2662,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
-
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
     const int kqsx = threadIdx.x % threads_per_row;
@@ -2539,24 +2702,24 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0);
             const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
         const int ls = bxi->scales[kqsx];
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
 #else
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2565,14 +2728,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2601,22 +2764,22 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
             const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
         const int ls = aux32 >> 28;
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0     + kqsx] = (ls*d + d/2)/2;
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4   + kqsx] = (ls*d + d/2)/2;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2625,14 +2788,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2668,22 +2831,22 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
             const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
         const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F);
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0     + kqsx] = ls*d;
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4   + kqsx] = ls*d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2692,14 +2855,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_ds = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2727,23 +2890,23 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid0 = (grid >> 0) & 0x0F0F0F0F;
             const int grid1 = (grid >> 4) & 0x0F0F0F0F;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+0)] = grid0;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+1)] = grid1;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
         const float  d1q   = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1);
         const float  delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_ds[i*MMQ_MMA_TILE_X_K_Q8_1     + kqsx] = make_half2(d1q, d1q*delta);
 #else
         x_ds[i*(MMQ_TILE_NE_K/4) + i/4   + kqsx] = make_half2(d1q, d1q*delta);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2752,14 +2915,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2779,13 +2942,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
         const int k0 = 8 * (kqsx / 4) + kqsx % 4;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 
     constexpr int rows_per_warp = warp_size / 8;
@@ -2804,11 +2967,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F)
             | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + threadIdx.x % 8] = d * (ls - 32);
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     }
 }
 
@@ -2848,7 +3011,7 @@ static __device__ __forceinline__ void mmq_write_back_mma(
     constexpr int granularity = mmq_get_granularity_device(mmq_x);
     constexpr int nwarps = mmq_get_nwarps_device();
 
-#if defined(AMD_MFMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr int tileC_IJ = mmq_get_granularity_device(0);
     typedef tile<tileC_IJ, tileC_IJ, int> tile_C;
     constexpr int rows_per_warp = granularity;
@@ -2859,11 +3022,11 @@ static __device__ __forceinline__ void mmq_write_back_mma(
     constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
 
     const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I);
-#if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y");
 #else
     GGML_UNUSED(nwarps);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
@@ -3063,13 +3226,13 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
     int * tile_y = data_mul_mat_q + mmq_x;
     int * tile_x = tile_y + GGML_PAD(mmq_x*MMQ_TILE_Y_K, nwarps*warp_size);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, need_check, type>::vec_dot_mma;
     constexpr mmq_write_back_t write_back = mmq_write_back_mma<type, mmq_x, mmq_y, need_check>;
 #else
     constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, need_check, type>::vec_dot_dp4a;
     constexpr mmq_write_back_t write_back = mmq_write_back_dp4a<mmq_x, mmq_y, need_check>;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
     constexpr int blocks_per_iter = MMQ_ITER_K / qk;
 
@@ -3494,7 +3657,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
     const int col_diff = col_high - col_low;
 
     for (int j = threadIdx.y*warp_size + threadIdx.x; j < mmq_x; j += nwarps*warp_size) {
-        ids_dst_shared[j] = ids_dst[col_low + j];
+        ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
     }
     __syncthreads();
 
@@ -3538,7 +3701,7 @@ static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int
     const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
     const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
     const size_t nbs_ids = mmq_x*sizeof(int);
-    const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
+    const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc) || amd_wmma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
     const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq);
     return nbs_ids + nbs_x + GGML_PAD(nbs_y, nwarps*warp_size*sizeof(int));
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cu b/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cu
index 57ab8393..6238ce7e 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cu
@@ -1,11 +1,12 @@
 #include "ggml.h"
 #include "common.cuh"
-#include "convert.cuh"
+#include "unary.cuh"
 #include "mmvf.cuh"
+#include "convert.cuh"
 
-template <typename T, typename type_acc, int ncols_dst, int block_size>
+template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false>
 static __global__ void mul_mat_vec_f(
-        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
         const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
@@ -24,58 +25,164 @@ static __global__ void mul_mat_vec_f(
     y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
     dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
 
+    bool use_gate = false;
+    bool use_bias = false;
+    bool use_gate_bias = false;
+    ggml_glu_op glu_op = ggml_glu_op::GGML_GLU_OP_SWIGLU;
+    const T * gate_x = nullptr;
+    const float * x_bias = nullptr;
+    const float * gate_bias = nullptr;
+
+    if constexpr (has_fusion) {
+        use_gate = fusion.gate != nullptr;
+        use_bias = fusion.x_bias != nullptr;
+        use_gate_bias = fusion.gate_bias != nullptr;
+        glu_op = fusion.glu_op;
+
+        if (use_gate) {
+            gate_x = static_cast<const T *>(fusion.gate);
+        }
+        if (use_bias) {
+            x_bias = static_cast<const float *>(fusion.x_bias);
+        }
+        if (use_gate_bias) {
+            gate_bias = static_cast<const float *>(fusion.gate_bias);
+            use_gate_bias = use_gate;
+        } else {
+            use_gate_bias = false;
+        }
+    }
+
+    if (use_gate) {
+        gate_x += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    }
+    if constexpr (has_fusion) {
+        const int channel_bias = ids ? channel_x : channel_dst;
+        if (use_bias) {
+            x_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
+        }
+        if (use_gate_bias) {
+            gate_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
+        }
+    }
+
     const float2 * y2 = (const float2 *) y;
 
     extern __shared__ char data_mmv[];
     float * buf_iw = (float *) data_mmv;
+    float * buf_iw_gate = nullptr;
+    if constexpr (has_fusion) {
+        buf_iw_gate = (float *) (data_mmv + warp_size*sizeof(float));
+    }
 
     if (block_size > warp_size) {
         if (tid < warp_size) {
             buf_iw[tid] = 0.0f;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    buf_iw_gate[tid] = 0.0f;
+                }
+            }
         }
         __syncthreads();
     }
 
     float sumf[ncols_dst] = {0.0f};
+    float sumf_gate[ncols_dst];
+    if constexpr (has_fusion) {
+#pragma unroll
+        for (int j = 0; j < ncols_dst; ++j) {
+            sumf_gate[j] = 0.0f;
+        }
+    }
 
     if constexpr (std::is_same_v<T, float>) {
         const float2 * x2 = (const float2 *) x;
+        const float2 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const float2 *) gate_x;
+            }
+        }
 
         for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const float2 tmpx = x2[col2];
+            float2 tmpx_gate = make_float2(0.0f, 0.0f);
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
 
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
                 ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
                 ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                    }
+                }
             }
         }
     } else if constexpr (std::is_same_v<T, half>) {
         const half2 * x2 = (const half2 *) x;
+        const half2 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const half2 *) gate_x;
+            }
+        }
 
         if (std::is_same_v<type_acc, float>) {
             for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const float2 tmpx = __half22float2(x2[col2]);
-
+                float2 tmpx_gate = make_float2(0.0f, 0.0f);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmpx_gate = __half22float2(gate_x2[col2]);
+                    }
+                }
 #pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
                     const float2 tmpy = y2[j*stride_col_y2 + col2];
                     ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
                     ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                            ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                        }
+                    }
                 }
             }
         } else {
 #ifdef FP16_AVAILABLE
             half2 sumh2[ncols_dst] = {{0.0f, 0.0f}};
+            half2 sumh2_gate[ncols_dst] = {{0.0f, 0.0f}};
 
             for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const half2 tmpx = x2[col2];
-
+                half2 tmpx_gate = make_half2(0.0f, 0.0f);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmpx_gate = gate_x2[col2];
+                    }
+                }
 #pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
                     const float2 tmpy = y2[j*stride_col_y2 + col2];
                     sumh2[j] += tmpx * make_half2(tmpy.x, tmpy.y);
+
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            sumh2_gate[j] += tmpx_gate * make_half2(tmpy.x, tmpy.y);
+                        }
+                    }
                 }
             }
 
@@ -83,6 +190,15 @@ static __global__ void mul_mat_vec_f(
             for (int j = 0; j < ncols_dst; ++j) {
                 sumf[j] = __low2float(sumh2[j]) + __high2float(sumh2[j]);
             }
+
+            if constexpr (has_fusion) {
+                if (use_gate) {
+#pragma unroll
+                    for (int j = 0; j < ncols_dst; ++j) {
+                        sumf_gate[j] = __low2float(sumh2_gate[j]) + __high2float(sumh2_gate[j]);
+                    }
+                }
+            }
 #else
             NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
@@ -91,8 +207,20 @@ static __global__ void mul_mat_vec_f(
 //TODO: add support for ggml_cuda_mad for hip_bfloat162
 #if defined(GGML_USE_HIP)
         const int * x2 = (const int *) x;
+        const int * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const int *) gate_x;
+            }
+        }
         for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const int tmpx = x2[col2];
+            int tmpx_gate = 0;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
@@ -100,17 +228,45 @@ static __global__ void mul_mat_vec_f(
                 const float tmpx1 = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]);
                 ggml_cuda_mad(sumf[j], tmpx0, tmpy.x);
                 ggml_cuda_mad(sumf[j], tmpx1, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        const float tmpx0_gate = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx_gate)[0]);
+                        const float tmpx1_gate = ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx_gate)[1]);
+                        ggml_cuda_mad(sumf_gate[j], tmpx0_gate, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx1_gate, tmpy.y);
+                    }
+                }
             }
         }
 #else
         const nv_bfloat162 * x2 = (const nv_bfloat162 *) x;
+        const nv_bfloat162 * gate_x2 = nullptr;
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                gate_x2 = (const nv_bfloat162 *) gate_x;
+            }
+        }
         for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const nv_bfloat162 tmpx = x2[col2];
+            nv_bfloat162 tmpx_gate;
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = gate_x2[col2];
+                }
+            }
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
                 ggml_cuda_mad(sumf[j], tmpx.x, tmpy.x);
                 ggml_cuda_mad(sumf[j], tmpx.y, tmpy.y);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.x, tmpy.x);
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate.y, tmpy.y);
+                    }
+                }
             }
         }
 #endif
@@ -122,13 +278,31 @@ static __global__ void mul_mat_vec_f(
     for (int j = 0; j < ncols_dst; ++j) {
         sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
 
+        if constexpr (has_fusion) {
+            if (use_gate) {
+                sumf_gate[j] = warp_reduce_sum<warp_size>(sumf_gate[j]);
+            }
+        }
+
         if (block_size > warp_size) {
             buf_iw[tid/warp_size] = sumf[j];
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    buf_iw_gate[tid/warp_size] = sumf_gate[j];
+                }
+            }
             __syncthreads();
             if (tid < warp_size) {
                 sumf[j] = buf_iw[tid];
                 sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        sumf_gate[j] = buf_iw_gate[tid];
+                        sumf_gate[j] = warp_reduce_sum<warp_size>(sumf_gate[j]);
+                    }
+                }
             }
+
             if (j < ncols_dst) {
                 __syncthreads();
             }
@@ -139,12 +313,74 @@ static __global__ void mul_mat_vec_f(
         return;
     }
 
-    dst[tid*stride_col_dst + row] = sumf[tid];
+    float value = sumf[tid];
+
+    if constexpr (has_fusion) {
+        if (use_bias) {
+            value += x_bias[tid*stride_col_dst + row];
+        }
+
+        if (use_gate) {
+            float gate_value = sumf_gate[tid];
+            if (use_gate_bias) {
+                gate_value += gate_bias[tid*stride_col_dst + row];
+            }
+            switch (glu_op) {
+                case GGML_GLU_OP_SWIGLU:
+                    value *= ggml_cuda_op_silu_single(gate_value);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                    value *= ggml_cuda_op_gelu_single(gate_value);
+                    break;
+                case GGML_GLU_OP_SWIGLU_OAI: {
+                    value = ggml_cuda_op_swiglu_oai_single(gate_value, value);
+                    break;
+                }
+                default:
+                    break;
+            }
+        }
+    }
+
+    dst[tid*stride_col_dst + row] = value;
+
+    if constexpr (!has_fusion) {
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, glu_op, gate_x, x_bias, gate_bias, sumf_gate);
+    }
+}
+
+template<typename T, typename type_acc, int ncols_dst, int block_size>
+static void mul_mat_vec_f_switch_fusion(
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
+        const int64_t ncols, const int64_t nrows,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const cudaStream_t stream) {
+
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    if constexpr (ncols_dst == 1) {
+        if (has_fusion) {
+            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            return;
+       }
+    }
+
+    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
+
+    mul_mat_vec_f<T, type_acc, ncols_dst, block_size><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+
 }
 
 template <typename T, typename type_acc, int ncols_dst>
-static void launch_mul_mat_vec_f_cuda(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+void launch_mul_mat_vec_f_cuda(
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
@@ -176,57 +412,59 @@ static void launch_mul_mat_vec_f_cuda(
         }
     }
 
-    const int nbytes_shared = warp_size*sizeof(float);
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+
+    const int nbytes_shared = warp_size*sizeof(float) + (has_fusion ? warp_size*sizeof(float) : 0);
     const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case   64: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case   96: {
-            mul_mat_vec_f<T, type_acc, ncols_dst,  96><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  128: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  160: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 160><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  192: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 192><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  224: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 224><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         case  256: {
-            mul_mat_vec_f<T, type_acc, ncols_dst, 256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -236,7 +474,7 @@ static void launch_mul_mat_vec_f_cuda(
 
 template <typename T, typename type_acc>
 static void mul_mat_vec_f_cuda_switch_ncols_dst(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
@@ -246,49 +484,49 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
     switch (ncols_dst) {
         case 1:
             launch_mul_mat_vec_f_cuda<T, type_acc, 1>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 2:
             launch_mul_mat_vec_f_cuda<T, type_acc, 2>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 3:
             launch_mul_mat_vec_f_cuda<T, type_acc, 3>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 4:
             launch_mul_mat_vec_f_cuda<T, type_acc, 4>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 5:
             launch_mul_mat_vec_f_cuda<T, type_acc, 5>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 6:
             launch_mul_mat_vec_f_cuda<T, type_acc, 6>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 7:
             launch_mul_mat_vec_f_cuda<T, type_acc, 7>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case 8:
             launch_mul_mat_vec_f_cuda<T, type_acc, 8>
-                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
@@ -300,29 +538,31 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
 
 template<typename T>
 static void mul_mat_vec_f_cuda(
-        const T * x, const float * y, const int32_t * ids, float * dst,
+        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
         const int64_t stride_row, const int64_t stride_col_y, const int stride_col_dst,
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         enum ggml_prec prec, cudaStream_t stream) {
+
     if constexpr(std::is_same_v<T, half>) {
         if (prec == GGML_PREC_DEFAULT) {
             mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
-                (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
-                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             return;
         }
     }
     mul_mat_vec_f_cuda_switch_ncols_dst<T, float>
-        (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
-         nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-         stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+        (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+        nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
 }
 
-void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+    const ggml_cuda_mm_fusion_args_host * fusion) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
     GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
     GGML_ASSERT(         dst->type == GGML_TYPE_F32);
@@ -348,6 +588,30 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
+
+    if (fusion) {
+        GGML_ASSERT( !ids || dst->ne[2] == 1);
+        GGML_ASSERT(  ids || dst->ne[1] == 1);
+        if (fusion->x_bias) {
+            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->x_bias->ne[1] == src0->ne[2]);
+            fusion_local.x_bias = fusion->x_bias->data;
+        }
+        if (fusion->gate) {
+            GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
+            fusion_local.gate = fusion->gate->data;
+        }
+        if (fusion->gate_bias) {
+            GGML_ASSERT(fusion->gate_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->gate_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
+            fusion_local.gate_bias = fusion->gate_bias->data;
+        }
+        fusion_local.glu_op = fusion->glu_op;
+    }
+
     const int64_t s01 = src0->nb[1] / ts_src0;
     const int64_t s11 = src1->nb[1] / ts_src1;
     const int64_t s1  =  dst->nb[1] / ts_dst;
@@ -370,19 +634,19 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
@@ -409,7 +673,6 @@ void ggml_cuda_op_mul_mat_vec_f(
     const int cc = ggml_cuda_info().devices[id].cc;
     const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
 
-
     // ggml_cuda_op provides single, contiguous matrices
     const int64_t stride_row         = ne00;
     const int64_t stride_col_y       = ne10;
@@ -426,22 +689,23 @@ void ggml_cuda_op_mul_mat_vec_f(
     const int64_t stride_sample_y    = 0;
     const int64_t stride_sample_dst  = 0;
 
+    ggml_cuda_mm_fusion_args_device empty{};
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
-            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
@@ -452,10 +716,23 @@ void ggml_cuda_op_mul_mat_vec_f(
     GGML_UNUSED_VARS(ctx, src1, dst, src1_ddq_i, src1_ncols, src1_padded_row_size);
 }
 
-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11) {
     if (src0_ne[0] % 2 != 0) {
         return false;
     }
+
+    const size_t ts = ggml_type_size(type);
+    if (src0_nb[0] != ts) {
+        return false;
+    }
+
+    // Pointers not aligned to the size of half2/nv_bfloat162/float2 would result in a crash:
+    for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+        if (src0_nb[i] % (2*ts) != 0) {
+            return false;
+        }
+    }
+
     switch (type) {
         case GGML_TYPE_F32:
             if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cuh
index 1da46099..a09fbdc7 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmvf.cuh
@@ -1,6 +1,7 @@
 #include "common.cuh"
 
-void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+    const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 
 void ggml_cuda_op_mul_mat_vec_f(
     ggml_backend_cuda_context & ctx,
@@ -8,4 +9,4 @@ void ggml_cuda_op_mul_mat_vec_f(
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
 
-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11);
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu b/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu
index 3bf0c9ed..d671551c 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cu
@@ -1,5 +1,6 @@
 #include "mmvq.cuh"
 #include "quantize.cuh"
+#include "unary.cuh"
 #include "vecdotq.cuh"
 
 #include <cstdint>
@@ -82,7 +83,7 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
-static constexpr __host__ __device__ int calc_nwarps(int ncols_dst,  mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
         switch (ncols_dst) {
             case 1:
@@ -136,11 +137,11 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-template <ggml_type type, int ncols_dst>
 // tell the compiler to use as many registers as it wants, see nwarps definition below
+template <ggml_type type, int ncols_dst, bool has_fusion>
 __launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
-        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
@@ -169,8 +170,56 @@ static __global__ void mul_mat_vec_q(
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
 
+    bool use_gate = false;
+    bool use_bias = false;
+    bool use_gate_bias = false;
+    const void * vgate = nullptr;
+    const float * x_bias = nullptr;
+    const float * gate_bias = nullptr;
+    ggml_glu_op active_glu;
+
+    if constexpr (has_fusion) {
+        use_gate      = fusion.gate      != nullptr;
+        use_bias      = fusion.x_bias    != nullptr;
+        use_gate_bias = fusion.gate_bias != nullptr && use_gate;
+        vgate         = fusion.gate;
+        x_bias        = (const float *) fusion.x_bias;
+        gate_bias     = (const float *) fusion.gate_bias;
+        active_glu    = fusion.glu_op;
+    }
+
+    const uint32_t channel_bias = ids ? channel_x : channel_dst;
+
+    float x_biases[ncols_dst]    = { 0.0f };
+    float gate_biases[ncols_dst] = { 0.0f };
+    if constexpr (has_fusion) {
+        if (use_bias) {
+            x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            // 1. Hide latency by prefetching bias and gate here
+            // 2. load only on threads that won't die after partial sum calculation
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    x_biases[j] = x_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
+        }
+        if (use_gate_bias) {
+            gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    gate_biases[j] = gate_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
+        }
+    }
+
     // partial sum for each thread
     float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
+    float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
@@ -187,17 +236,35 @@ static __global__ void mul_mat_vec_q(
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp[j][i] += vec_dot_q_cuda(
                     vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_gate[j][i] += vec_dot_q_cuda(
+                            vgate, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
+                    }
+                }
             }
         }
     }
 
     __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
+    __shared__ float tmp_shared_gate[(has_fusion && (nwarps-1 > 0)) ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
+    if constexpr (!has_fusion) {
+        (void) tmp_shared_gate;
+    } else if (!use_gate) {
+        (void) tmp_shared_gate;
+    }
+
     if (threadIdx.y > 0) {
 #pragma unroll
         for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_shared_gate[threadIdx.y-1][j][i][threadIdx.x] = tmp_gate[j][i];
+                    }
+                }
             }
         }
     }
@@ -216,14 +283,55 @@ static __global__ void mul_mat_vec_q(
 #pragma unroll
             for (int l = 0; l < nwarps-1; ++l) {
                 tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        tmp_gate[j][i] += tmp_shared_gate[l][j][i][threadIdx.x];
+                    }
+                }
             }
             tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmp_gate[j][i] = warp_reduce_sum<warp_size>(tmp_gate[j][i]);
+                }
+            }
         }
 
         if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
-            dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
+            float result = tmp[j][threadIdx.x];
+            if constexpr (has_fusion) {
+                if (use_bias) {
+                    result += x_biases[j];
+                }
+                if (use_gate) {
+                    float gate_value = tmp_gate[j][threadIdx.x];
+                    if (use_gate_bias) {
+                        gate_value += gate_biases[j];
+                    }
+                    switch (active_glu) {
+                        case GGML_GLU_OP_SWIGLU:
+                            result *= ggml_cuda_op_silu_single(gate_value);
+                            break;
+                        case GGML_GLU_OP_GEGLU:
+                            result *= ggml_cuda_op_gelu_single(gate_value);
+                            break;
+                        case GGML_GLU_OP_SWIGLU_OAI: {
+                            result = ggml_cuda_op_swiglu_oai_single(gate_value, result);
+                            break;
+                        }
+                        default:
+                            result = result * gate_value;
+                            break;
+                    }
+                }
+            }
+            dst[j*stride_col_dst + threadIdx.x] = result;
         }
     }
+
+    if constexpr (!has_fusion) {
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, active_glu, gate_bias, x_bias, tmp_gate);
+    }
 }
 
 static std::pair<dim3, dim3> calc_launch_params(
@@ -235,9 +343,37 @@ static std::pair<dim3, dim3> calc_launch_params(
     return {block_nums, block_dims};
 }
 
+template<ggml_type type, int c_ncols_dst>
+static void mul_mat_vec_q_switch_fusion(
+        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
+        const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
+        const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
+        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
+
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    if constexpr (c_ncols_dst == 1) {
+        if (has_fusion) {
+            mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            return;
+        }
+    }
+
+    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
+
+    mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+}
+
 template <ggml_type type>
 static void mul_mat_vec_q_switch_ncols_dst(
-        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int ncols_x, const int nrows_x, const int ncols_dst,
         const int stride_row_x, const int stride_col_y, const int stride_col_dst,
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
@@ -256,80 +392,83 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
     const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+
     GGML_ASSERT(!ids || ncols_dst == 1);
     switch (ncols_dst) {
         case 1: {
             constexpr int c_ncols_dst = 1;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 2: {
             constexpr int c_ncols_dst = 2;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 3: {
             constexpr int c_ncols_dst = 3;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 4: {
             constexpr int c_ncols_dst = 4;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 5: {
             constexpr int c_ncols_dst = 5;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 6: {
             constexpr int c_ncols_dst = 6;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 7: {
             constexpr int c_ncols_dst = 7;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         case 8: {
             constexpr int c_ncols_dst = 8;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 dims.first, dims.second, 0, stream);
         } break;
         default:
             GGML_ABORT("fatal error");
             break;
     }
-}
 
+    GGML_UNUSED(has_fusion);
+}
 static void mul_mat_vec_q_switch_type(
-        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, float * dst,
+        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int ncols_x, const int nrows_x, const int ncols_dst,
         const int stride_row_x, const int stride_col_y, const int stride_col_dst,
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
@@ -339,143 +478,123 @@ static void mul_mat_vec_q_switch_type(
     switch (type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q4_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q8_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_MXFP4:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q3_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q4_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q5_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_Q6_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ2_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ3_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ1_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ1_M:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ4_NL:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ4_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         case GGML_TYPE_IQ3_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
-                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -484,7 +603,8 @@ static void mul_mat_vec_q_switch_type(
 }
 
 void ggml_cuda_mul_mat_vec_q(
-        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+        const ggml_cuda_mm_fusion_args_host * fusion) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
     GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
     GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
@@ -508,6 +628,31 @@ void ggml_cuda_mul_mat_vec_q(
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
+
+    if (fusion) {
+        GGML_ASSERT( !ids || dst->ne[2] == 1);
+        GGML_ASSERT(  ids || dst->ne[1] == 1);
+
+        if (fusion->x_bias) {
+            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->x_bias->ne[1] == src0->ne[2]);
+            fusion_local.x_bias = fusion->x_bias->data;
+        }
+        if (fusion->gate) {
+            GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
+            fusion_local.gate = fusion->gate->data;
+        }
+        if (fusion->gate_bias) {
+            GGML_ASSERT(fusion->gate_bias->type == GGML_TYPE_F32);
+            GGML_ASSERT(fusion->gate_bias->ne[0] == dst->ne[0]);
+            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
+            fusion_local.gate_bias = fusion->gate_bias->data;
+        }
+        fusion_local.glu_op = fusion->glu_op;
+    }
+
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
         const size_t size_data  = ggml_nbytes(src0);
@@ -549,10 +694,10 @@ void ggml_cuda_mul_mat_vec_q(
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
     mul_mat_vec_q_switch_type(
-        src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
+        src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, dst_d, ne00,
         ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
         ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-        ne03,              ne3,           s03, s13,              s3,                 stream);
+        ne03,              ne3,           s03, s13,              s3,               stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -578,8 +723,9 @@ void ggml_cuda_op_mul_mat_vec_q(
     const int stride_row_x = ne00 / ggml_blck_size(src0->type);
     const int stride_col_y = src1_padded_row_size / QK8_1;
 
+    ggml_cuda_mm_fusion_args_device fusion_local{};
     mul_mat_vec_q_switch_type(
-        src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
+        src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
         1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
 
     GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cuh
index 39dc7d33..4bb10cfa 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/mmvq.cuh
@@ -3,7 +3,7 @@
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
 
 void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 
 void ggml_cuda_op_mul_mat_vec_q(
     ggml_backend_cuda_context & ctx,
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/rope.cu b/ml/backend/ggml/ggml/src/ggml-cuda/rope.cu
index 287fe9d2..71ca6021 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/rope.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/rope.cu
@@ -1,3 +1,6 @@
+#include "convert.cuh"
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
 #include "rope.cuh"
 
 struct rope_corr_dims {
@@ -37,11 +40,23 @@ static __device__ void rope_yarn(
     }
 }
 
-template<bool forward, bool has_ff, typename T>
-static __global__ void rope_norm(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
-        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors) {
+template <bool forward, bool has_ff, typename T, typename D>
+static __global__ void rope_norm(const T *            x,
+                                 D *                  dst,
+                                 const int            ne0,
+                                 const int            ne1,
+                                 const int            s1,
+                                 const int            s2,
+                                 const int            n_dims,
+                                 const int32_t *      pos,
+                                 const float          freq_scale,
+                                 const float          ext_factor,
+                                 const float          attn_factor,
+                                 const rope_corr_dims corr_dims,
+                                 const float          theta_scale,
+                                 const float *        freq_factors,
+                                 const int64_t *      row_indices,
+                                 const int            set_rows_stride) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -53,13 +68,27 @@ static __global__ void rope_norm(
     const int row_x     = row_dst % ne1;
     const int channel_x = row_dst / ne1;
 
-    const int idst = row_dst*ne0 + i0;
+    int       idst = row_dst * ne0 + i0;
     const int ix   = channel_x*s2 + row_x*s1 + i0;
 
-    if (i0 >= n_dims) {
-        dst[idst + 0] = x[ix + 0];
-        dst[idst + 1] = x[ix + 1];
+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
+    // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
+    if (set_rows_stride != 0) {
+        idst = row_x * ne0 + i0;
+        idst += row_indices[channel_x] * set_rows_stride;
+    }
 
+    const auto & store_coaelsced = [&](float x0, float x1) {
+        if constexpr (std::is_same_v<float, D>) {
+            float2 v = make_float2(x0, x1);
+            ggml_cuda_memcpy_1<8>(dst + idst, &v);
+        } else if constexpr (std::is_same_v<half, D>) {
+            half2 v = make_half2(x0, x1);
+            ggml_cuda_memcpy_1<4>(dst + idst, &v);
+        }
+    };
+    if (i0 >= n_dims) {
+        store_coaelsced(x[ix + 0], x[ix + 1]);
         return;
     }
 
@@ -75,15 +104,26 @@ static __global__ void rope_norm(
     const float x0 = x[ix + 0];
     const float x1 = x[ix + 1];
 
-    dst[idst + 0] = x0*cos_theta - x1*sin_theta;
-    dst[idst + 1] = x0*sin_theta + x1*cos_theta;
+    store_coaelsced(x0 * cos_theta - x1 * sin_theta, x0 * sin_theta + x1 * cos_theta);
 }
 
-template<bool forward, bool has_ff, typename T>
-static __global__ void rope_neox(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims,
-        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors) {
+template <bool forward, bool has_ff, typename T, typename D>
+static __global__ void rope_neox(const T *            x,
+                                 D *                  dst,
+                                 const int            ne0,
+                                 const int            ne1,
+                                 const int            s1,
+                                 const int            s2,
+                                 const int            n_dims,
+                                 const int32_t *      pos,
+                                 const float          freq_scale,
+                                 const float          ext_factor,
+                                 const float          attn_factor,
+                                 const rope_corr_dims corr_dims,
+                                 const float          theta_scale,
+                                 const float *        freq_factors,
+                                 const int64_t *      row_indices,
+                                 const int            set_rows_stride) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -95,12 +135,19 @@ static __global__ void rope_neox(
     const int row_x     = row_dst % ne1;
     const int channel_x = row_dst / ne1;
 
-    const int idst = row_dst*ne0 + i0/2;
+    int       idst = row_dst * ne0 + i0 / 2;
     const int ix   = channel_x*s2 + row_x*s1 + i0/2;
 
+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
+    // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
+    if (set_rows_stride != 0) {
+        idst = row_x * ne0 + i0 / 2;
+        idst += row_indices[channel_x] * set_rows_stride;
+    }
+
     if (i0 >= n_dims) {
-        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
-        dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];
+        dst[idst + i0 / 2 + 0] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 0]);
+        dst[idst + i0 / 2 + 1] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 1]);
 
         return;
     }
@@ -117,15 +164,15 @@ static __global__ void rope_neox(
     const float x0 = x[ix + 0];
     const float x1 = x[ix + n_dims/2];
 
-    dst[idst + 0]        = x0*cos_theta - x1*sin_theta;
-    dst[idst + n_dims/2] = x0*sin_theta + x1*cos_theta;
+    dst[idst + 0]          = ggml_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }
 
 template<bool forward, bool has_ff, typename T>
 static __global__ void rope_multi(
         const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2,
         const int n_dims, const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections) {
+        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections, const bool is_imrope) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -151,12 +198,30 @@ static __global__ void rope_multi(
     const int sec_w = sections.v[1] + sections.v[0];
     const int sector = (i0 / 2) % sect_dims;
 
-    float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
-    if (sector % 3 == 1 && sector < 1 + 3 * sections.v[1]) {
-        theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
-    }
-    else if (sector % 3 == 2 && sector < 2 + 3 * sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+    float theta_base = 0.0;
+    if (is_imrope) {
+        if (sector % 3 == 1 && sector < 1 + 3 * sections.v[1]) { // h
+            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 2 + 3 * sections.v[2]) { // w
+            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
+            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+        // } else {
+        //     theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < sections.v[0]) {
+            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sections.v[0] && sector < sec_w) {
+            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+        }
     }
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
@@ -220,11 +285,25 @@ static __global__ void rope_vision(
     dst[idst + n_dims] = x0*sin_theta + x1*cos_theta;
 }
 
-template<bool forward, typename T>
-static void rope_norm_cuda(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+template <bool forward, typename T, typename D>
+static void rope_norm_cuda(const T *            x,
+                           D *                  dst,
+                           const int            ne0,
+                           const int            ne1,
+                           const int            s1,
+                           const int            s2,
+                           const int            n_dims,
+                           const int            nr,
+                           const int32_t *      pos,
+                           const float          freq_scale,
+                           const float          freq_base,
+                           const float          ext_factor,
+                           const float          attn_factor,
+                           const rope_corr_dims corr_dims,
+                           const float *        freq_factors,
+                           const int64_t *      row_indices,
+                           const int            set_rows_stride,
+                           cudaStream_t         stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -234,20 +313,34 @@ static void rope_norm_cuda(
 
     if (freq_factors == nullptr) {
         rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
     } else {
         rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
     }
 }
 
-template<bool forward, typename T>
-static void rope_neox_cuda(
-        const T * x, T * dst, const int ne0, const int ne1, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+template <bool forward, typename T, typename D>
+static void rope_neox_cuda(const T *            x,
+                           D *                  dst,
+                           const int            ne0,
+                           const int            ne1,
+                           const int            s1,
+                           const int            s2,
+                           const int            n_dims,
+                           const int            nr,
+                           const int32_t *      pos,
+                           const float          freq_scale,
+                           const float          freq_base,
+                           const float          ext_factor,
+                           const float          attn_factor,
+                           const rope_corr_dims corr_dims,
+                           const float *        freq_factors,
+                           const int64_t *      row_indices,
+                           const int            set_rows_stride,
+                           cudaStream_t         stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -256,13 +349,13 @@ static void rope_neox_cuda(
     const float theta_scale = powf(freq_base, -2.0f/n_dims);
 
     if (freq_factors == nullptr) {
-        rope_neox<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+        rope_neox<forward, false><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
     } else {
-        rope_neox<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors);
+        rope_neox<forward, true><<<block_nums, block_dims, 0, stream>>>(
+            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
+            freq_factors, row_indices, set_rows_stride);
     }
 }
 
@@ -270,7 +363,7 @@ template<bool forward, typename T>
 static void rope_multi_cuda(
         const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
         const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, cudaStream_t stream) {
+        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, const bool is_imrope, cudaStream_t stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -281,11 +374,11 @@ static void rope_multi_cuda(
     if (freq_factors == nullptr) {
         rope_multi<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
             x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections);
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     } else {
         rope_multi<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
             x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections);
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     }
 }
 
@@ -315,7 +408,9 @@ static void rope_vision_cuda(
 }
 
 template <bool forward>
-void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
+                            ggml_tensor *               dst,
+                            const ggml_tensor *         set_rows = nullptr) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * src2 = dst->src[2];
@@ -323,12 +418,25 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     const float * src0_d = (const float *)src0->data;
     const float * src1_d = (const float *)src1->data;
 
-    float * dst_d = (float *)dst->data;
+    void *          dst_d           = dst->data;
+    const int64_t * row_indices     = nullptr;
+    ggml_type       dst_type        = dst->type;
+    int             set_rows_stride = 0;
+
+    if (set_rows != nullptr) {
+        GGML_ASSERT(forward);
+        dst_d           = set_rows->data;
+        row_indices     = (const int64_t *) set_rows->src[1]->data;
+        dst_type        = set_rows->type;
+        set_rows_stride = set_rows->nb[1] / ggml_type_size(set_rows->type);
+    }
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
     GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
-    GGML_ASSERT(src0->type == dst->type);
+    // When not fused, src0 and dst types must match
+    // When fused (ROPE+VIEW+SET_ROWS), src0 may be F32 and dst may be F16
+    GGML_ASSERT(src0->type == dst->type || (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16));
 
     const int64_t ne00 = src0->ne[0]; // head dims
     const int64_t ne01 = src0->ne[1]; // num heads
@@ -363,6 +471,7 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -385,14 +494,18 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 
     // compute
     if (is_neox) {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_neox_cuda<forward>(
-                (const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_neox_cuda<forward>(
-                (const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
+            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                  freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
+            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                 freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
+            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
+                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                freq_factors, row_indices, set_rows_stride, stream);
         } else {
             GGML_ABORT("fatal error");
         }
@@ -400,11 +513,11 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
         if (src0->type == GGML_TYPE_F32) {
             rope_multi_cuda<forward>(
                 (const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
         } else if (src0->type == GGML_TYPE_F16) {
             rope_multi_cuda<forward>(
                 (const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
         } else {
             GGML_ABORT("fatal error");
         }
@@ -421,14 +534,18 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
             GGML_ABORT("fatal error");
         }
     } else {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_norm_cuda<forward>(
-                (const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_norm_cuda<forward>(
-                (const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
+            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                  freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
+            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
+                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                 freq_factors, row_indices, set_rows_stride, stream);
+        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
+            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
+                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                                                freq_factors, row_indices, set_rows_stride, stream);
         } else {
             GGML_ABORT("fatal error");
         }
@@ -442,3 +559,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_rope_impl<false>(ctx, dst);
 }
+
+void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rope, ggml_tensor * set_rows) {
+    ggml_cuda_op_rope_impl<true>(ctx, rope, set_rows);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/rope.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/rope.cuh
index 9139f3b2..72af086c 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/rope.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/rope.cuh
@@ -5,3 +5,5 @@
 void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * set_rows);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/set-rows.cu b/ml/backend/ggml/ggml/src/ggml-cuda/set-rows.cu
index 1525a159..631de7e8 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/set-rows.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/set-rows.cu
@@ -4,30 +4,53 @@
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
 // Generic quantized set_rows kernel template
-template<typename idx_t, typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
-static __global__ void k_set_rows_quant(
-        const float * __restrict__ src0, const idx_t * __restrict__ src1, block_type * __restrict__ dst,
-        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
-        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
-        const int64_t s01, const int64_t s02, const int64_t s03,
-        const int64_t s10, const int64_t s11, const int64_t s12,
-        const int64_t s1, const int64_t s2, const int64_t s3) {
-
+template <typename idx_t, typename block_type, int qk, void (*quantize_func)(const float *, block_type *)>
+static __global__ void k_set_rows_quant(const float * __restrict__ src0,
+                                        const idx_t * __restrict__ src1,
+                                        block_type * __restrict__ dst,
+                                        const int64_t ne_total,
+                                        const int64_t ne10,
+                                        const int64_t ne11,
+                                        const int64_t ne12,
+                                        const int64_t ne13,
+                                        const int64_t s01,
+                                        const int64_t s02,
+                                        const int64_t s03,
+                                        const int64_t s10,
+                                        const int64_t s11,
+                                        const int64_t s12,
+                                        const int64_t s1,
+                                        const int64_t s2,
+                                        const int64_t s3,
+                                        const uint3   ne00,
+                                        const uint3   ne01,
+                                        const uint3   ne02,
+                                        const uint3   ne11_fd,
+                                        const uint3   ne12_fd) {
     const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
-    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
 
     if (i >= ne_total) {
         return;
     }
 
     const int64_t i_base = i * qk;
-    const int64_t i03 = i_base / (ne00 * ne01 * ne02);
-    const int64_t i02 = (i_base - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
-    const int64_t i01 = (i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
-    const int64_t i00 = i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+    uint32_t      tmp    = (uint32_t) i_base;
+    uint2         div_mod;
 
-    const int64_t i12 = i03 % ne12;
-    const int64_t i11 = i02 % ne11;
+    div_mod           = fast_div_modulo(tmp, ne00);
+    const int64_t i00 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne01);
+    const int64_t i01 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne02);
+    const int64_t i02 = div_mod.y;
+    const int64_t i03 = div_mod.x;
+
+    const int64_t i12 = fastmodulo((uint32_t) i03, ne12_fd);
+    const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
@@ -41,6 +64,8 @@ static __global__ void k_set_rows_quant(
     quantize_func(src_block, dst_block);
 
     GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12);
     GGML_UNUSED(ne13);
 }
 
@@ -71,40 +96,65 @@ static void set_rows_cuda_quant(
     const int64_t s2  = nb2;
     const int64_t s3  = nb3;
 
-    if (ne_total > 0) {
+    if (ne_total > 0 && ne00 > 0 && ne01 > 0 && ne02 > 0 && ne11 > 0 && ne12 > 0) {
+        const uint3 ne00_fd = init_fastdiv_values((uint32_t) ne00);
+        const uint3 ne01_fd = init_fastdiv_values((uint32_t) ne01);
+        const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
+        const uint3 ne11_fd = init_fastdiv_values((uint32_t) ne11);
+        const uint3 ne12_fd = init_fastdiv_values((uint32_t) ne12);
+
         k_set_rows_quant<idx_t, block_type, qk, quantize_func><<<grid_size, block_size, 0, stream>>>(
-            src0_d, src1_d, dst_d,
-            ne00, ne01, ne02, ne03,
-            ne10, ne11, ne12, ne13,
-            s01, s02, s03,
-            s10, s11, s12,
-            s1, s2, s3);
+            src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01, s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd,
+            ne01_fd, ne02_fd, ne11_fd, ne12_fd);
     }
 }
 
-template<typename src_t, typename idx_t, typename dst_t>
-static __global__ void k_set_rows(
-        const src_t * __restrict__ src0, const idx_t * __restrict__ src1, dst_t * __restrict__ dst,
-        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
-        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
-        const int64_t s01, const int64_t s02, const int64_t s03,
-        const int64_t s10, const int64_t s11, const int64_t s12,
-        const int64_t s1, const int64_t s2, const int64_t s3) {
-
+template <typename src_t, typename idx_t, typename dst_t>
+static __global__ void k_set_rows(const src_t * __restrict__ src0,
+                                  const idx_t * __restrict__ src1,
+                                  dst_t * __restrict__ dst,
+                                  const int64_t ne_total,
+                                  const int64_t ne10,
+                                  const int64_t ne11,
+                                  const int64_t ne12,
+                                  const int64_t ne13,
+                                  const int64_t s01,
+                                  const int64_t s02,
+                                  const int64_t s03,
+                                  const int64_t s10,
+                                  const int64_t s11,
+                                  const int64_t s12,
+                                  const int64_t s1,
+                                  const int64_t s2,
+                                  const int64_t s3,
+                                  const uint3   ne00,
+                                  const uint3   ne01,
+                                  const uint3   ne02,
+                                  const uint3   ne11_fd,
+                                  const uint3   ne12_fd) {
     const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
-    const int64_t ne_total = ne00 * ne01 * ne02 * ne03;
 
     if (i >= ne_total) {
         return;
     }
 
-    const int64_t i03 = i / (ne00 * ne01 * ne02);
-    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
-    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
-    const int64_t i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+    uint32_t tmp = (uint32_t) i;
+    uint2    div_mod;
 
-    const int64_t i12 = i03 % ne12;
-    const int64_t i11 = i02 % ne11;
+    div_mod           = fast_div_modulo(tmp, ne00);
+    const int64_t i00 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne01);
+    const int64_t i01 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne02);
+    const int64_t i02 = div_mod.y;
+    const int64_t i03 = div_mod.x;
+
+    const int64_t i12 = fastmodulo((uint32_t) i03, ne12_fd);
+    const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
@@ -115,6 +165,8 @@ static __global__ void k_set_rows(
     dst_row_ptr[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 
     GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12);
     GGML_UNUSED(ne13);
 }
 
@@ -144,14 +196,16 @@ static void set_rows_cuda(
     const int64_t s2  = nb2/sizeof(dst_t);
     const int64_t s3  = nb3/sizeof(dst_t);
 
-    if (ne_total > 0) {
-        k_set_rows<<<grid_size, block_size, 0, stream>>>(
-            src0_d, src1_d, dst_d,
-            ne00, ne01, ne02, ne03,
-            ne10, ne11, ne12, ne13,
-            s01, s02, s03,
-            s10, s11, s12,
-            s1, s2, s3);
+    if (ne_total > 0 && ne00 > 0 && ne01 > 0 && ne02 > 0 && ne11 > 0 && ne12 > 0) {
+        const uint3 ne00_fd = init_fastdiv_values((uint32_t) ne00);
+        const uint3 ne01_fd = init_fastdiv_values((uint32_t) ne01);
+        const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
+        const uint3 ne11_fd = init_fastdiv_values((uint32_t) ne11);
+        const uint3 ne12_fd = init_fastdiv_values((uint32_t) ne12);
+
+        k_set_rows<<<grid_size, block_size, 0, stream>>>(src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01,
+                                                         s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd, ne01_fd, ne02_fd,
+                                                         ne11_fd, ne12_fd);
     }
 }
 
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/set.cu b/ml/backend/ggml/ggml/src/ggml-cuda/set.cu
new file mode 100644
index 00000000..04bfe07b
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/set.cu
@@ -0,0 +1,39 @@
+#include "set.cuh"
+#include "cpy.cuh"
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32));
+    GGML_ASSERT(src1->type == src0->type);
+    GGML_ASSERT(dst ->type == src0->type);
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const size_t nb1    = ((int32_t *) dst->op_params)[0];
+    const size_t nb2    = ((int32_t *) dst->op_params)[1];
+    const size_t nb3    = ((int32_t *) dst->op_params)[2];
+    const size_t offset = ((int32_t *) dst->op_params)[3];
+    const bool   inplace= (bool)     ((int32_t *) dst->op_params)[4];
+
+    if (!inplace) {
+        ggml_cuda_cpy(ctx, src0, dst);
+    }
+
+    ggml_tensor dst_view = *dst;
+    dst_view.data  = (void *)((char *)dst->data + offset);
+    dst_view.ne[0] = src1->ne[0];
+    dst_view.ne[1] = src1->ne[1];
+    dst_view.ne[2] = src1->ne[2];
+    dst_view.ne[3] = src1->ne[3];
+
+    dst_view.nb[0] = ggml_element_size(dst);
+    dst_view.nb[1] = nb1;
+    dst_view.nb[2] = nb2;
+    dst_view.nb[3] = nb3;
+
+    ggml_cuda_cpy(ctx, src1, &dst_view);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/set.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/set.cuh
new file mode 100644
index 00000000..dd09529f
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/set.cuh
@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+#define CUDA_SET_BLOCK_SIZE 256
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cu b/ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cu
new file mode 100644
index 00000000..2e2b3972
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cu
@@ -0,0 +1,203 @@
+#include "common.cuh"
+#include "ggml.h"
+#include "solve_tri.cuh"
+
+#define MAX_N_FAST 64
+#define MAX_K_FAST 32
+
+// ======================
+// Fast Kernel (n <= 64, k <= 32) - Warp-based parallel reduction
+// ======================
+// When ncols_template == 0 the bounds for the loops in this function are not
+// known and can't be unrolled. As we want to keep pragma unroll for all other
+// cases we supress the clang transformation warning here.
+#ifdef __clang__
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wpass-failed"
+#endif  // __clang__
+template <int n_template, int k_template>
+static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
+                                          const float * __restrict__ B,
+                                          float * __restrict__ X,
+                                          const uint3  ne02,
+                                          const size_t nb02,
+                                          const size_t nb03,
+                                          const size_t nb12,
+                                          const size_t nb13,
+                                          const size_t nb2,
+                                          const size_t nb3,
+                                          const int    n_arg,
+                                          const int    k_arg) {
+    const int n = n_template == 0 ? n_arg : n_template;
+    const int k = k_template == 0 ? k_arg : k_template;
+
+    const int batch_idx = blockIdx.x;
+    const int lane      = threadIdx.x;
+    const int col_idx   = threadIdx.y;
+
+    if (col_idx >= k) {
+        return;
+    }
+
+    const uint2   i02_i03 = fast_div_modulo(batch_idx, ne02);
+    const int64_t i02     = i02_i03.y;
+    const int64_t i03     = i02_i03.x;
+
+    const float * const A_batch = (const float *) (A + i02 * nb02 + i03 * nb03);
+    const float * const B_batch = (const float *) (B + i02 * nb12 + i03 * nb13);
+    float *             X_batch = (float *) (X + i02 * nb2 + i03 * nb3);
+
+    __shared__ float sA[MAX_N_FAST * MAX_N_FAST];
+    __shared__ float sXt[MAX_N_FAST * (MAX_K_FAST + 1)];
+
+    const int offset = threadIdx.x + threadIdx.y * blockDim.x;
+
+#pragma unroll
+    for (int i = 0; i < n * n; i += k * WARP_SIZE) {
+        int i0 = i + offset;
+        if (i0 < n * n) {
+            sA[i0] = A_batch[i0];
+        }
+    }
+
+    const int rows_per_warp = (n + WARP_SIZE - 1) / WARP_SIZE;
+
+#pragma unroll
+    for (int i = 0; i < rows_per_warp; i++) {
+        const int i0 = lane + i * WARP_SIZE;
+        if (i0 < n) {
+            sXt[col_idx * n + i0] = B_batch[i0 * k + col_idx];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int row = 0; row < n; ++row) {
+        float sum = 0.0f;
+
+        {
+            int j = lane;
+            if (j < row) {
+                sum += sA[row * n + j] * sXt[col_idx * n + j];
+            }
+        }
+        if (row >= WARP_SIZE) {
+            int j = WARP_SIZE + lane;
+            if (j < row) {
+                sum += sA[row * n + j] * sXt[col_idx * n + j];
+            }
+        }
+
+        sum = warp_reduce_sum(sum);
+
+        if (lane == 0) {
+            const float b_val      = sXt[col_idx * n + row];
+            const float a_diag     = sA[row * n + row];
+            // no safeguards for division by zero because that indicates corrupt
+            // data anyway
+            sXt[col_idx * n + row] = (b_val - sum) / a_diag;
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int i = 0; i < rows_per_warp; i++) {
+        const int i0 = lane + i * WARP_SIZE;
+        if (i0 < n) {
+            X_batch[i0 * k + col_idx] = sXt[col_idx * n + i0];
+        }
+    }
+}
+#ifdef __clang__
+#    pragma clang diagnostic pop
+#endif  // __clang__
+
+static void solve_tri_f32_cuda(const float * A,
+                               const float * B,
+                               float *       X,
+                               int           n,
+                               int           k,
+                               int64_t       ne02,
+                               int64_t       ne03,
+                               size_t        nb02,
+                               size_t        nb03,
+                               size_t        nb12,
+                               size_t        nb13,
+                               size_t        nb2,
+                               size_t        nb3,
+                               cudaStream_t  stream) {
+    const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
+    dim3        threads(WARP_SIZE, k);
+    dim3        grid(ne02 * ne03);
+    if (n == 64) {
+        switch (k) {
+            case 32:
+                solve_tri_f32_fast<64, 32>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 16:
+                solve_tri_f32_fast<64, 16>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 14:
+                solve_tri_f32_fast<64, 14>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 12:
+                solve_tri_f32_fast<64, 12>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 10:
+                solve_tri_f32_fast<64, 10>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 8:
+                solve_tri_f32_fast<64, 8>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 6:
+                solve_tri_f32_fast<64, 6>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 4:
+                solve_tri_f32_fast<64, 4>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 2:
+                solve_tri_f32_fast<64, 2>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            case 1:
+                solve_tri_f32_fast<64, 1>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, 0, 0);
+                break;
+            default:
+                solve_tri_f32_fast<0, 0>
+                    <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
+        }
+    } else {  // run general case
+        solve_tri_f32_fast<0, 0>
+            <<<grid, threads, 0, stream>>>(A, B, X, ne02_fd, nb02, nb03, nb12, nb13, nb2, nb3, n, k);
+    }
+}
+
+void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];  // A (triangular n x x matrix)
+    const ggml_tensor * src1 = dst->src[1];  // B (right hand side of n x k equation columns)
+
+    ggml_is_contiguous(src0);
+    ggml_is_contiguous(src1);
+
+    const int64_t n = src0->ne[0];
+    const int64_t k = src1->ne[0];
+
+    GGML_ASSERT(n <= 64);
+    GGML_ASSERT(k <= 32);
+
+    solve_tri_f32_cuda((const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k, src0->ne[2],
+                       src0->ne[3], src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
+                       src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
+                       dst->nb[3] / sizeof(float), ctx.stream());
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cuh
new file mode 100644
index 00000000..63999239
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu b/ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu
new file mode 100644
index 00000000..8f9d5315
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(72, 72);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cu b/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cu
index e28c810a..572379fc 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cu
@@ -2,6 +2,7 @@
 #include "ggml.h"
 #include "topk-moe.cuh"
 
+#include <cmath>
 #include <initializer_list>
 
 // Warp-local softmax used for both the pre-top-k logits and the post-top-k delayed path.
@@ -63,7 +64,8 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
                                                                   float *       weights,
                                                                   int32_t *     ids,
                                                                   const int     n_rows,
-                                                                  const int     n_expert_used) {
+                                                                  const int     n_expert_used,
+                                                                  const float   clamp_val) {
     const int row = blockIdx.x * blockDim.y + threadIdx.y;
     if (row >= n_rows) {
         return;
@@ -139,6 +141,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
 
     if constexpr (with_norm) {
         wt_sum              = warp_reduce_sum(wt_sum);
+        wt_sum              = max(wt_sum, clamp_val);
         const float inv_sum = 1.0f / wt_sum;
 
         for (int i = 0; i < experts_per_thread; i++) {
@@ -157,6 +160,10 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
             weights[idx] = output_weights[i];
         }
     }
+
+    if (!with_norm) {
+        GGML_UNUSED(clamp_val);
+    }
 }
 
 template <bool with_norm, bool delayed_softmax = false>
@@ -166,9 +173,9 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
                                  int32_t *                   ids,
                                  const int                   n_rows,
                                  const int                   n_expert,
-                                 const int                   n_expert_used) {
+                                 const int                   n_expert_used,
+                                 const float                 clamp_val) {
     static_assert(!(with_norm && delayed_softmax), "delayed softmax is not supported with weight normalization");
-
     const int    rows_per_block = 4;
     dim3         grid_dims((n_rows + rows_per_block - 1) / rows_per_block, 1, 1);
     dim3         block_dims(WARP_SIZE, rows_per_block, 1);
@@ -177,43 +184,43 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
     switch (n_expert) {
         case 1:
             topk_moe_cuda<1, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 2:
             topk_moe_cuda<2, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 4:
             topk_moe_cuda<4, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 8:
             topk_moe_cuda<8, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 16:
             topk_moe_cuda<16, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 32:
             topk_moe_cuda<32, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 64:
             topk_moe_cuda<64, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 128:
             topk_moe_cuda<128, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 256:
             topk_moe_cuda<256, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         case 512:
             topk_moe_cuda<512, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used);
+                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
             break;
         default:
             GGML_ASSERT(false && "fatal error");
@@ -226,7 +233,8 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
                            ggml_tensor *               weights,
                            ggml_tensor *               ids,
                            const bool                  with_norm,
-                           const bool                  delayed_softmax) {
+                           const bool                  delayed_softmax,
+                           ggml_tensor *               clamp) {
     GGML_ASSERT(logits->type == GGML_TYPE_F32);
     GGML_ASSERT(weights->type == GGML_TYPE_F32);
     GGML_ASSERT(ids->type == GGML_TYPE_I32);
@@ -242,18 +250,25 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
 
     const int n_expert_used = weights->ne[1];
 
+    float clamp_val = -INFINITY;
     if (with_norm) {
-        launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used);
+        if (clamp) {
+            clamp_val = ggml_get_op_params_f32(clamp, 0);
+        }
+        launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used, clamp_val);
     } else {
+        GGML_ASSERT(clamp == nullptr);
         if (delayed_softmax) {
-            launch_topk_moe_cuda<false, true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used);
+            launch_topk_moe_cuda<false, true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used,
+                                              clamp_val);
         } else {
-            launch_topk_moe_cuda<false, false>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used);
+            launch_topk_moe_cuda<false, false>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used,
+                                               clamp_val);
         }
     }
 }
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights) {
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights, const ggml_tensor * clamp) {
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
@@ -279,13 +294,26 @@ bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tenso
         return false;
     }
 
+    if (clamp) {
+        if (clamp->op != GGML_OP_CLAMP) {
+            return false;
+        }
+        float max_val = ggml_get_op_params_f32(clamp, 1);
+
+        if (max_val != INFINITY) {
+            return false;
+        }
+    }
+
+
     return true;
 }
 
 std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool norm, bool delayed_softmax) {
     static std::initializer_list<enum ggml_op> norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
                                                             GGML_OP_VIEW,     GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
-                                                            GGML_OP_SUM_ROWS, GGML_OP_DIV,      GGML_OP_RESHAPE };
+                                                            GGML_OP_SUM_ROWS, GGML_OP_CLAMP,    GGML_OP_DIV,
+                                                            GGML_OP_RESHAPE };
 
     static std::initializer_list<enum ggml_op> no_norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE, GGML_OP_ARGSORT,
                                                                GGML_OP_VIEW, GGML_OP_GET_ROWS };
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cuh
index cc2fbfe9..2eff408b 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/topk-moe.cuh
@@ -8,8 +8,9 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
                            ggml_tensor *               weights,
                            ggml_tensor *               ids,
                            const bool                  with_norm,
-                           const bool                  delayed_softmax = false);
+                           const bool                  delayed_softmax = false,
+                           ggml_tensor *               weight_clamp    = nullptr);
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights);
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights, const ggml_tensor * clamp = nullptr);
 
 std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool with_norm, bool delayed_softmax = false);
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/unary.cu b/ml/backend/ggml/ggml/src/ggml-cuda/unary.cu
index 3c564566..d4866067 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/unary.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/unary.cu
@@ -18,10 +18,7 @@ static __device__ __forceinline__ float op_step(float x) {
 }
 
 static __device__ __forceinline__ float op_gelu(float x) {
-    const float GELU_COEF_A    = 0.044715f;
-    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-
-    return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+    return ggml_cuda_op_gelu_single(x);
 }
 
 static __device__ __forceinline__ float op_gelu_erf(float x) {
@@ -37,7 +34,7 @@ static __device__ __forceinline__ float op_gelu_quick(float x) {
 }
 
 static __device__ __forceinline__ float op_silu(float x) {
-    return x / (1.0f + expf(-x));
+    return ggml_cuda_op_silu_single(x);
 }
 
 static __device__ __forceinline__ float op_tanh(float x) {
@@ -84,10 +81,34 @@ static __device__ __forceinline__ float op_log(float x) {
     return logf(x);
 }
 
+static __device__ __forceinline__ float op_expm1(float x) {
+    return expm1f(x);
+}
+
+static __device__ __forceinline__ float op_softplus(float x) {
+    return (x > 20.0f) ? x : logf(1.0f + expf(x));
+}
+
 static __device__ __forceinline__ float op_elu(float x) {
     return (x > 0.f) ? x : expm1f(x);
 }
 
+static __device__ __forceinline__ float op_floor(float x) {
+    return floorf(x);
+}
+
+static __device__ __forceinline__ float op_ceil(float x) {
+    return ceilf(x);
+}
+
+static __device__ __forceinline__ float op_round(float x) {
+    return round(x);
+}
+
+static __device__ __forceinline__ float op_trunc(float x) {
+    return trunc(x);
+}
+
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -204,6 +225,30 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_elu>(ctx, dst);
 }
+
+void ggml_cuda_op_floor(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_floor>(ctx, dst);
+}
+
+void ggml_cuda_op_ceil(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_ceil>(ctx, dst);
+}
+
+void ggml_cuda_op_round(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_round>(ctx, dst);
+}
+
+void ggml_cuda_op_trunc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_trunc>(ctx, dst);
+}
+
+void ggml_cuda_op_expm1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_expm1>(ctx, dst);
+}
+
+void ggml_cuda_op_softplus(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_softplus>(ctx, dst);
+}
 /* gated ops */
 
 template <float (*op)(float), typename T>
@@ -317,13 +362,8 @@ static __global__ void swiglu_oai_kernel(const T * x, const T * g, T * dst, cons
 
     float xi = x[j0];
     float gi = g[j1];
-    xi = fminf(xi, limit);
-    gi = fmaxf(fminf(gi, limit), -limit);
 
-    float out_glu = xi / (1.0f + expf(-xi * alpha));
-    out_glu = out_glu * (1.0f + gi);
-
-    dst[i] = out_glu;
+    dst[i] = ggml_cuda_op_swiglu_oai_single(xi, gi, alpha, limit);
 }
 
 template <typename T>
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/unary.cuh b/ml/backend/ggml/ggml/src/ggml-cuda/unary.cuh
index 8e7644fc..609046e5 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/unary.cuh
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/unary.cuh
@@ -1,3 +1,4 @@
+#pragma once
 #include "common.cuh"
 
 #define CUDA_NEG_BLOCK_SIZE 256
@@ -60,8 +61,20 @@ void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_expm1(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_softplus(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_floor(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_ceil(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_round(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_trunc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
@@ -75,3 +88,23 @@ void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+__device__ __forceinline__ float ggml_cuda_op_silu_single(float x) {
+    return x / (1.0f + expf(-x));
+}
+
+__device__ __forceinline__ float ggml_cuda_op_gelu_single(float x) {
+    const float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+
+    return 0.5f * x * (1.0f + tanhf(SQRT_2_OVER_PI * x * (1.0f + GELU_COEF_A * x * x)));
+}
+
+__device__ __forceinline__ float ggml_cuda_op_swiglu_oai_single(float x, float g, float alpha = 1.702f, float limit = 7.0f) {
+    x = fminf(x, limit);
+    g = fmaxf(fminf(g, limit), -limit);
+
+    float out_glu = x / (1.0f + expf(-x * alpha));
+    out_glu = out_glu * (1.0f + g);
+    return out_glu;
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/upscale.cu b/ml/backend/ggml/ggml/src/ggml-cuda/upscale.cu
index ef48aa5f..687c6693 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/upscale.cu
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/upscale.cu
@@ -81,6 +81,70 @@ static __global__ void upscale_f32_bilinear(const float * x, float * dst,
     dst[index] = result;
 }
 
+namespace bicubic_interpolation {
+// https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+__device__ const float a = -0.75f; // use alpha = -0.75 (same as PyTorch)
+
+static __device__ float weight1(float x) { return ((a + 2) * x - (a + 3)) * x * x + 1; };
+static __device__ float weight2(float x) { return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a; };
+
+static __device__ float bicubic(float p0, float p1, float p2, float p3, float x) {
+    const float w0 = weight2(x + 1);
+    const float w1 = weight1(x + 0);
+    const float w2 = weight1(1 - x);
+    const float w3 = weight2(2 - x);
+    return p0 * w0 + p1 * w1 + p2 * w2 + p3 * w3;
+};
+} // namespace bicubic_interpolation
+
+static __global__ void upscale_f32_bicubic(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset) {
+    using bicubic_interpolation::bicubic;
+
+    const int64_t index              = threadIdx.x + blockIdx.x * blockDim.x;
+    const int64_t dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    const int i10_dst = index % ne10_dst;
+    const int i11_dst = (index / ne10_dst) % ne11_dst;
+    const int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    const int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    const int i02_src = (int)(i12_dst / sf2);
+    const int i03_src = (int)(i13_dst / sf3);
+
+    const float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    const int y0_src    = (int)floorf(y_src_f);
+    const float dy      = y_src_f - (float)y0_src;
+
+    const float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    const int x0_src    = (int)floorf(x_src_f);
+    const float dx      = x_src_f - (float)x0_src;
+
+    const char * x_base = (const char *)x + (int64_t)i02_src * nb02 + (int64_t)i03_src * nb03;
+
+    auto load = [=](int x_off, int y_off) -> float {
+        int i00_src = max(0, min(x0_src + x_off, ne00_src - 1));
+        int i01_src = max(0, min(y0_src + y_off, ne01_src - 1));
+        return *(const float *)(x_base + (int64_t)i00_src * nb00 + (int64_t)i01_src * nb01);
+    };
+
+    const float result = bicubic(
+        bicubic(load(-1,-1), load(0,-1), load(1,-1), load(2,-1), dx),
+        bicubic(load(-1, 0), load(0, 0), load(1, 0), load(2, 0), dx),
+        bicubic(load(-1, 1), load(0, 1), load(1, 1), load(2, 1), dx),
+        bicubic(load(-1, 2), load(0, 2), load(1, 2), load(2, 2), dx), dy);
+
+    dst[index] = result;
+}
+
 static void upscale_f32_cuda(const float * x, float * dst,
         const int nb00, const int nb01, const int nb02, const int nb03,
         const int ne10, const int ne11, const int ne12, const int ne13,
@@ -104,6 +168,18 @@ static void upscale_f32_bilinear_cuda(const float * x, float * dst,
     upscale_f32_bilinear<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
 }
 
+static void upscale_f32_bicubic_cuda(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset, cudaStream_t stream) {
+    const int64_t dst_size   = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    const int64_t num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+
+    upscale_f32_bicubic<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+}
+
 void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;
@@ -121,17 +197,22 @@ void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float sf2 = (float)dst->ne[2]/src0->ne[2];
     const float sf3 = (float)dst->ne[3]/src0->ne[3];
 
+    float pixel_offset = 0.5f;
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        sf0          = dst->ne[0] > 1 && src0->ne[0] > 1 ? (float)(dst->ne[0] - 1) / (src0->ne[0] - 1) : sf0;
+        sf1          = dst->ne[1] > 1 && src0->ne[1] > 1 ? (float)(dst->ne[1] - 1) / (src0->ne[1] - 1) : sf1;
+        pixel_offset = 0.0f;
+    }
+
     if (mode == GGML_SCALE_MODE_NEAREST) {
         upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
-        float pixel_offset = 0.5f;
-        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-            sf0          = (float)(dst->ne[0] - 1) / (src0->ne[0] - 1);
-            sf1          = (float)(dst->ne[1] - 1) / (src0->ne[1] - 1);
-            pixel_offset = 0.0f;
-        }
         upscale_f32_bilinear_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
                                  src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
                                  sf0, sf1, sf2, sf3, pixel_offset, stream);
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        upscale_f32_bicubic_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                                 src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+                                 sf0, sf1, sf2, sf3, pixel_offset, stream);
     }
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h b/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h
index 2f9ef2dc..5ad5623a 100644
--- a/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h
@@ -109,7 +109,7 @@
 #define cudaStreamNonBlocking hipStreamNonBlocking
 #define cudaStreamPerThread hipStreamPerThread
 #define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaStreamWaitEvent hipStreamWaitEvent
 #define cudaGraphExec_t hipGraphExec_t
 #define cudaGraphNode_t hipGraphNode_t
 #define cudaKernelNodeParams hipKernelNodeParams
diff --git a/ml/backend/ggml/ggml/src/ggml-hip/CMakeLists.txt b/ml/backend/ggml/ggml/src/ggml-hip/CMakeLists.txt
index 6b499320..23b68899 100644
--- a/ml/backend/ggml/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ml/backend/ggml/ggml/src/ggml-hip/CMakeLists.txt
@@ -29,10 +29,11 @@ if (CXX_IS_HIPCC)
     endif()
 else()
     # Forward (AMD)GPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+    if(AMDGPU_TARGETS AND NOT GPU_TARGETS)
+        set(GPU_TARGETS ${AMDGPU_TARGETS})
+    endif()
     if(GPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
         set(CMAKE_HIP_ARCHITECTURES ${GPU_TARGETS})
-    elseif(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
-        set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
     endif()
     cmake_minimum_required(VERSION 3.21)
     enable_language(HIP)
diff --git a/ml/backend/ggml/ggml/src/ggml-impl.h b/ml/backend/ggml/ggml/src/ggml-impl.h
index e5c446d1..0da3e065 100644
--- a/ml/backend/ggml/ggml/src/ggml-impl.h
+++ b/ml/backend/ggml/ggml/src/ggml-impl.h
@@ -102,7 +102,7 @@ static bool ggml_op_is_empty(enum ggml_op op) {
     }
 }
 
-static inline float ggml_softplus(float input) {
+static inline float ggml_compute_softplus_f32(float input) {
     return (input > 20.0f) ? input : logf(1 + expf(input));
 }
 //
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-context.m b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-context.m
index 052efb7a..e6664628 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-context.m
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-context.m
@@ -35,7 +35,6 @@ struct ggml_metal {
     // additional, inference-time compiled pipelines
     ggml_metal_pipelines_t pipelines_ext;
 
-    bool use_bfloat;
     bool use_fusion;
     bool use_concurrency;
     bool use_graph_optimize;
@@ -121,11 +120,10 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
         }
     }
 
-    const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
+    //const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
 
     res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
-    res->use_bfloat      = props_dev->has_bfloat;
     res->use_fusion      = getenv("GGML_METAL_FUSION_DISABLE") == nil;
     res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil;
 
@@ -147,7 +145,6 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
 
     memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt));
 
-    GGML_LOG_INFO("%s: use bfloat         = %s\n", __func__, res->use_bfloat         ? "true" : "false");
     GGML_LOG_INFO("%s: use fusion         = %s\n", __func__, res->use_fusion         ? "true" : "false");
     GGML_LOG_INFO("%s: use concurrency    = %s\n", __func__, res->use_concurrency    ? "true" : "false");
     GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");
@@ -292,7 +289,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
 
         // queue the copy operation into the queue of the Metal context
         // this will be queued at the end, after any currently ongoing GPU operations
-        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
         id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
 
         [encoder copyFromBuffer:buf_src
@@ -303,6 +300,7 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
 
         [encoder endEncoding];
         [cmd_buf commit];
+        [buf_src release];
 
         // do not wait here for completion
         //[cmd_buf waitUntilCompleted];
@@ -333,7 +331,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
 
         // queue the copy operation into the queue of the Metal context
         // this will be queued at the end, after any currently ongoing GPU operations
-        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
         id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
 
         [encoder copyFromBuffer:bid_src.metal
@@ -344,6 +342,7 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
 
         [encoder endEncoding];
         [cmd_buf commit];
+        [buf_dst release];
 
         // do not wait here for completion
         //[cmd_buf waitUntilCompleted];
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.cpp b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.cpp
index c78082ac..329500a0 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.cpp
@@ -318,6 +318,44 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum_rows(ggml_metal_librar
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_blk(ggml_metal_library_t lib, const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_CUMSUM);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_cumsum_blk_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_add(ggml_metal_library_t lib, const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_CUMSUM);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_cumsum_add_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_soft_max(ggml_metal_library_t lib, const ggml_tensor * op) {
     GGML_ASSERT(!op->src[1] || op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32);
 
@@ -943,6 +981,92 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort(ggml_metal_library
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_ARGSORT);
+
+    char base[256];
+    char name[256];
+
+    ggml_sort_order order = (ggml_sort_order) op->op_params[0];
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_merge_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
+// note: reuse the argsort kernel for top_k
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    // note: the top_k kernel is always descending order
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_merge(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_TOP_K);
+
+    char base[256];
+    char name[256];
+
+    ggml_sort_order order = GGML_SORT_ORDER_DESC;
+
+    const char * order_str = "undefined";
+    switch (order) {
+        case GGML_SORT_ORDER_ASC:  order_str = "asc";  break;
+        case GGML_SORT_ORDER_DESC: order_str = "desc"; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_argsort_merge_%s_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->type), order_str);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
@@ -1332,11 +1456,12 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope(ggml_metal_library_t
 
     const bool is_neox   = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope  = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_neox) {
         snprintf(base, 256, "kernel_rope_neox_%s", ggml_type_name(op->src[0]->type));
-    } else if (is_mrope && !is_vision) {
+    } else if ((is_mrope || is_imrope) && !is_vision) {
         GGML_ASSERT(op->src[1]->ne[0]*4 >= op->src[0]->ne[2]); // need at least 4 pos per token
         snprintf(base, 256, "kernel_rope_multi_%s", ggml_type_name(op->src[0]->type));
     } else if (is_vision) {
@@ -1346,14 +1471,20 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope(ggml_metal_library_t
         snprintf(base, 256, "kernel_rope_norm_%s", ggml_type_name(op->src[0]->type));
     }
 
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_imrope=%d", base, is_imrope ? 1 : 0);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
     if (res) {
         return res;
     }
 
-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_bool(cv, is_imrope, FC_ROPE + 0);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
 
     return res;
 }
@@ -1431,6 +1562,30 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_2d(ggml_met
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_2d(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_CONV_2D);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type         == GGML_TYPE_F32);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_conv_2d_%s_%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->src[1]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_upscale(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_UPSCALE);
 
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.h b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.h
index 4d582974..3976e622 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.h
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.h
@@ -95,7 +95,9 @@ void ggml_metal_encoder_end_encoding(ggml_metal_encoder_t encoder);
 
 typedef struct ggml_metal_library * ggml_metal_library_t;
 
-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev);
+ggml_metal_library_t ggml_metal_library_init            (ggml_metal_device_t dev);
+ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose);
+
 void ggml_metal_library_free(ggml_metal_library_t lib);
 
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline    (ggml_metal_library_t lib, const char * name);
@@ -111,6 +113,8 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_unary             (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_blk        (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -123,6 +127,9 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id         (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -131,6 +138,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope              (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_im2col            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_2d (ggml_metal_library_t lib, const struct ggml_tensor * op);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -193,6 +201,7 @@ struct ggml_metal_device_props {
     bool has_simdgroup_mm;
     bool has_unified_memory;
     bool has_bfloat;
+    bool has_tensor;
     bool use_residency_sets;
     bool use_shared_buffers;
 
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.m b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.m
index 360fbe19..09b1b503 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.m
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-device.m
@@ -21,8 +21,9 @@
 #define GGML_METAL_HAS_RESIDENCY_SETS 1
 #endif
 
-// overload of MTLGPUFamilyMetal3 (not available in some environments)
+// overload of MTLGPUFamilyMetalX (not available in some environments)
 static const NSInteger MTLGPUFamilyMetal3_GGML = 5001;
+static const NSInteger MTLGPUFamilyMetal4_GGML = 5002;
 
 // virtual address for GPU memory allocations
 static atomic_uintptr_t g_addr_device = 0x000000400ULL;
@@ -261,6 +262,10 @@ ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
                     [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
                 }
 
+                if (ggml_metal_device_get_props(dev)->has_tensor) {
+                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+                }
+
 #if GGML_METAL_EMBED_LIBRARY
                 [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
 #endif
@@ -298,6 +303,72 @@ ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
     return res;
 }
 
+ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
+    if (source == NULL) {
+        GGML_LOG_ERROR("%s: source is NULL\n", __func__);
+        return NULL;
+    }
+
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+    id<MTLLibrary> library = nil;
+    NSError * error = nil;
+
+    const int64_t t_start = ggml_time_us();
+
+    NSString * src = [[NSString alloc] initWithBytes:source
+                                              length:strlen(source)
+                                            encoding:NSUTF8StringEncoding];
+    if (!src) {
+        GGML_LOG_ERROR("%s: failed to create NSString from source\n", __func__);
+        return NULL;
+    }
+
+    @autoreleasepool {
+        NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+
+        MTLCompileOptions * options = [MTLCompileOptions new];
+        options.preprocessorMacros = prep;
+
+        library = [device newLibraryWithSource:src options:options error:&error];
+        if (error) {
+            if (verbose) {
+                GGML_LOG_ERROR("%s: error compiling source: %s\n", __func__, [[error description] UTF8String]);
+            } else {
+                GGML_LOG_ERROR("%s: error compiling source\n", __func__);
+            }
+            library = nil;
+        }
+
+        [options release];
+    }
+
+    [src release];
+
+    if (!library) {
+        if (verbose) {
+            GGML_LOG_ERROR("%s: failed to create Metal library from source\n", __func__);
+        }
+
+        return NULL;
+    }
+
+    if (verbose) {
+        GGML_LOG_INFO("%s: compiled in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+    }
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    if (!res) {
+        GGML_LOG_ERROR("%s: calloc failed\n", __func__);
+        return NULL;
+    }
+
+    res->obj       = library;
+    res->device    = device;
+    res->pipelines = ggml_metal_pipelines_init();
+
+    return res;
+}
+
 void ggml_metal_library_free(ggml_metal_library_t lib) {
     if (!lib) {
         return;
@@ -345,9 +416,9 @@ ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t l
         if (!mtl_function) {
             ggml_critical_section_end();
 
-            GGML_LOG_ERROR("%s: error: failed to compile pipeline: base = '%s', name = '%s'\n", __func__, base, name);
+            GGML_LOG_ERROR("%s: failed to compile pipeline: base = '%s', name = '%s'\n", __func__, base, name);
             if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                GGML_LOG_ERROR("%s: %s\n", __func__, [[error description] UTF8String]);
             }
 
             return nil;
@@ -355,13 +426,21 @@ ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t l
 
         res->obj = [lib->device newComputePipelineStateWithFunction:mtl_function error:&error];
 
-        ggml_metal_pipelines_add(lib->pipelines, name, res);
-
         [mtl_function release];
 
         GGML_LOG_DEBUG("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, name, (void *) res->obj,
                 (int) res->obj.maxTotalThreadsPerThreadgroup,
                 (int) res->obj.threadExecutionWidth);
+
+        if (res->obj.maxTotalThreadsPerThreadgroup == 0 || res->obj.threadExecutionWidth == 0) {
+            ggml_critical_section_end();
+
+            GGML_LOG_ERROR("%s: incompatible pipeline %s\n", __func__, name);
+
+            return nil;
+        }
+
+        ggml_metal_pipelines_add(lib->pipelines, name, res);
     }
 
     ggml_critical_section_end();
@@ -469,6 +548,128 @@ ggml_metal_device_t ggml_metal_device_init(void) {
 
             dev->props.has_bfloat  = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
             dev->props.has_bfloat |= [dev->mtl_device supportsFamily:MTLGPUFamilyApple6];
+            if (getenv("GGML_METAL_BF16_DISABLE") != NULL) {
+                dev->props.has_bfloat = false;
+            }
+
+            dev->props.has_tensor = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal4_GGML];
+            if (getenv("GGML_METAL_TENSOR_DISABLE") != NULL) {
+                dev->props.has_tensor = false;
+            }
+
+            // note: disable the tensor API by default for old chips because with the current implementation it is not useful
+            // - M2 Ultra:   ~5% slower
+            // - M4, M4 Max: no significant difference
+            //
+            // TODO: try to update the tensor API kernels to at least match the simdgroup performance
+            if (getenv("GGML_METAL_TENSOR_ENABLE") == NULL &&
+                ![[dev->mtl_device name] containsString:@"M5"] &&
+                ![[dev->mtl_device name] containsString:@"M6"] &&
+                ![[dev->mtl_device name] containsString:@"A19"] &&
+                ![[dev->mtl_device name] containsString:@"A20"]) {
+                GGML_LOG_WARN("%s: tensor API disabled for pre-M5 and pre-A19 devices\n", __func__);
+                dev->props.has_tensor = false;
+            }
+
+            // double-check that the tensor API compiles
+            if (dev->props.has_tensor) {
+                const char * src_tensor_f16 = "\n"
+                    "#include <metal_stdlib> \n"
+                    "#include <metal_tensor> \n"
+                    "#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h> \n"
+                    " \n"
+                    "using namespace metal; \n"
+                    "using namespace mpp::tensor_ops; \n"
+                    " \n"
+                    "kernel void dummy_kernel( \n"
+                    "    tensor<device  half, dextents<int32_t, 2>> A [[buffer(0)]], \n"
+                    "    tensor<device  half, dextents<int32_t, 2>> B [[buffer(1)]], \n"
+                    "    device float * C [[buffer(2)]], \n"
+                    "    uint2 tgid [[threadgroup_position_in_grid]]) \n"
+                    "{ \n"
+                    "    auto tA = A.slice(0, (int)tgid.y); \n"
+                    "    auto tB = B.slice((int)tgid.x, 0); \n"
+                    " \n"
+                    "    matmul2d< \n"
+                    "        matmul2d_descriptor(8, 8, dynamic_extent), \n"
+                    "        execution_simdgroups<4>> mm; \n"
+                    " \n"
+                    "    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>(); \n"
+                    " \n"
+                    "    auto sA = tA.slice(0, 0); \n"
+                    "    auto sB = tB.slice(0, 0); \n"
+                    "    mm.run(sB, sA, cT); \n"
+                    " \n"
+                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
+                    " \n"
+                    "    cT.store(tC); \n"
+                    "}";
+
+                GGML_LOG_INFO("%s: testing tensor API for f16 support\n", __func__);
+                ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_f16, false);
+                if (lib == NULL) {
+                    GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__);
+                    dev->props.has_tensor = false;
+                } else {
+                    ggml_metal_pipeline_t ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
+                    if (!ppl) {
+                        GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__);
+                        dev->props.has_tensor = false;
+                    }
+
+                    ggml_metal_library_free(lib);
+                }
+            }
+
+            // try to compile a dummy kernel to determine if the tensor API is supported for bfloat
+            if (dev->props.has_tensor && dev->props.has_bfloat) {
+                const char * src_tensor_bf16 = "\n"
+                    "#include <metal_stdlib> \n"
+                    "#include <metal_tensor> \n"
+                    "#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h> \n"
+                    " \n"
+                    "using namespace metal; \n"
+                    "using namespace mpp::tensor_ops; \n"
+                    " \n"
+                    "kernel void dummy_kernel( \n"
+                    "    tensor<device bfloat, dextents<int32_t, 2>> A [[buffer(0)]], \n"
+                    "    tensor<device bfloat, dextents<int32_t, 2>> B [[buffer(1)]], \n"
+                    "    device float * C [[buffer(2)]], \n"
+                    "    uint2 tgid [[threadgroup_position_in_grid]]) \n"
+                    "{ \n"
+                    "    auto tA = A.slice(0, (int)tgid.y); \n"
+                    "    auto tB = B.slice((int)tgid.x, 0); \n"
+                    " \n"
+                    "    matmul2d< \n"
+                    "        matmul2d_descriptor(8, 8, dynamic_extent), \n"
+                    "        execution_simdgroups<4>> mm; \n"
+                    " \n"
+                    "    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>(); \n"
+                    " \n"
+                    "    auto sA = tA.slice(0, 0); \n"
+                    "    auto sB = tB.slice(0, 0); \n"
+                    "    mm.run(sB, sA, cT); \n"
+                    " \n"
+                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
+                    " \n"
+                    "    cT.store(tC); \n"
+                    "}";
+
+                GGML_LOG_INFO("%s: testing tensor API for bfloat support\n", __func__);
+                ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_bf16, false);
+                if (lib == NULL) {
+                    GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__);
+                    dev->props.has_bfloat = false;
+                } else {
+                    ggml_metal_pipeline_t ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
+                    if (!ppl) {
+                        GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__);
+                        dev->props.has_bfloat = false;
+                    }
+
+                    ggml_metal_library_free(lib);
+                }
+            }
 
             dev->props.use_residency_sets = true;
 #if defined(GGML_METAL_HAS_RESIDENCY_SETS)
@@ -476,7 +677,6 @@ ggml_metal_device_t ggml_metal_device_init(void) {
 #endif
 
             dev->props.use_shared_buffers = dev->props.has_unified_memory;
-
             if (getenv("GGML_METAL_SHARED_BUFFERS_DISABLE") != NULL) {
                 dev->props.use_shared_buffers = false;
             }
@@ -529,6 +729,7 @@ ggml_metal_device_t ggml_metal_device_init(void) {
             GGML_LOG_INFO("%s: simdgroup matrix mul. = %s\n", __func__, dev->props.has_simdgroup_mm        ? "true" : "false");
             GGML_LOG_INFO("%s: has unified memory    = %s\n", __func__, dev->props.has_unified_memory      ? "true" : "false");
             GGML_LOG_INFO("%s: has bfloat            = %s\n", __func__, dev->props.has_bfloat              ? "true" : "false");
+            GGML_LOG_INFO("%s: has tensor            = %s\n", __func__, dev->props.has_tensor              ? "true" : "false");
             GGML_LOG_INFO("%s: use residency sets    = %s\n", __func__, dev->props.use_residency_sets      ? "true" : "false");
             GGML_LOG_INFO("%s: use shared buffers    = %s\n", __func__, dev->props.use_shared_buffers      ? "true" : "false");
 
@@ -669,6 +870,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_SUM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_CUMSUM:
         case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
@@ -684,6 +886,11 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             return true;
         case GGML_OP_IM2COL:
             return ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32 && (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
+        case GGML_OP_CONV_2D:
+            return ggml_is_contiguous(op->src[0]) &&
+                   op->src[1]->type == GGML_TYPE_F32 &&
+                   op->type == GGML_TYPE_F32 &&
+                   (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
         case GGML_OP_POOL_1D:
             return false;
         case GGML_OP_UPSCALE:
@@ -698,8 +905,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_LEAKY_RELU:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ARGSORT:
-            // TODO: Support arbitrary column width
-            return op->src[0]->ne[0] <= 1024;
+        case GGML_OP_TOP_K:
         case GGML_OP_ARANGE:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
@@ -707,6 +913,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             if (op->src[0]->ne[0] != 32 &&
                 op->src[0]->ne[0] != 40 &&
                 op->src[0]->ne[0] != 64 &&
+                op->src[0]->ne[0] != 72 &&
                 op->src[0]->ne[0] != 80 &&
                 op->src[0]->ne[0] != 96 &&
                 op->src[0]->ne[0] != 112 &&
@@ -783,7 +990,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                                 return false;
                         }
                     case GGML_TYPE_I32:
-                        return op->type == GGML_TYPE_F32;
+                        return op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_I32;
                     default:
                         return false;
                 };
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
index 135266c7..da4c2bb0 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
@@ -1961,6 +1961,7 @@ GGML_TABLE_END()
 #define FC_FLASH_ATTN_EXT_VEC_REDUCE   500
 #define FC_MUL_MV                      600
 #define FC_MUL_MM                      700
+#define FC_ROPE                        800
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPTG 8
@@ -2412,6 +2413,36 @@ typedef struct {
     uint64_t nb2;
 } ggml_metal_kargs_conv_transpose_2d;
 
+typedef struct {
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  IW;
+    int32_t  IH;
+    int32_t  KW;
+    int32_t  KH;
+    int32_t  IC;
+    int32_t  OC;
+    int32_t  OW;
+    int32_t  OH;
+    int32_t  N;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  p0;
+    int32_t  p1;
+    int32_t  d0;
+    int32_t  d1;
+} ggml_metal_kargs_conv_2d;
+
 typedef struct {
     uint64_t  ofs0;
     uint64_t  ofs1;
@@ -2466,6 +2497,45 @@ typedef struct {
     uint64_t nb3;
 } ggml_metal_kargs_sum_rows;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  net0;
+    int64_t  net1;
+    int64_t  net2;
+    int64_t  net3;
+    uint64_t nbt0;
+    uint64_t nbt1;
+    uint64_t nbt2;
+    uint64_t nbt3;
+    bool     outb;
+} ggml_metal_kargs_cumsum_blk;
+
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  net0;
+    int64_t  net1;
+    int64_t  net2;
+    int64_t  net3;
+    uint64_t nbt0;
+    uint64_t nbt1;
+    uint64_t nbt2;
+    uint64_t nbt3;
+} ggml_metal_kargs_cumsum_add;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -2647,10 +2717,38 @@ typedef struct {
 } ggml_metal_kargs_leaky_relu;
 
 typedef struct {
-    int64_t  ncols;
-    int64_t  ncols_pad;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
 } ggml_metal_kargs_argsort;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
+    int32_t  len;
+} ggml_metal_kargs_argsort_merge;
+
 typedef struct {
     int64_t  ne0;
     float    start;
@@ -2688,6 +2786,12 @@ typedef struct {
 
 #include <metal_stdlib>
 
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
 using namespace metal;
 
 #define MAX(x, y) ((x) > (y) ? (x) : (y))
@@ -4421,7 +4525,7 @@ kernel void kernel_op_sum_f32(
 
     float sumf = 0;
 
-    for (int64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
         sumf += src0[i0];
     }
 
@@ -4505,6 +4609,117 @@ typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
 template [[host_name("kernel_sum_rows_f32")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
 template [[host_name("kernel_mean_f32")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
 
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;
+
 template<typename T>
 kernel void kernel_soft_max(
         constant ggml_metal_kargs_soft_max & args,
@@ -6388,6 +6603,8 @@ template [[host_name("kernel_mul_mv_bf16_f32_short")]]  kernel mul_mv_t_t_short_
 template [[host_name("kernel_mul_mv_bf16_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short<bfloat, bfloat>;
 #endif
 
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+
 static float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -6567,11 +6784,27 @@ kernel void kernel_rope_multi(
             const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
             const int sector    = ic % sect_dims;
 
-            float theta_base = (float) pos[i2];
-            if (sector % 3 == 1 && sector < 1 + 3 * args.sect_1) {
-                theta_base = (float) pos[i2 + args.ne02];
-            } else if (sector % 3 == 2 && sector < 2 + 3 * args.sect_2) {
-                theta_base = (float) pos[i2 + args.ne02 * 2];
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 1 + 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 2 + 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                // } else { // e
+                //     theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
             }
             // end of mrope
 
@@ -6801,6 +7034,120 @@ template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
 //template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
 //template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
 
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
 typedef void (conv_transpose_1d_t)(
         constant ggml_metal_kargs_conv_transpose_1d & args,
         device const float * src0,
@@ -7082,115 +7429,127 @@ kernel void kernel_timestep_embedding_f32(
 // bitonic sort implementation following the CUDA kernels as reference
 typedef void (argsort_t)(
         constant   ggml_metal_kargs_argsort & args,
-        device  const float * x,
+        device   const char * src0,
         device      int32_t * dst,
-        threadgroup int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]);
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
 
 template<ggml_sort_order order>
 kernel void kernel_argsort_f32_i32(
         constant   ggml_metal_kargs_argsort & args,
-        device const float  * x,
+        device   const char * src0,
         device      int32_t * dst,
-        threadgroup int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]) {
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
     // bitonic sort
-    int col = tpitg[0];
-    int row = tgpig[1];
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
 
-    if (col >= args.ncols_pad) return;
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
 
-    device const float   * x_row   = x + row * args.ncols;
-    threadgroup int32_t  * dst_row = shared_values;
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 
     // initialize indices
-    dst_row[col] = col;
+    shmem_i32[col] = i00 + col;
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (int k = 2; k <= args.ncols_pad; k *= 2) {
+    for (int k = 2; k <= ntg.x; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (dst_row[col] >= args.ncols ||
-                        (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 } else {
-                    if (dst_row[ixj] >= args.ncols ||
-                        (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 }
             }
+
             threadgroup_barrier(mem_flags::mem_threadgroup);
         }
     }
 
+    const int64_t i0 = ib*args.top_k;
+
     // copy the result to dst without the padding
-    if (col < args.ncols) {
-        dst[row * args.ncols + col] = dst_row[col];
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
     }
 }
 
 typedef void (i32_argsort_t)(
         constant   ggml_metal_kargs_argsort & args,
-        device  const int32_t * x,
+        device  const int32_t * src0,
         device        int32_t * dst,
-        threadgroup   int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]);
+        threadgroup   int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
 
 template<ggml_sort_order order>
 kernel void kernel_argsort_i32_i32(
         constant   ggml_metal_kargs_argsort & args,
-        device const int32_t * x,
+        device const int32_t * src0,
         device       int32_t * dst,
-        threadgroup int32_t  * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]) {
+        threadgroup int32_t  * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
     // bitonic sort
-    int col = tpitg[0];
-    int row = tgpig[1];
+    const int col = tpitg[0];
 
-    if (col >= args.ncols_pad) return;
+    const int i00 = (tgpig[0]/args.ne01)*ntg.x;
+    const int i01 =  tgpig[0]%args.ne01;
+    const int i02 =  tgpig[1];
+    const int i03 =  tgpig[2];
 
-    device const int32_t * x_row   = x + row * args.ncols;
-    threadgroup int32_t  * dst_row = shared_values;
+    device const int32_t * src0_row = (device const int32_t *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 
     // initialize indices
-    dst_row[col] = col;
+    shmem_i32[col] = i00 + col;
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (int k = 2; k <= args.ncols_pad; k *= 2) {
+    for (int k = 2; k <= ntg.x; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (dst_row[col] >= args.ncols ||
-                        (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    if (shmem_i32[col] >= args.ne00 ||
+                        (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 } else {
-                    if (dst_row[ixj] >= args.ncols ||
-                        (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                        (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 }
             }
@@ -7199,8 +7558,10 @@ kernel void kernel_argsort_i32_i32(
     }
 
     // copy the result to dst without the padding
-    if (col < args.ncols) {
-        dst[row * args.ncols + col] = dst_row[col];
+    if (i00 + col < args.ne00) {
+        dst += i00 + args.ne00*i01 + args.ne00*args.ne01*i02 + args.ne00*args.ne01*args.ne02*i03;
+
+        dst[col] = shmem_i32[col];
     }
 }
 
@@ -7209,6 +7570,162 @@ template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_ar
 template [[host_name("kernel_argsort_i32_i32_asc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_ASC>;
 template [[host_name("kernel_argsort_i32_i32_desc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_DESC>;
 
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;
+
 kernel void kernel_leaky_relu_f32(
         constant     ggml_metal_kargs_leaky_relu & args,
         device const float * src0,
@@ -8087,6 +8604,7 @@ typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, hal
 template [[host_name("kernel_flash_attn_ext_f32_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f32_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f32_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f32_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f32_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f32_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f32_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  112, 112>;
@@ -8099,6 +8617,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f16_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f16_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f16_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f16_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f16_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f16_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  112, 112>;
@@ -8112,6 +8631,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
@@ -8125,6 +8645,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 112, 112>;
@@ -8137,6 +8658,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 112, 112>;
@@ -8149,6 +8671,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 112, 112>;
@@ -8161,6 +8684,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 112, 112>;
@@ -8173,6 +8697,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q8_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 112, 112>;
@@ -8184,6 +8709,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_dk576_dv512")]] kernel flash_at
 
 #undef FA_TYPES
 #undef FA_TYPES_BF
+#undef FA_TYPES_F32
 
 constant bool FC_flash_attn_ext_vec_has_mask  [[function_constant(FC_FLASH_ATTN_EXT_VEC + 0)]];
 constant bool FC_flash_attn_ext_vec_has_sinks [[function_constant(FC_FLASH_ATTN_EXT_VEC + 1)]];
@@ -8805,6 +9331,7 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk576_dv512")]] kernel flas
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk576_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 576, 512, 2>;
 
 #undef FA_TYPES
+#undef FA_TYPES_F32
 
 constant int32_t FC_flash_attn_ext_vec_reduce_DV  [[function_constant(FC_FLASH_ATTN_EXT_VEC_REDUCE + 0)]];
 constant int32_t FC_flash_attn_ext_vec_reduce_NWG [[function_constant(FC_FLASH_ATTN_EXT_VEC_REDUCE + 1)]];
@@ -8886,6 +9413,7 @@ template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_
 template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
 template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
 template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
 #if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
 #endif
@@ -10858,17 +11386,6 @@ kernel void kernel_set_rows_f(
 constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
 constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
 
-#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
-#define BLOCK_SIZE_K 32
-#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
-#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
-#define THREAD_PER_BLOCK 128
-#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
-#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
-#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
-#define SG_MAT_ROW 8
-
 // each block_q contains 16*nl weights
 template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
 kernel void kernel_mul_mm(
@@ -10884,18 +11401,48 @@ kernel void kernel_mul_mm(
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
-    const int r0 = tgpig.y;
-    const int r1 = tgpig.x;
+    threadgroup float * sc = (threadgroup float *)(shmem);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
     const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
 
     // if this block is of 64x32 shape or smaller
-    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    const short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
 
     // a thread shouldn't load data outside of the matrix
-    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
 
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im%args.ne12;
+    const int i13 = im/args.ne12;
+
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
     S0_8x8 ma[4];
     S1_8x8 mb[2];
 
@@ -10904,36 +11451,36 @@ kernel void kernel_mul_mm(
     for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
 
-    short il = (tiitg % THREAD_PER_ROW);
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
 
-    const int i12 = im%args.ne12;
-    const int i13 = im/args.ne12;
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
 
-    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
-    const short    offset1 = il/nl;
-
-    device const block_q * x = (device const block_q *)(src0
-        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
-
-    const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
-
-    device const T1 * y = (device const T1 *)(src1
-        + args.nb13*i13
-        + args.nb12*i12
-        + args.nb11*(r1*BLOCK_SIZE_N + thread_col)
-        + args.nb10*iy);
-
-    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
         // load data and store to threadgroup memory
         if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             // no need for dequantization
             for (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
             }
         } else {
             S0_4x4 temp_a;
@@ -10942,91 +11489,203 @@ kernel void kernel_mul_mm(
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             FOR_UNROLL (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
             }
         }
 
         if (FC_mul_mm_bc_inp) {
             for (short i = 0; i < 8; ++i) {
-                sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
             }
         } else {
-            *(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            const short dx = sx;
+            const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
         }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
-        y += BLOCK_SIZE_K;
+
+        y += NK;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
+#ifndef GGML_METAL_HAS_TENSOR
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
 
-        #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(4)
-            for (short i = 0; i < 4; i++) {
-                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
-            }
-
-            #pragma unroll(2)
-            for (short i = 0; i < 2; i++) {
-                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
             }
 
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(8)
-            for (short i = 0; i < 8; i++){
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
 
-            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
-            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsma += 8*64;
+            lsmb += 4*64;
         }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
     }
 
-    if (!FC_mul_mm_bc_out || ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1)) {
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
         // if no bounds checks on the output are needed, we can directly write to device memory
+#ifdef GGML_METAL_HAS_TENSOR
         device float * C = (device float *) dst +
-            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
-            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+            r0 + \
+            r1 * args.ne0 + im*args.ne1*args.ne0;
+
+        auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(args.ne0, NR1));
+        cT.store(tC);
+#else
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
 
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
         }
+#endif
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+#ifdef GGML_METAL_HAS_TENSOR
+        auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+        cT.store(tC);
+#else
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
         }
+#endif
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         if (sgitg == 0) {
-            for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.ne0 + im*args.ne1*args.ne0;
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
                 device float4 * D4 = (device float4 *) D;
 
-                threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
+                threadgroup float  * C  = temp_str + (j*NR0);
                 threadgroup float4 * C4 = (threadgroup float4 *) C;
 
                 int i = 0;
-                for (; i < n_rows/4; i++) {
+                for (; i < nr0/4; i++) {
                     *(D4 + i) = *(C4 + i);
                 }
 
                 i *= 4;
-                for (; i < n_rows; i++) {
+                for (; i < nr0; i++) {
                     *(D + i) = *(C + i);
                 }
             }
@@ -11111,31 +11770,63 @@ kernel void kernel_mul_mm_id(
         ushort tiitg[[thread_index_in_threadgroup]],
         ushort tiisg[[thread_index_in_simdgroup]],
         ushort sgitg[[simdgroup_index_in_threadgroup]]) {
-
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
-    const int r0 = tgpig.y;
-    const int r1 = tgpig.x;
+    threadgroup float * sc = (threadgroup float *)(shmem);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
     const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
 
     device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
     device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
 
     const int32_t neh1 = tpe_u32[im];
 
-    if (r1*BLOCK_SIZE_N >= neh1) {
+    if (r1 >= neh1) {
         return;
     }
 
     // if this block is of 64x32 shape or smaller
-    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    const short n_cols = (    neh1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (    neh1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
 
     // a thread shouldn't load data outside of the matrix
-    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
 
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
     S0_8x8 ma[4];
     S1_8x8 mb[2];
 
@@ -11144,39 +11835,36 @@ kernel void kernel_mul_mm_id(
     for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
 
-    short il = (tiitg % THREAD_PER_ROW);
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
 
-    const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + thread_col];
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
 
-    const short i11 = (id % args.ne20) % args.ne11;
-    const short i12 = (id / args.ne20);
-    const short i13 = 0;
-
-    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
-    const short    offset1 = il/nl;
-
-    device const block_q * x = (device const block_q *)(src0
-        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
-
-    const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
-
-    device const T1 * y = (device const T1 *)(src1
-        + args.nb13*i13
-        + args.nb12*i12
-        + args.nb11*i11
-        + args.nb10*iy);
-
-    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
         // load data and store to threadgroup memory
         if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             // no need for dequantization
             for (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
             }
         } else {
             S0_4x4 temp_a;
@@ -11185,85 +11873,188 @@ kernel void kernel_mul_mm_id(
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             FOR_UNROLL (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
             }
         }
 
         if (FC_mul_mm_bc_inp) {
             for (short i = 0; i < 8; ++i) {
-                sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
             }
         } else {
-            *(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            const short dx = sx;
+            const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
         }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
-        y += BLOCK_SIZE_K;
+
+        y += NK;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
+#ifndef GGML_METAL_HAS_TENSOR
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
 
-        #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
-            #pragma unroll(4)
-            for (short i = 0; i < 4; i++) {
-                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
             }
 
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(2)
-            for (short i = 0; i < 2; i++) {
-                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
             }
 
-            #pragma unroll(8)
-            for (short i = 0; i < 8; i++){
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
 
-            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
-            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsma += 8*64;
+            lsmb += 4*64;
         }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
     }
 
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                 + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
 
-    #pragma unroll(8)
     for (short i = 0; i < 8; i++) {
-        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
     }
+#endif
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (short j = sgitg; j < n_cols; j += 4) {
-        const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + j];
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
 
         const short ide = id % args.ne20;
         const short idt = id / args.ne20;
 
-        device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
         device float4 * D4 = (device float4 *) D;
 
-        threadgroup float  * C  = (threadgroup float  *) shmem + (j*BLOCK_SIZE_M);
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
         threadgroup float4 * C4 = (threadgroup float4 *) C;
 
         int i = tiisg;
-        for (; i < n_rows/4; i += 32) {
+        for (; i < nr0/4; i += 32) {
             *(D4 + i) = *(C4 + i);
         }
 
-        i = (4*(n_rows/4)) + tiisg;
-        for (; i < n_rows; i += 32) {
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
             *(D + i) = *(C + i);
         }
     }
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-impl.h b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-impl.h
index 96f43d26..342dc4f8 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-impl.h
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-impl.h
@@ -76,6 +76,7 @@
 #define FC_FLASH_ATTN_EXT_VEC_REDUCE   500
 #define FC_MUL_MV                      600
 #define FC_MUL_MM                      700
+#define FC_ROPE                        800
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPTG 8
@@ -527,6 +528,36 @@ typedef struct {
     uint64_t nb2;
 } ggml_metal_kargs_conv_transpose_2d;
 
+typedef struct {
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  IW;
+    int32_t  IH;
+    int32_t  KW;
+    int32_t  KH;
+    int32_t  IC;
+    int32_t  OC;
+    int32_t  OW;
+    int32_t  OH;
+    int32_t  N;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  p0;
+    int32_t  p1;
+    int32_t  d0;
+    int32_t  d1;
+} ggml_metal_kargs_conv_2d;
+
 typedef struct {
     uint64_t  ofs0;
     uint64_t  ofs1;
@@ -581,6 +612,45 @@ typedef struct {
     uint64_t nb3;
 } ggml_metal_kargs_sum_rows;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  net0;
+    int64_t  net1;
+    int64_t  net2;
+    int64_t  net3;
+    uint64_t nbt0;
+    uint64_t nbt1;
+    uint64_t nbt2;
+    uint64_t nbt3;
+    bool     outb;
+} ggml_metal_kargs_cumsum_blk;
+
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  net0;
+    int64_t  net1;
+    int64_t  net2;
+    int64_t  net3;
+    uint64_t nbt0;
+    uint64_t nbt1;
+    uint64_t nbt2;
+    uint64_t nbt3;
+} ggml_metal_kargs_cumsum_add;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -762,10 +832,38 @@ typedef struct {
 } ggml_metal_kargs_leaky_relu;
 
 typedef struct {
-    int64_t  ncols;
-    int64_t  ncols_pad;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
 } ggml_metal_kargs_argsort;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    int32_t  top_k;
+    int32_t  len;
+} ggml_metal_kargs_argsort_merge;
+
 typedef struct {
     int64_t  ne0;
     float    start;
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.cpp
index 7a85edbd..9871e976 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.cpp
@@ -10,6 +10,8 @@
 
 #include <cassert>
 #include <algorithm>
+#include <limits>
+#include <cmath>
 
 static ggml_metal_buffer_id ggml_metal_get_buffer_id(const ggml_tensor * t) {
     if (!t) {
@@ -310,6 +312,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_sum_rows(ctx, idx);
             } break;
+        case GGML_OP_CUMSUM:
+            {
+                n_fuse = ggml_metal_op_cumsum(ctx, idx);
+            } break;
         case GGML_OP_SOFT_MAX:
             {
                 n_fuse = ggml_metal_op_soft_max(ctx, idx);
@@ -364,6 +370,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_im2col(ctx, idx);
             } break;
+        case GGML_OP_CONV_2D:
+            {
+                n_fuse = ggml_metal_op_conv_2d(ctx, idx);
+            } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
                 n_fuse = ggml_metal_op_conv_transpose_1d(ctx, idx);
@@ -396,6 +406,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_argsort(ctx, idx);
             } break;
+        case GGML_OP_TOP_K:
+            {
+                n_fuse = ggml_metal_op_top_k(ctx, idx);
+            } break;
         case GGML_OP_LEAKY_RELU:
             {
                 n_fuse = ggml_metal_op_leaky_relu(ctx, idx);
@@ -534,7 +548,7 @@ int ggml_metal_op_repeat(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_repeat(lib, op->type);
 
@@ -580,7 +594,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
@@ -689,7 +703,7 @@ int ggml_metal_op_scale(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float scale;
     float bias;
@@ -728,7 +742,7 @@ int ggml_metal_op_clamp(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float min;
     float max;
@@ -767,7 +781,7 @@ int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     int64_t n = ggml_nelements(op);
 
@@ -797,7 +811,7 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     if (op->src[1]) {
         GGML_ASSERT(ggml_are_same_shape(op->src[0], op->src[1]));
@@ -829,18 +843,6 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
 
     const int32_t nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00/2);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-    //if (src1) {
-    //    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-    //} else {
-    //    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-    //}
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:3];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
@@ -902,7 +904,7 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_sum_rows args = {
         /*.ne00 =*/ ne00,
@@ -936,14 +938,6 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
 
     const size_t smem = ggml_metal_pipeline_get_smem(pipeline);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:0];
-    //[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-    //[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
@@ -956,6 +950,149 @@ int ggml_metal_op_sum_rows(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_cumsum(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_pipeline_t pipeline_blk = ggml_metal_library_get_pipeline_cumsum_blk(lib, op);
+
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_blk)) {
+        nth *= 2;
+    }
+
+    GGML_ASSERT(ne00 <= nth*nth);
+
+    const int64_t net0 = (ne00 + nth - 1) / nth;
+    const int64_t net1 = ne01;
+    const int64_t net2 = ne02;
+    const int64_t net3 = ne03;
+
+    const uint64_t nbt0 = sizeof(float);
+    const uint64_t nbt1 = net0*nbt0;
+    const uint64_t nbt2 = net1*nbt1;
+    const uint64_t nbt3 = net2*nbt2;
+
+    const size_t smem = GGML_PAD(32*sizeof(float), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += ggml_nbytes(op);
+
+    {
+        ggml_metal_kargs_cumsum_blk args = {
+            /*.ne00 =*/ ne00,
+            /*.ne01 =*/ ne01,
+            /*.ne02 =*/ ne02,
+            /*.ne03 =*/ ne03,
+            /*.nb00 =*/ nb00,
+            /*.nb01 =*/ nb01,
+            /*.nb02 =*/ nb02,
+            /*.nb03 =*/ nb03,
+            /*.net0 =*/ net0,
+            /*.net1 =*/ net1,
+            /*.net2 =*/ net2,
+            /*.net3 =*/ net3,
+            /*.nbt0 =*/ nbt0,
+            /*.nbt1 =*/ nbt1,
+            /*.nbt2 =*/ nbt2,
+            /*.nbt3 =*/ nbt3,
+            /*.outb =*/ ne00 > nth,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_blk);
+        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
+
+        ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, net0*ne01, ne02, ne03, nth, 1, 1);
+    }
+
+    if (ne00 > nth) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        {
+            ggml_metal_kargs_cumsum_blk args = {
+                /*.ne00 =*/ net0,
+                /*.ne01 =*/ net1,
+                /*.ne02 =*/ net2,
+                /*.ne03 =*/ net3,
+                /*.nb00 =*/ nbt0,
+                /*.nb01 =*/ nbt1,
+                /*.nb02 =*/ nbt2,
+                /*.nb03 =*/ nbt3,
+                /*.net0 =*/ net0,
+                /*.net1 =*/ net1,
+                /*.net2 =*/ net2,
+                /*.net3 =*/ net3,
+                /*.nbt0 =*/ nbt0,
+                /*.nbt1 =*/ nbt1,
+                /*.nbt2 =*/ nbt2,
+                /*.nbt3 =*/ nbt3,
+                /*.outb =*/ false,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline_blk);
+            ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 3);
+
+            ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, net1, net2, net3, nth, 1, 1);
+        }
+
+        ggml_metal_op_concurrency_reset(ctx);
+
+        {
+            ggml_metal_pipeline_t pipeline_add = ggml_metal_library_get_pipeline_cumsum_add(lib, op);
+
+            ggml_metal_kargs_cumsum_add args = {
+                /*.ne00 =*/ ne00,
+                /*.ne01 =*/ ne01,
+                /*.ne02 =*/ ne02,
+                /*.ne03 =*/ ne03,
+                /*.nb00 =*/ nb00,
+                /*.nb01 =*/ nb01,
+                /*.nb02 =*/ nb02,
+                /*.nb03 =*/ nb03,
+                /*.net0 =*/ net0,
+                /*.net1 =*/ net1,
+                /*.net2 =*/ net2,
+                /*.net3 =*/ net3,
+                /*.nbt0 =*/ nbt0,
+                /*.nbt1 =*/ nbt1,
+                /*.nbt2 =*/ nbt2,
+                /*.nbt3 =*/ nbt3,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline_add);
+            ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_tmp, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_dst, 2);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, net0*ne01, ne02, ne03, nth, 1, 1);
+        }
+    }
+
+    return 1;
+}
+
 int ggml_metal_op_get_rows(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -967,7 +1104,7 @@ int ggml_metal_op_get_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_get_rows(lib, op->src[0]->type);
 
@@ -1012,7 +1149,7 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op->src[1]->type, op->type);
 
@@ -1076,7 +1213,7 @@ int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float scale;
     float max_bias;
@@ -1164,7 +1301,7 @@ int ggml_metal_op_ssm_conv(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_ssm_conv args = {
         /*.ne00 =*/ ne00,
@@ -1219,7 +1356,7 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne6, op->src[6], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb6, op->src[6], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const ggml_tensor * src3 = op->src[3];
     const ggml_tensor * src4 = op->src[4];
@@ -1305,7 +1442,7 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int64_t B = op->op == GGML_OP_RWKV_WKV6 ? op->src[5]->ne[1] : op->src[6]->ne[1];
     const int64_t T = op->src[0]->ne[2];
@@ -1346,7 +1483,7 @@ int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_cpy(lib, op->src[0]->type, op->type);
 
@@ -1419,7 +1556,7 @@ int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t * opts = op->op_params;
     ggml_op_pool op_pool = (ggml_op_pool) opts[0];
@@ -1483,7 +1620,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     GGML_ASSERT(ne00 == ne10);
 
@@ -1724,7 +1861,7 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     // src2 = ids
     GGML_ASSERT(op->src[2]->type == GGML_TYPE_I32);
@@ -1970,7 +2107,9 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
     const bool has_mask = op->src[3] != nullptr;
 
     if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
-        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+        // note: always reserve the padding space to avoid graph reallocations
+        //const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+        const bool has_kvpad = true;
 
         if (has_kvpad) {
             res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
@@ -1979,7 +2118,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
                 (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
         }
     } else {
-        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+        //const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+        const bool has_kvpad = true;
 
         if (has_kvpad) {
             res += OP_FLASH_ATTN_EXT_NCPSG*(
@@ -2015,9 +2155,10 @@ size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
     const bool is_vec = ggml_metal_op_flash_attn_ext_use_vec(op);
 
     // this optimization is not useful for the vector kernels
-    if (is_vec) {
-        return res;
-    }
+    // note: always reserve the blk buffer to avoid graph reallocations
+    //if (is_vec) {
+    //    return res;
+    //}
 
     const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
     const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;
@@ -2044,13 +2185,16 @@ size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
 
     size_t res = 0;
 
-    if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+    // note: always reserve the temp buffer to avoid graph reallocations
+    //if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+    if (true) {
         const int64_t nwg = 32;
+        const int64_t ne01_max = std::min(ne01, 32);
 
         // temp buffer for writing the results from each workgroup
         // - ne20: the size of the Value head
         // -  + 2: the S and M values for each intermediate result
-        res += ggml_type_size(GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+        res += ggml_type_size(GGML_TYPE_F32)*(ne01_max*ne02*ne03*nwg*(ne20 + 2));
     }
 
     return res;
@@ -2179,8 +2323,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
 
             need_sync = true;
-        } else {
-            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
         }
 
         if (has_mask) {
@@ -2210,8 +2352,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_dispatch_threadgroups(enc, nblk0, nblk1, ne32*ne33, 32, 1, 1);
 
             need_sync = true;
-        } else {
-            assert(ggml_metal_op_flash_attn_ext_extra_blk(op) == 0);
         }
 
         if (need_sync) {
@@ -2351,8 +2491,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
 
             need_sync = true;
-        } else {
-            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
         }
 
         if (need_sync) {
@@ -2683,7 +2821,7 @@ int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float eps;
     memcpy(&eps, op->op_params, sizeof(float));
@@ -2731,7 +2869,7 @@ int ggml_metal_op_group_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t ngrp = ((const int32_t *) op->op_params)[0];
 
@@ -2786,7 +2924,7 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float eps;
     memcpy(&eps, op->op_params, sizeof(float));
@@ -2922,7 +3060,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     // make sure we have one or more position id(ne10) per token(ne02)
     GGML_ASSERT(ne10 % ne02 == 0);
@@ -3016,7 +3154,7 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
     const int32_t s1 = ((const int32_t *)(op->op_params))[1];
@@ -3077,6 +3215,84 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_conv_2d(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t *) op->op_params)[0];
+    const int32_t s1 = ((const int32_t *) op->op_params)[1];
+    const int32_t p0 = ((const int32_t *) op->op_params)[2];
+    const int32_t p1 = ((const int32_t *) op->op_params)[3];
+    const int32_t d0 = ((const int32_t *) op->op_params)[4];
+    const int32_t d1 = ((const int32_t *) op->op_params)[5];
+
+    ggml_metal_kargs_conv_2d args = {
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+        /*.IW   =*/ ne10,
+        /*.IH   =*/ ne11,
+        /*.KW   =*/ ne00,
+        /*.KH   =*/ ne01,
+        /*.IC   =*/ ne02,
+        /*.OC   =*/ ne03,
+        /*.OW   =*/ ne0,
+        /*.OH   =*/ ne1,
+        /*.N    =*/ ne3,
+        /*.s0   =*/ s0,
+        /*.s1   =*/ s1,
+        /*.p0   =*/ p0,
+        /*.p1   =*/ p1,
+        /*.d0   =*/ d0,
+        /*.d1   =*/ d1,
+    };
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_conv_2d(lib, op);
+
+    int nth = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
+    nth = std::min(nth, 256);
+    nth = std::max(nth, 1);
+
+    const uint64_t n_out = ggml_nelements(op);
+
+    uint64_t tg = (n_out + nth - 1)/nth;
+    tg = std::max<uint64_t>(tg, 1);
+    tg = std::min<uint64_t>(tg, (uint64_t) std::numeric_limits<int>::max());
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, tg, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -3088,7 +3304,7 @@ int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
 
@@ -3133,7 +3349,7 @@ int ggml_metal_op_conv_transpose_2d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int32_t s0 = ((const int32_t *)(op->op_params))[0];
 
@@ -3187,7 +3403,7 @@ int ggml_metal_op_upscale(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const float sf0 = (float)ne0/op->src[0]->ne[0];
     const float sf1 = (float)ne1/op->src[0]->ne[1];
@@ -3240,7 +3456,7 @@ int ggml_metal_op_pad(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_pad args = {
         /*.ne00 =*/ ne00,
@@ -3284,7 +3500,7 @@ int ggml_metal_op_pad_reflect_1d(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_pad_reflect_1d args = {
         /*.ne00 =*/ ne00,
@@ -3328,7 +3544,7 @@ int ggml_metal_op_arange(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_t enc = ctx->enc;
 
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float start;
     float step;
@@ -3346,12 +3562,6 @@ int ggml_metal_op_arange(ggml_metal_op_t ctx, int idx) {
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_arange(lib, op);
 
-    //[encoder setComputePipelineState:pipeline];
-    //[encoder setBuffer:id_dst  offset:offs_dst  atIndex:0];
-    //[encoder setBytes:&args length:sizeof(args) atIndex:1];
-
-    //[encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op), 1);
@@ -3370,7 +3580,7 @@ int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     const int dim        = op->op_params[0];
     const int max_period = op->op_params[1];
@@ -3404,7 +3614,7 @@ int ggml_metal_op_argmax(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_kargs_argmax args = {
         /*.ne00 = */ ne00,
@@ -3440,38 +3650,215 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
 
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
-
-    // bitonic sort requires the number of elements to be power of 2
-    int64_t ne00_padded = 1;
-    while (ne00_padded < ne00) {
-        ne00_padded *= 2;
-    }
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_argsort(lib, op);
 
-    const int64_t nrows = ggml_nrows(op->src[0]);
+    // bitonic sort requires the number of elements to be power of 2
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    const int npr = (ne00 + nth - 1)/nth;
 
     // Metal kernels require the buffer size to be multiple of 16 bytes
     // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
-    const size_t smem = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
+    const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += ggml_nbytes(op);
+
+    if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
+        std::swap(bid_dst, bid_tmp);
+    }
 
     ggml_metal_kargs_argsort args = {
-        /*.ncols =*/ ne00,
-        /*.ncols_pad =*/ ne00_padded
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ nth,
     };
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
 
     ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, 1, nrows, 1, ne00_padded, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
+
+    ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_argsort_merge(lib, op);
+
+    int len = nth;
+
+    while (len < ne00) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        ggml_metal_kargs_argsort_merge args_merge = {
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ ne00,
+            /*.len   =*/ len,
+        };
+
+        // merges per row
+        const int nm = (ne00 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge));
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
+        ggml_metal_encoder_set_bytes   (enc, &args_merge, sizeof(args_merge), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, nm*ne01, ne02, ne03, nth, 1, 1);
+
+        std::swap(bid_dst, bid_tmp);
+
+        len <<= 1;
+    }
+
+    return 1;
+}
+
+int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_top_k(lib, op);
+
+    // bitonic sort requires the number of elements to be power of 2
+    int nth = 1;
+    while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        nth *= 2;
+    }
+
+    // blocks per row
+    const int npr = (ne00 + nth - 1)/nth;
+
+    const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_dst;
+    bid_tmp.offs += sizeof(int32_t)*ggml_nelements(op->src[0]);
+
+    if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
+        std::swap(bid_dst, bid_tmp);
+    }
+
+    const int top_k = ne0;
+
+    ggml_metal_kargs_argsort args = {
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.top_k =*/ std::min(nth, top_k), // for each block, keep just the top_k indices
+    };
+
+    if (npr > 1) {
+        args.ne0 = (npr - 1)*args.top_k + std::min(ne00 - (npr - 1)*nth, args.top_k);
+    }
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
+
+    ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_top_k_merge(lib, op);
+
+    int len = args.top_k;
+
+    while (len < args.ne0) {
+        ggml_metal_op_concurrency_reset(ctx);
+
+        // merges per row
+        const int nm = (args.ne0 + 2*len - 1) / (2*len);
+
+        const int nth = std::min(512, std::min(len, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge)));
+
+        ggml_metal_kargs_argsort_merge args_merge = {
+            /*.ne00  =*/ ne00,
+            /*.ne01  =*/ ne01,
+            /*.ne02  =*/ ne02,
+            /*.ne03  =*/ ne03,
+            /*.nb00  =*/ nb00,
+            /*.nb01  =*/ nb01,
+            /*.nb02  =*/ nb02,
+            /*.nb03  =*/ nb03,
+            /*.ne0   =*/ args.ne0,
+            /*.ne1   =*/ ne1,
+            /*.ne2   =*/ ne2,
+            /*.ne3   =*/ ne3,
+            /*.top_k =*/ nm == 1 ? top_k : args.ne0, // the final merge outputs top_k elements
+            /*.len   =*/ len,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
+        ggml_metal_encoder_set_bytes   (enc, &args_merge, sizeof(args_merge), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+        ggml_metal_encoder_set_buffer  (enc, bid_tmp,  3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, nm*ne01, ne02, ne03, nth, 1, 1);
+
+        std::swap(bid_dst, bid_tmp);
+
+        len <<= 1;
+    }
 
     return 1;
 }
@@ -3485,7 +3872,7 @@ int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     float slope;
     memcpy(&slope, op->op_params, sizeof(float));
@@ -3521,7 +3908,7 @@ int ggml_metal_op_opt_step_adamw(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_opt_step_adamw(lib, op);
 
@@ -3557,7 +3944,7 @@ int ggml_metal_op_opt_step_sgd(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
     ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_opt_step_sgd(lib, op);
 
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.h b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.h
index 0d9cb8af..b5546146 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.h
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-ops.h
@@ -52,6 +52,7 @@ int ggml_metal_op_unary             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_glu               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_sum               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_sum_rows          (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
@@ -70,6 +71,7 @@ int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rope              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_im2col            (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_conv_2d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_1d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_2d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_upscale           (ggml_metal_op_t ctx, int idx);
@@ -79,6 +81,7 @@ int ggml_metal_op_arange            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argmax            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argsort           (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_top_k             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_leaky_relu        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_adamw    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_sgd      (ggml_metal_op_t ctx, int idx);
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.cpp b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.cpp
index 032dee76..f6f8f7a1 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.cpp
@@ -199,6 +199,15 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
                 res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
                 res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
             } break;
+        case GGML_OP_CUMSUM:
+        case GGML_OP_ARGSORT:
+            {
+                res *= 2;
+            } break;
+        case GGML_OP_TOP_K:
+            {
+                res = 2*sizeof(int32_t)*ggml_nelements(tensor->src[0]);
+            } break;
         default:
             break;
     }
diff --git a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.metal b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.metal
index 65a3183c..a489de43 100644
--- a/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.metal
@@ -9,6 +9,12 @@ __embed_ggml-common.h__
 
 #include <metal_stdlib>
 
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
 using namespace metal;
 
 #define MAX(x, y) ((x) > (y) ? (x) : (y))
@@ -1742,7 +1748,7 @@ kernel void kernel_op_sum_f32(
 
     float sumf = 0;
 
-    for (int64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
         sumf += src0[i0];
     }
 
@@ -1826,6 +1832,117 @@ typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
 template [[host_name("kernel_sum_rows_f32")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
 template [[host_name("kernel_mean_f32")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
 
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;
+
 template<typename T>
 kernel void kernel_soft_max(
         constant ggml_metal_kargs_soft_max & args,
@@ -3709,6 +3826,8 @@ template [[host_name("kernel_mul_mv_bf16_f32_short")]]  kernel mul_mv_t_t_short_
 template [[host_name("kernel_mul_mv_bf16_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short<bfloat, bfloat>;
 #endif
 
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+
 static float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -3888,11 +4007,27 @@ kernel void kernel_rope_multi(
             const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
             const int sector    = ic % sect_dims;
 
-            float theta_base = (float) pos[i2];
-            if (sector % 3 == 1 && sector < 1 + 3 * args.sect_1) {
-                theta_base = (float) pos[i2 + args.ne02];
-            } else if (sector % 3 == 2 && sector < 2 + 3 * args.sect_2) {
-                theta_base = (float) pos[i2 + args.ne02 * 2];
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 1 + 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 2 + 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                // } else { // e
+                //     theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
             }
             // end of mrope
 
@@ -4122,6 +4257,120 @@ template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
 //template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
 //template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
 
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
 typedef void (conv_transpose_1d_t)(
         constant ggml_metal_kargs_conv_transpose_1d & args,
         device const float * src0,
@@ -4403,115 +4652,127 @@ kernel void kernel_timestep_embedding_f32(
 // bitonic sort implementation following the CUDA kernels as reference
 typedef void (argsort_t)(
         constant   ggml_metal_kargs_argsort & args,
-        device  const float * x,
+        device   const char * src0,
         device      int32_t * dst,
-        threadgroup int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]);
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
 
 template<ggml_sort_order order>
 kernel void kernel_argsort_f32_i32(
         constant   ggml_metal_kargs_argsort & args,
-        device const float  * x,
+        device   const char * src0,
         device      int32_t * dst,
-        threadgroup int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]) {
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
     // bitonic sort
-    int col = tpitg[0];
-    int row = tgpig[1];
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
 
-    if (col >= args.ncols_pad) return;
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
 
-    device const float   * x_row   = x + row * args.ncols;
-    threadgroup int32_t  * dst_row = shared_values;
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 
     // initialize indices
-    dst_row[col] = col;
+    shmem_i32[col] = i00 + col;
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (int k = 2; k <= args.ncols_pad; k *= 2) {
+    for (int k = 2; k <= ntg.x; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (dst_row[col] >= args.ncols ||
-                        (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 } else {
-                    if (dst_row[ixj] >= args.ncols ||
-                        (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 }
             }
+
             threadgroup_barrier(mem_flags::mem_threadgroup);
         }
     }
 
+    const int64_t i0 = ib*args.top_k;
+
     // copy the result to dst without the padding
-    if (col < args.ncols) {
-        dst[row * args.ncols + col] = dst_row[col];
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
     }
 }
 
 typedef void (i32_argsort_t)(
         constant   ggml_metal_kargs_argsort & args,
-        device  const int32_t * x,
+        device  const int32_t * src0,
         device        int32_t * dst,
-        threadgroup   int32_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]);
+        threadgroup   int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
 
 template<ggml_sort_order order>
 kernel void kernel_argsort_i32_i32(
         constant   ggml_metal_kargs_argsort & args,
-        device const int32_t * x,
+        device const int32_t * src0,
         device       int32_t * dst,
-        threadgroup int32_t  * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]]) {
+        threadgroup int32_t  * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
     // bitonic sort
-    int col = tpitg[0];
-    int row = tgpig[1];
+    const int col = tpitg[0];
 
-    if (col >= args.ncols_pad) return;
+    const int i00 = (tgpig[0]/args.ne01)*ntg.x;
+    const int i01 =  tgpig[0]%args.ne01;
+    const int i02 =  tgpig[1];
+    const int i03 =  tgpig[2];
 
-    device const int32_t * x_row   = x + row * args.ncols;
-    threadgroup int32_t  * dst_row = shared_values;
+    device const int32_t * src0_row = (device const int32_t *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
 
     // initialize indices
-    dst_row[col] = col;
+    shmem_i32[col] = i00 + col;
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (int k = 2; k <= args.ncols_pad; k *= 2) {
+    for (int k = 2; k <= ntg.x; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (dst_row[col] >= args.ncols ||
-                        (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    if (shmem_i32[col] >= args.ne00 ||
+                        (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 } else {
-                    if (dst_row[ixj] >= args.ncols ||
-                        (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                        (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
                     ) {
-                        SWAP(dst_row[col], dst_row[ixj]);
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
                     }
                 }
             }
@@ -4520,8 +4781,10 @@ kernel void kernel_argsort_i32_i32(
     }
 
     // copy the result to dst without the padding
-    if (col < args.ncols) {
-        dst[row * args.ncols + col] = dst_row[col];
+    if (i00 + col < args.ne00) {
+        dst += i00 + args.ne00*i01 + args.ne00*args.ne01*i02 + args.ne00*args.ne01*args.ne02*i03;
+
+        dst[col] = shmem_i32[col];
     }
 }
 
@@ -4530,6 +4793,162 @@ template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_ar
 template [[host_name("kernel_argsort_i32_i32_asc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_ASC>;
 template [[host_name("kernel_argsort_i32_i32_desc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_DESC>;
 
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;
+
 kernel void kernel_leaky_relu_f32(
         constant     ggml_metal_kargs_leaky_relu & args,
         device const float * src0,
@@ -5408,6 +5827,7 @@ typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, hal
 template [[host_name("kernel_flash_attn_ext_f32_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f32_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f32_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f32_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f32_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f32_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f32_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  112, 112>;
@@ -5420,6 +5840,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f16_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f16_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f16_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f16_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f16_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f16_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  112, 112>;
@@ -5433,6 +5854,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
@@ -5446,6 +5868,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 112, 112>;
@@ -5458,6 +5881,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 112, 112>;
@@ -5470,6 +5894,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 112, 112>;
@@ -5482,6 +5907,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 112, 112>;
@@ -5494,6 +5920,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q8_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 112, 112>;
@@ -5505,6 +5932,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_dk576_dv512")]] kernel flash_at
 
 #undef FA_TYPES
 #undef FA_TYPES_BF
+#undef FA_TYPES_F32
 
 constant bool FC_flash_attn_ext_vec_has_mask  [[function_constant(FC_FLASH_ATTN_EXT_VEC + 0)]];
 constant bool FC_flash_attn_ext_vec_has_sinks [[function_constant(FC_FLASH_ATTN_EXT_VEC + 1)]];
@@ -6126,6 +6554,7 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk576_dv512")]] kernel flas
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk576_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 576, 512, 2>;
 
 #undef FA_TYPES
+#undef FA_TYPES_F32
 
 constant int32_t FC_flash_attn_ext_vec_reduce_DV  [[function_constant(FC_FLASH_ATTN_EXT_VEC_REDUCE + 0)]];
 constant int32_t FC_flash_attn_ext_vec_reduce_NWG [[function_constant(FC_FLASH_ATTN_EXT_VEC_REDUCE + 1)]];
@@ -6207,6 +6636,7 @@ template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_
 template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
 template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
 template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
 #if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
 #endif
@@ -8179,17 +8609,6 @@ kernel void kernel_set_rows_f(
 constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
 constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
 
-#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
-#define BLOCK_SIZE_K 32
-#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
-#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
-#define THREAD_PER_BLOCK 128
-#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
-#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
-#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
-#define SG_MAT_ROW 8
-
 // each block_q contains 16*nl weights
 template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
 kernel void kernel_mul_mm(
@@ -8205,18 +8624,48 @@ kernel void kernel_mul_mm(
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
-    const int r0 = tgpig.y;
-    const int r1 = tgpig.x;
+    threadgroup float * sc = (threadgroup float *)(shmem);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
     const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
 
     // if this block is of 64x32 shape or smaller
-    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    const short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
 
     // a thread shouldn't load data outside of the matrix
-    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
 
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im%args.ne12;
+    const int i13 = im/args.ne12;
+
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
     S0_8x8 ma[4];
     S1_8x8 mb[2];
 
@@ -8225,36 +8674,36 @@ kernel void kernel_mul_mm(
     for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
 
-    short il = (tiitg % THREAD_PER_ROW);
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
 
-    const int i12 = im%args.ne12;
-    const int i13 = im/args.ne12;
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
 
-    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
-    const short    offset1 = il/nl;
-
-    device const block_q * x = (device const block_q *)(src0
-        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
-
-    const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
-
-    device const T1 * y = (device const T1 *)(src1
-        + args.nb13*i13
-        + args.nb12*i12
-        + args.nb11*(r1*BLOCK_SIZE_N + thread_col)
-        + args.nb10*iy);
-
-    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
         // load data and store to threadgroup memory
         if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             // no need for dequantization
             for (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
             }
         } else {
             S0_4x4 temp_a;
@@ -8263,91 +8712,203 @@ kernel void kernel_mul_mm(
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             FOR_UNROLL (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
             }
         }
 
         if (FC_mul_mm_bc_inp) {
             for (short i = 0; i < 8; ++i) {
-                sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
             }
         } else {
-            *(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            const short dx = sx;
+            const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
         }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
-        y += BLOCK_SIZE_K;
+
+        y += NK;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
+#ifndef GGML_METAL_HAS_TENSOR
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
 
-        #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(4)
-            for (short i = 0; i < 4; i++) {
-                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
-            }
-
-            #pragma unroll(2)
-            for (short i = 0; i < 2; i++) {
-                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
             }
 
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(8)
-            for (short i = 0; i < 8; i++){
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
 
-            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
-            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsma += 8*64;
+            lsmb += 4*64;
         }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
     }
 
-    if (!FC_mul_mm_bc_out || ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1)) {
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
         // if no bounds checks on the output are needed, we can directly write to device memory
+#ifdef GGML_METAL_HAS_TENSOR
         device float * C = (device float *) dst +
-            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
-            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+            r0 + \
+            r1 * args.ne0 + im*args.ne1*args.ne0;
+
+        auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(args.ne0, NR1));
+        cT.store(tC);
+#else
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
 
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
         }
+#endif
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+#ifdef GGML_METAL_HAS_TENSOR
+        auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+        cT.store(tC);
+#else
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
         }
+#endif
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         if (sgitg == 0) {
-            for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.ne0 + im*args.ne1*args.ne0;
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
                 device float4 * D4 = (device float4 *) D;
 
-                threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
+                threadgroup float  * C  = temp_str + (j*NR0);
                 threadgroup float4 * C4 = (threadgroup float4 *) C;
 
                 int i = 0;
-                for (; i < n_rows/4; i++) {
+                for (; i < nr0/4; i++) {
                     *(D4 + i) = *(C4 + i);
                 }
 
                 i *= 4;
-                for (; i < n_rows; i++) {
+                for (; i < nr0; i++) {
                     *(D + i) = *(C + i);
                 }
             }
@@ -8432,31 +8993,63 @@ kernel void kernel_mul_mm_id(
         ushort tiitg[[thread_index_in_threadgroup]],
         ushort tiisg[[thread_index_in_simdgroup]],
         ushort sgitg[[simdgroup_index_in_threadgroup]]) {
-
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
-    const int r0 = tgpig.y;
-    const int r1 = tgpig.x;
+    threadgroup float * sc = (threadgroup float *)(shmem);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
     const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
 
     device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
     device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
 
     const int32_t neh1 = tpe_u32[im];
 
-    if (r1*BLOCK_SIZE_N >= neh1) {
+    if (r1 >= neh1) {
         return;
     }
 
     // if this block is of 64x32 shape or smaller
-    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    const short n_cols = (    neh1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (    neh1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
 
     // a thread shouldn't load data outside of the matrix
-    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
 
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
     S0_8x8 ma[4];
     S1_8x8 mb[2];
 
@@ -8465,39 +9058,36 @@ kernel void kernel_mul_mm_id(
     for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
 
-    short il = (tiitg % THREAD_PER_ROW);
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
 
-    const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + thread_col];
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
 
-    const short i11 = (id % args.ne20) % args.ne11;
-    const short i12 = (id / args.ne20);
-    const short i13 = 0;
-
-    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
-    const short    offset1 = il/nl;
-
-    device const block_q * x = (device const block_q *)(src0
-        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
-
-    const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
-
-    device const T1 * y = (device const T1 *)(src1
-        + args.nb13*i13
-        + args.nb12*i12
-        + args.nb11*i11
-        + args.nb10*iy);
-
-    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
         // load data and store to threadgroup memory
         if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             // no need for dequantization
             for (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
             }
         } else {
             S0_4x4 temp_a;
@@ -8506,85 +9096,188 @@ kernel void kernel_mul_mm_id(
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             FOR_UNROLL (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
             }
         }
 
         if (FC_mul_mm_bc_inp) {
             for (short i = 0; i < 8; ++i) {
-                sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
             }
         } else {
-            *(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            const short dx = sx;
+            const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
         }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
-        y += BLOCK_SIZE_K;
+
+        y += NK;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
+#ifndef GGML_METAL_HAS_TENSOR
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
 
-        #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
-            #pragma unroll(4)
-            for (short i = 0; i < 4; i++) {
-                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
             }
 
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(2)
-            for (short i = 0; i < 2; i++) {
-                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
             }
 
-            #pragma unroll(8)
-            for (short i = 0; i < 8; i++){
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
 
-            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
-            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsma += 8*64;
+            lsmb += 4*64;
         }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
     }
 
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                 + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
 
-    #pragma unroll(8)
     for (short i = 0; i < 8; i++) {
-        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
     }
+#endif
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (short j = sgitg; j < n_cols; j += 4) {
-        const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + j];
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
 
         const short ide = id % args.ne20;
         const short idt = id / args.ne20;
 
-        device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
         device float4 * D4 = (device float4 *) D;
 
-        threadgroup float  * C  = (threadgroup float  *) shmem + (j*BLOCK_SIZE_M);
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
         threadgroup float4 * C4 = (threadgroup float4 *) C;
 
         int i = tiisg;
-        for (; i < n_rows/4; i += 32) {
+        for (; i < nr0/4; i += 32) {
             *(D4 + i) = *(C4 + i);
         }
 
-        i = (4*(n_rows/4)) + tiisg;
-        for (; i < n_rows; i += 32) {
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
             *(D + i) = *(C + i);
         }
     }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ml/backend/ggml/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 903050b0..c98f98c7 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -32,6 +32,7 @@ DispatchLoaderDynamic & ggml_vk_default_dispatcher();
 #include <memory>
 #include <limits>
 #include <map>
+#include <set>
 #include <unordered_map>
 #include <memory>
 #include <mutex>
@@ -130,10 +131,10 @@ struct vk_pipeline_struct {
     uint32_t align;
     // true if fields have been set by ggml_vk_create_pipeline
     bool initialized {};
-    // set to true to request the pipeline is compiled after the dryrun
-    bool needed {};
+    // set to true to request the pipeline is compiled
+    std::atomic<bool> needed {};
     // set to true when the shader has been compiled
-    bool compiled {};
+    std::atomic<bool> compiled {};
     // number of registers used, extracted from pipeline executable properties
     uint32_t register_count {};
 };
@@ -235,6 +236,7 @@ class vk_memory_logger;
 #endif
 class vk_perf_logger;
 static void ggml_vk_destroy_buffer(vk_buffer& buf);
+static void ggml_vk_synchronize(ggml_backend_vk_context * ctx);
 static std::string ggml_vk_get_device_id(int device);
 
 static constexpr uint32_t mul_mat_vec_max_cols = 8;
@@ -353,6 +355,12 @@ enum vk_conv_shapes {
     CONV_SHAPE_COUNT,
 };
 
+uint32_t conv_shapes_wg_denoms[][3] = {
+    { 128, 128, 1 },
+    {  64,  32, 1 },
+    {  32, 256, 1 },
+};
+
 enum dmmv_wg_sizes {
     DMMV_WG_SIZE_SUBGROUP,
     DMMV_WG_SIZE_LARGE,
@@ -381,6 +389,30 @@ struct vk_fa_pipeline_state {
     }
 };
 
+struct vk_conv2d_pipeline_state {
+    vk_conv2d_pipeline_state(uint32_t s0, uint32_t s1, uint32_t p0, uint32_t p1, uint32_t d0, uint32_t d1, uint32_t KW, uint32_t KH)
+        : s0(s0), s1(s1), p0(p0), p1(p1), d0(d0), d1(d1), KW(KW), KH(KH) {}
+
+    uint32_t s0, s1, p0, p1, d0, d1, KW, KH;
+
+    bool operator<(const vk_conv2d_pipeline_state &b) const {
+        return std::tie(s0, s1, p0, p1, d0, d1, KW, KH) <
+               std::tie(b.s0, b.s1, b.p0, b.p1, b.d0, b.d1, b.KW, b.KH);
+    }
+};
+
+struct vk_solve_tri_pipeline_state {
+    vk_solve_tri_pipeline_state(uint32_t N, uint32_t K)
+        : N(N), K(K) {}
+
+    uint32_t N, K;
+
+    bool operator<(const vk_solve_tri_pipeline_state &b) const {
+        return std::tie(N, K) <
+               std::tie(b.N, b.K);
+    }
+};
+
 enum shader_reduction_mode {
     SHADER_REDUCTION_MODE_SHMEM,
     SHADER_REDUCTION_MODE_HYBRID,
@@ -388,9 +420,10 @@ enum shader_reduction_mode {
     SHADER_REDUCTION_MODE_COUNT,
 };
 
+// argsort pipelines for up to 1<<10 invocations per workgroup
 static constexpr uint32_t num_argsort_pipelines = 11;
-static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
 static constexpr uint32_t num_topk_moe_pipelines = 10;
+static constexpr uint32_t num_topk_pipelines = 11;
 
 static constexpr std::initializer_list<ggml_op> topk_moe_early_softmax_norm{ GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
                                                                              GGML_OP_VIEW,     GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
@@ -468,6 +501,14 @@ static constexpr std::initializer_list<std::array<int, 3>> rope_view_set_rows_ed
     { 2, 0, 1 }, // set_rows->src[0] == view
 };
 
+static constexpr std::initializer_list<std::array<int, 3>> rms_norm_mul_rope_view_set_rows_edges {
+    { 1, 0, 0 }, // mul->src[0]      == rms
+    { 2, 0, 1 }, // rope->src[0]     == mul
+    { 3, 0, 2 }, // view->src[0]     == rope
+    { 4, 0, 3 }, // set_rows->src[0] == view
+};
+
+
 struct vk_device_struct {
     std::recursive_mutex mutex;
 
@@ -487,7 +528,9 @@ struct vk_device_struct {
     vk_queue compute_queue;
     vk_queue transfer_queue;
     bool single_queue;
+    bool support_async;
     uint32_t subgroup_size;
+    uint32_t subgroup_size_log2;
     uint32_t shader_core_count;
     bool uma;
     bool prefer_host_memory;
@@ -496,9 +539,11 @@ struct vk_device_struct {
     bool subgroup_shuffle;
     bool subgroup_ballot;
     bool subgroup_clustered;
+    bool subgroup_vote;
     bool multi_add;
     bool shader_int64;
     bool buffer_device_address;
+    bool vulkan_memory_model;
 
     bool add_rms_fusion;
     uint32_t partials_binding_alignment;
@@ -512,6 +557,9 @@ struct vk_device_struct {
     uint32_t subgroup_max_size;
     bool subgroup_require_full_support;
 
+    // floor(log2(maxComputeWorkGroupInvocations))
+    uint32_t max_workgroup_size_log2 {};
+
     bool coopmat_support;
     bool coopmat_acc_f32_support {};
     bool coopmat_acc_f16_support {};
@@ -541,9 +589,6 @@ struct vk_device_struct {
     bool mul_mat_id_m[GGML_TYPE_COUNT];
     bool mul_mat_id_s[GGML_TYPE_COUNT];
 
-    // set to true to indicate that some shaders need to be compiled after the dryrun
-    bool need_compiles {};
-
     vk::DescriptorSetLayout dsl;
 
     vk_matmul_pipeline pipeline_matmul_f32 {};
@@ -565,15 +610,15 @@ struct vk_device_struct {
     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_matmul_split_k_reduce;
-    vk_pipeline pipeline_quantize_q8_1;
     vk_pipeline pipeline_quantize_q8_1_x4;
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
-    vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_dequant_mul_mat_vec_q8_1_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_id_q8_1_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
     vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
@@ -600,12 +645,14 @@ struct vk_device_struct {
     vk_pipeline pipeline_add_id_f32;
 
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
-    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bilinear_ac_f32;
+    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
     vk_pipeline pipeline_sqrt_f32;
     vk_pipeline pipeline_sin_f32;
     vk_pipeline pipeline_cos_f32;
+    vk_pipeline pipeline_log[2];
+    vk_pipeline pipeline_tri[2];
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_roll_f32;
@@ -614,6 +661,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16, pipeline_contig_cpy_f32_i32, pipeline_contig_cpy_i32_f32;
     vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_cpy_transpose_16, pipeline_cpy_transpose_32;
     vk_pipeline pipeline_set_rows_i32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_set_rows_i64[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
@@ -622,6 +670,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_rms_norm_mul_f32;
     vk_pipeline pipeline_rms_norm_partials_f32;
     vk_pipeline pipeline_rms_norm_mul_partials_f32;
+    vk_pipeline pipeline_rms_norm_mul_rope_f32_f32;
+    vk_pipeline pipeline_rms_norm_mul_rope_f32_f16;
     vk_pipeline pipeline_rms_norm_back_f32;
     vk_pipeline pipeline_l2_norm_f32;
 
@@ -632,10 +682,26 @@ struct vk_device_struct {
     vk_pipeline pipeline_gelu_quick[2];
     vk_pipeline pipeline_silu[2];
     vk_pipeline pipeline_relu[2];
+    vk_pipeline pipeline_neg[2];
     vk_pipeline pipeline_tanh[2];
     vk_pipeline pipeline_sigmoid[2];
     vk_pipeline pipeline_hardsigmoid[2];
     vk_pipeline pipeline_hardswish[2];
+    vk_pipeline pipeline_abs[2];
+    vk_pipeline pipeline_softplus[2];
+    vk_pipeline pipeline_step[2];
+    vk_pipeline pipeline_round[2];
+    vk_pipeline pipeline_ceil[2];
+    vk_pipeline pipeline_floor[2];
+    vk_pipeline pipeline_trunc[2];
+
+    vk_pipeline pipeline_add1_f16_f16;
+    vk_pipeline pipeline_add1_f16_f32;
+    vk_pipeline pipeline_add1_f32_f32;
+
+    vk_pipeline pipeline_arange_f32;
+
+    vk_pipeline pipeline_fill_f32;
 
     vk_pipeline pipeline_geglu[2];
     vk_pipeline pipeline_reglu[2];
@@ -655,9 +721,13 @@ struct vk_device_struct {
     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
     vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
     vk_pipeline pipeline_argsort_f32[num_argsort_pipelines];
+    vk_pipeline pipeline_argsort_large_f32[num_argsort_pipelines];
+    vk_pipeline pipeline_topk_f32[num_topk_pipelines];
     vk_pipeline pipeline_sum_rows_f32;
+    vk_pipeline pipeline_cumsum_f32;
     vk_pipeline pipeline_argmax_f32;
     vk_pipeline pipeline_count_equal_i32;
+    std::map<vk_solve_tri_pipeline_state, vk_pipeline> pipeline_solve_tri_f32;
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
     vk_pipeline pipeline_im2col_3d_f32, pipeline_im2col_3d_f32_f16;
     vk_pipeline pipeline_timestep_embedding_f32;
@@ -670,10 +740,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_ssm_conv_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
     vk_pipeline pipeline_opt_step_sgd_f32;
-    vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
-    vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
-    vk_pipeline pipeline_conv_transpose_2d_f32[CONV_SHAPE_COUNT];
-    vk_pipeline pipeline_conv_transpose_2d_f16_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv2d_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f16_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_dw_whcn_f32, pipeline_conv2d_dw_whcn_f16_f32;
     vk_pipeline pipeline_conv2d_dw_cwhn_f32, pipeline_conv2d_dw_cwhn_f16_f32;
 
@@ -798,10 +868,51 @@ struct vk_mat_mat_push_constants {
     uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
     uint32_t padded_N;
 };
+
+#define MAT_VEC_FUSION_FLAGS_BIAS0 0x1
+#define MAT_VEC_FUSION_FLAGS_BIAS1 0x2
+#define MAT_VEC_FUSION_FLAGS_SCALE0 0x4
+#define MAT_VEC_FUSION_FLAGS_SCALE1 0x8
+
 struct vk_mat_vec_push_constants {
-    uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
-    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
-    uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
+    uint32_t ncols;
+    uint32_t stride_a;
+    uint32_t stride_b;
+    uint32_t stride_d;
+    uint32_t batch_stride_a;
+    uint32_t batch_stride_b;
+    uint32_t batch_stride_d;
+    uint32_t fusion_flags;
+    uint32_t ne02;
+    uint32_t ne12;
+    uint32_t broadcast2;
+    uint32_t broadcast3;
+};
+
+struct vk_mat_vec_p021_push_constants {
+    uint32_t ncols_x;
+    uint32_t nrows_x;
+    uint32_t nchannels_x;
+    uint32_t nchannels_y;
+    uint32_t b_offset;
+    uint32_t d_offset;
+    uint32_t fusion_flags;
+};
+
+struct vk_mat_vec_nc_push_constants {
+    uint32_t ncols_x;
+    uint32_t nrows_x;
+    uint32_t row_stride_x;
+    uint32_t channel_stride_x;
+    uint32_t channel_stride_y;
+    uint32_t channel_x_divisor;
+    uint32_t ne12;
+    uint32_t b_offset;
+    uint32_t d_offset;
+    uint32_t nb03;
+    uint32_t nb13;
+    uint32_t nb23;
+    uint32_t fusion_flags;
 };
 
 struct vk_mat_mat_id_push_constants {
@@ -812,9 +923,16 @@ struct vk_mat_mat_id_push_constants {
     uint32_t padded_N;
 };
 struct vk_mat_vec_id_push_constants {
-    uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
-    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
-    uint32_t nei0; uint32_t ne11;
+    uint32_t ncols;
+    uint32_t stride_a;
+    uint32_t stride_b;
+    uint32_t stride_d;
+    uint32_t batch_stride_a;
+    uint32_t batch_stride_b;
+    uint32_t batch_stride_d;
+    uint32_t fusion_flags;
+    uint32_t nei0;
+    uint32_t ne11;
 };
 
 struct vk_flash_attn_push_constants {
@@ -1049,6 +1167,7 @@ struct vk_op_diag_mask_push_constants {
 };
 
 struct vk_op_rope_push_constants {
+    uint32_t rope_mode;
     uint32_t ncols;
     uint32_t n_dims;
     float freq_scale;
@@ -1063,10 +1182,17 @@ struct vk_op_rope_push_constants {
     uint32_t s1;
     uint32_t s2;
     int32_t sections[4];
+    uint32_t is_imrope;
     uint32_t is_back;
     uint32_t set_rows_stride;
 };
 
+// For fused rms_norm+mul+rope(+view+set_rows)
+struct vk_op_rms_norm_mul_rope_push_constants {
+    vk_op_binary_push_constants bin;
+    vk_op_rope_push_constants rope;
+};
+
 struct vk_op_soft_max_push_constants {
     uint32_t KX;
     uint32_t KY;
@@ -1089,8 +1215,23 @@ struct vk_op_soft_max_push_constants {
 
 struct vk_op_argsort_push_constants {
     uint32_t ncols;
+    uint32_t ncols_padded;
+    uint32_t ncols_padded_log2;
     uint32_t nrows;
-    int32_t order;
+    uint32_t order;
+    uint32_t outer_start;
+    uint32_t outer_end;
+    uint32_t inner_start;
+    uint32_t inner_end;
+};
+
+struct vk_op_topk_push_constants {
+    uint32_t orig_ncols;
+    uint32_t ncols_input;
+    uint32_t ncols_output;
+    uint32_t nrows;
+    uint32_t first_pass;
+    uint32_t last_pass;
 };
 
 struct vk_op_im2col_push_constants {
@@ -1229,17 +1370,13 @@ struct vk_op_conv2d_push_constants {
     uint32_t nb2;
     uint32_t nb3;
 
-    // init_fastdiv_values constants for dividing by KW, KW*KH, OW, OW*OH
-    uint32_t KWmp;   uint32_t KWL;
-    uint32_t KWKHmp; uint32_t KWKHL;
+    // init_fastdiv_values constants for dividing by OW, OW*OH
     uint32_t OWmp;   uint32_t OWL;
     uint32_t OWOHmp; uint32_t OWOHL;
 };
 
 template <> void init_pushconst_fastdiv(vk_op_conv2d_push_constants &p) {
-    // Compute magic values to divide by KW, KW*KH, OW, OW*OH
-    init_fastdiv_values(p.KW,       p.KWmp,    p.KWL);
-    init_fastdiv_values(p.KW*p.KH,  p.KWKHmp,  p.KWKHL);
+    // Compute magic values to divide by OW, OW*OH
     init_fastdiv_values(p.OW,       p.OWmp,    p.OWL);
     init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
 }
@@ -1275,23 +1412,15 @@ struct vk_op_conv_transpose_2d_push_constants {
     uint32_t nb2;
     uint32_t nb3;
 
-    // init_fastdiv_values constants for dividing by KW, KW*KH, OW, OW*OH, s0, s1
-    uint32_t KWmp;   uint32_t KWL;
-    uint32_t KWKHmp; uint32_t KWKHL;
+    // init_fastdiv_values constants for dividing by OW, OW*OH
     uint32_t OWmp;   uint32_t OWL;
     uint32_t OWOHmp; uint32_t OWOHL;
-    uint32_t s0mp; uint32_t s0L;
-    uint32_t s1mp; uint32_t s1L;
 };
 
 template <> void init_pushconst_fastdiv(vk_op_conv_transpose_2d_push_constants &p) {
-    // Compute magic values to divide by KW, KW*KH, OW, OW*OH, s0, s1
-    init_fastdiv_values(p.KW,       p.KWmp,    p.KWL);
-    init_fastdiv_values(p.KW*p.KH,  p.KWKHmp,  p.KWKHL);
+    // Compute magic values to divide by OW, OW*OH
     init_fastdiv_values(p.OW,       p.OWmp,    p.OWL);
     init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
-    init_fastdiv_values(p.s0,       p.s0mp,    p.s0L);
-    init_fastdiv_values(p.s1,       p.s1mp,    p.s1L);
 }
 
 struct vk_op_conv2d_dw_push_constants {
@@ -1318,6 +1447,7 @@ struct vk_op_upscale_push_constants {
     uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
     float sf0; float sf1; float sf2; float sf3;
+    float pixel_offset;
 };
 
 struct vk_op_sum_rows_push_constants
@@ -1392,6 +1522,10 @@ struct ggml_vk_garbage_collector {
     std::vector<vk_context> contexts;
 };
 
+static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_context subctx);
+static void ggml_vk_load_shaders(vk_device& device);
+static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);
+
 #if defined(GGML_VULKAN_MEMORY_DEBUG) || defined(GGML_VULKAN_DEBUG)
 #define VK_LOG_MEMORY(msg) std::cerr << "ggml_vulkan memory: " << msg << std::endl
 
@@ -1480,7 +1614,7 @@ class vk_perf_logger {
         }
         if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
             const uint64_t m     = node->src[0]->ne[1];
-            const uint64_t n     = node->ne[1];
+            const uint64_t n     = (node->op == GGML_OP_MUL_MAT) ? node->ne[1] : node->ne[2];
             const uint64_t k     = node->src[1]->ne[0];
             const uint64_t batch = node->src[1]->ne[2] * node->src[1]->ne[3];
             std::string    name  = ggml_op_name(node->op);
@@ -1525,6 +1659,22 @@ class vk_perf_logger {
             timings[name].push_back(time);
             return;
         }
+        if (node->op == GGML_OP_FLASH_ATTN_EXT) {
+            const ggml_tensor * dst = node;
+            const ggml_tensor * q = node->src[0];
+            const ggml_tensor * k = node->src[1];
+            const ggml_tensor * v = node->src[2];
+            const ggml_tensor * m = node->src[3];
+            std::stringstream name;
+            name << ggml_op_name(node->op) <<
+                " dst(" << dst->ne[0] << "," << dst->ne[1] << "," << dst->ne[2] << "," << dst->ne[3] << "), " <<
+                " q(" << q->ne[0] << "," << q->ne[1] << "," << q->ne[2] << "," << q->ne[3] << "), " <<
+                " k(" << k->ne[0] << "," << k->ne[1] << "," << k->ne[2] << "," << k->ne[3] << "), " <<
+                " v(" << v->ne[0] << "," << v->ne[1] << "," << v->ne[2] << "," << v->ne[3] << "), " <<
+                " m(" << (m?m->ne[0]:0) << "," << (m?m->ne[1]:0) << "," << (m?m->ne[2]:0) << "," << (m?m->ne[3]:0) << ")";
+            timings[name.str()].push_back(time);
+            return;
+        }
         timings[ggml_op_name(node->op)].push_back(time);
     }
   private:
@@ -1540,13 +1690,17 @@ struct ggml_backend_vk_context {
     size_t semaphore_idx, event_idx;
     ggml_vk_garbage_collector gc;
     size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k, prealloc_size_add_rms_partials, prealloc_size_add_rms_partials_offset;
-    vk_buffer prealloc_x, prealloc_y, prealloc_split_k, prealloc_add_rms_partials;
+    vk_buffer prealloc_x, prealloc_y, prealloc_split_k, prealloc_add_rms_partials, sync_staging;
     vk::Fence fence, almost_ready_fence;
+    bool submit_pending {};
     bool almost_ready_fence_pending {};
     // Set before op_add and unset after op_rms_norm to indicate that the add should
     // write partial sums to accumulate the square of the vector components
+    bool do_add_rms_partials_offset_calculation;
     bool do_add_rms_partials;
 
+    uint64_t last_total_mul_mat_bytes {};
+
     // Cache most recent tensor that was converted into prealloc_y, and what pipeline it used to convert.
     vk_pipeline_struct * prealloc_y_last_pipeline_used {};
     const ggml_tensor * prealloc_y_last_tensor_used {};
@@ -1590,6 +1744,50 @@ static uint64_t vk_tensor_offset(const ggml_tensor * tensor) {
     return (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
 }
 
+static uint32_t get_misalign_bytes(const ggml_backend_vk_context * ctx, const ggml_tensor * t)
+{
+    return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
+}
+
+template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    GGML_UNUSED(p);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
+    GGML_UNUSED(dst);
+    static_assert(!std::is_const<T>::value, "unexpected type");
+    GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0);
+    GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0);
+    GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0);
+    GGML_ASSERT(!src3 || get_misalign_bytes(ctx, src3) == 0);
+    GGML_ASSERT(!dst  || get_misalign_bytes(ctx, dst) == 0);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_mat_vec_p021_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.b_offset = b_offset;
+    p.d_offset = d_offset;
+
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_mat_vec_nc_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.b_offset = b_offset;
+    p.d_offset = d_offset;
+
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
+}
+
 struct ggml_backend_vk_buffer_context {
     vk_device_ref device;
     vk_buffer dev_buffer;
@@ -1822,10 +2020,7 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         }
     }
 
-    {
-        std::lock_guard<std::recursive_mutex> guard(device->mutex);
-        device->all_pipelines.push_back(pipeline);
-    }
+    device->all_pipelines.push_back(pipeline);
 
     {
         std::lock_guard<std::mutex> guard(compile_count_mutex);
@@ -1849,8 +2044,9 @@ static void ggml_pipeline_request_descriptor_sets(ggml_backend_vk_context *ctx,
     ctx->pipeline_descriptor_set_requirements += n;
     if (!pipeline->compiled) {
         pipeline->needed = true;
-        ctx->device->need_compiles = true;
+        ggml_vk_load_shaders(ctx->device);
     }
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 }
 
 static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx) {
@@ -1862,7 +2058,9 @@ static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx
 
     vk_device& device = ctx->device;
 
-    uint32_t to_alloc = ctx->pipeline_descriptor_set_requirements - ctx->descriptor_sets.size();
+    // Grow by 50% to avoid frequent allocations
+    uint32_t needed = std::max(3 * ctx->descriptor_sets.size() / 2, size_t{ctx->pipeline_descriptor_set_requirements});
+    uint32_t to_alloc = needed - ctx->descriptor_sets.size();
     uint32_t pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE - ctx->descriptor_sets.size() % VK_DEVICE_DESCRIPTOR_POOL_SIZE;
     uint32_t pool_idx = ctx->descriptor_sets.size() / VK_DEVICE_DESCRIPTOR_POOL_SIZE;
 
@@ -2115,17 +2313,18 @@ static void ggml_vk_queue_command_pools_cleanup(vk_device& device) {
     }
 }
 
+static std::vector<uint32_t> ggml_vk_find_memory_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
+    std::vector<uint32_t> indices;
 
-static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
     for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
         vk::MemoryType memory_type = mem_props->memoryTypes[i];
         if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
             (flags & memory_type.propertyFlags) == flags &&
             mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
-            return static_cast<int32_t>(i);
+            indices.push_back(i);
         }
     }
-    return UINT32_MAX;
+    return indices;
 }
 
 static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std::initializer_list<vk::MemoryPropertyFlags> & req_flags_list) {
@@ -2168,24 +2367,33 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std
     for (auto it = req_flags_list.begin(); it != req_flags_list.end(); it++) {
         const auto & req_flags = *it;
 
-        uint32_t memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
+        const std::vector<uint32_t> memory_type_indices = ggml_vk_find_memory_properties(&mem_props, &mem_req, req_flags);
 
-        if (memory_type_index == UINT32_MAX) {
+        if (memory_type_indices.empty()) {
             continue;
         }
         buf->memory_property_flags = req_flags;
 
-        try {
-            buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index, &mem_flags_info });
-            break;
-        } catch (const vk::SystemError& e) {
-            // loop and retry
-            // during last attempt throw the exception
-            if (it + 1 == req_flags_list.end()) {
-                device->device.destroyBuffer(buf->buffer);
-                throw e;
+        bool done = false;
+
+        for (auto mtype_it = memory_type_indices.begin(); mtype_it != memory_type_indices.end(); mtype_it++) {
+            try {
+                buf->device_memory = device->device.allocateMemory({ mem_req.size, *mtype_it, &mem_flags_info });
+                done = true;
+                break;
+            } catch (const vk::SystemError& e) {
+                // loop and retry
+                // during last attempt throw the exception
+                if (it + 1 == req_flags_list.end() && mtype_it + 1 == memory_type_indices.end()) {
+                    device->device.destroyBuffer(buf->buffer);
+                    throw e;
+                }
             }
         }
+
+        if (done) {
+            break;
+        }
     }
 
     if (!buf->device_memory) {
@@ -2323,9 +2531,11 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
 
-static uint32_t get_fa_scalar_num_large_rows(uint32_t hsv) {
+static uint32_t get_fa_scalar_num_large_rows(uint32_t hsk, uint32_t hsv) {
     if (hsv >= 192) {
         return 2;
+    } else if ((hsv | hsk) & 8) {
+        return 4;
     } else {
         return 8;
     }
@@ -2357,9 +2567,9 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
             if ((hsv | hsk) & 8) {
                 // HSV/HSK not being a multiple of 16 makes D_split smaller, which makes cols_per_iter
                 // larger, and Bc needs to be >= cols_per_thread. 64 is large enough, 32 is not.
-                return {get_fa_scalar_num_large_rows(hsv), 64};
+                return {get_fa_scalar_num_large_rows(hsk, hsv), 64};
             } else {
-                return {get_fa_scalar_num_large_rows(hsv), 32};
+                return {get_fa_scalar_num_large_rows(hsk, hsv), 32};
             }
         }
     }
@@ -2513,6 +2723,7 @@ static uint32_t get_subgroup_size(const std::string &pipeline_name, const vk_dev
 static void ggml_vk_load_shaders(vk_device& device) {
     VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
 
+    std::lock_guard<std::recursive_mutex> guard(device->mutex);
     // some shaders have a minimum subgroup size
     const uint32_t subgroup_size_8 = std::max(device->subgroup_size, 8u);
     const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);
@@ -2706,6 +2917,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         if (!pipeline->needed || pipeline->compiled) {
             return;
         }
+        // TODO: We're no longer benefitting from the async compiles (shaders are
+        // compiled individually, as needed) and this complexity can be removed.
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -2763,15 +2976,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
             if (path == FAPATH) { \
                 if (aligned) { \
                     if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } \
                 } else { \
                     if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1,                                        true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1,                                        true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } \
                 } \
             } \
@@ -3315,13 +3528,18 @@ static void ggml_vk_load_shaders(vk_device& device) {
     // the number of rows computed per shader depends on GPU model and quant
     uint32_t rm_stdq = 1;
     uint32_t rm_kq = 2;
+    uint32_t rm_stdq_int = 1;
+    uint32_t rm_kq_int = 1;
     if (device->vendor_id == VK_VENDOR_ID_AMD) {
         if (device->architecture == AMD_GCN) {
             rm_stdq = 2;
             rm_kq = 4;
+            rm_stdq_int = 4;
         }
-    } else if (device->vendor_id == VK_VENDOR_ID_INTEL)
+    } else if (device->vendor_id == VK_VENDOR_ID_INTEL) {
         rm_stdq = 2;
+        rm_stdq_int = 2;
+    }
     uint32_t rm_iq = 2 * rm_kq;
 
     const bool use_subgroups = device->subgroup_arithmetic && device->architecture != vk_device_architecture::AMD_GCN;
@@ -3333,6 +3551,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     const uint32_t force_subgroup_size = use_subgroups ? subgroup_size : 0;
     const uint32_t force_subgroup_size16 = use_subgroups16 ? subgroup_size16 : 0;
+    static constexpr uint32_t mul_mat_vec_num_bindings = 5;
+    static constexpr uint32_t mul_mat_vec_id_num_bindings = 6;
 
     for (uint32_t w = 0; w < DMMV_WG_SIZE_COUNT; ++w) {
         const uint32_t wg_size_subgroup   = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size : (subgroup_size * 4);
@@ -3347,92 +3567,126 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                               SHADER_REDUCTION_MODE_SHMEM;
 
         for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32",  arr_dmmv_f32_f32_f32_len[reduc],  arr_dmmv_f32_f32_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32",  arr_dmmv_f16_f32_f32_len[reduc],  arr_dmmv_f16_f32_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32", arr_dmmv_bf16_f32_f32_len[reduc], arr_dmmv_bf16_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32", arr_dmmv_q4_0_f32_f32_len[reduc], arr_dmmv_q4_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32", arr_dmmv_q4_1_f32_f32_len[reduc], arr_dmmv_q4_1_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32", arr_dmmv_q5_0_f32_f32_len[reduc], arr_dmmv_q5_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[reduc], arr_dmmv_q5_1_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[reduc], arr_dmmv_q8_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[reduc16], arr_dmmv_q2_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[reduc16], arr_dmmv_q3_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[reduc16], arr_dmmv_q4_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[reduc16], arr_dmmv_q5_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32", arr_dmmv_q6_k_f32_f32_len[reduc16], arr_dmmv_q6_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32",   arr_dmmv_iq1_s_f32_f32_len[reduc16],   arr_dmmv_iq1_s_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32",   arr_dmmv_iq1_m_f32_f32_len[reduc16],   arr_dmmv_iq1_m_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32", arr_dmmv_iq2_xxs_f32_f32_len[reduc16], arr_dmmv_iq2_xxs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32",  arr_dmmv_iq2_xs_f32_f32_len[reduc16],  arr_dmmv_iq2_xs_f32_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32",   arr_dmmv_iq2_s_f32_f32_len[reduc16],   arr_dmmv_iq2_s_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32", arr_dmmv_iq3_xxs_f32_f32_len[reduc16], arr_dmmv_iq3_xxs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32",   arr_dmmv_iq3_s_f32_f32_len[reduc16],   arr_dmmv_iq3_s_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32",  arr_dmmv_iq4_xs_f32_f32_len[reduc16],  arr_dmmv_iq4_xs_f32_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32",  arr_dmmv_iq4_nl_f32_f32_len[reduc16],  arr_dmmv_iq4_nl_f32_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32",   arr_dmmv_mxfp4_f32_f32_len[reduc16],   arr_dmmv_mxfp4_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32",  arr_dmmv_f32_f32_f32_len[reduc],  arr_dmmv_f32_f32_f32_data[reduc],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32",  arr_dmmv_f16_f32_f32_len[reduc],  arr_dmmv_f16_f32_f32_data[reduc],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32", arr_dmmv_bf16_f32_f32_len[reduc], arr_dmmv_bf16_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32", arr_dmmv_q4_0_f32_f32_len[reduc], arr_dmmv_q4_0_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32", arr_dmmv_q4_1_f32_f32_len[reduc], arr_dmmv_q4_1_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32", arr_dmmv_q5_0_f32_f32_len[reduc], arr_dmmv_q5_0_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[reduc], arr_dmmv_q5_1_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[reduc], arr_dmmv_q8_0_f32_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[reduc16], arr_dmmv_q2_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[reduc16], arr_dmmv_q3_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[reduc16], arr_dmmv_q4_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[reduc16], arr_dmmv_q5_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32", arr_dmmv_q6_k_f32_f32_len[reduc16], arr_dmmv_q6_k_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32",   arr_dmmv_iq1_s_f32_f32_len[reduc16],   arr_dmmv_iq1_s_f32_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32",   arr_dmmv_iq1_m_f32_f32_len[reduc16],   arr_dmmv_iq1_m_f32_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32", arr_dmmv_iq2_xxs_f32_f32_len[reduc16], arr_dmmv_iq2_xxs_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32",  arr_dmmv_iq2_xs_f32_f32_len[reduc16],  arr_dmmv_iq2_xs_f32_f32_data[reduc16],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32",   arr_dmmv_iq2_s_f32_f32_len[reduc16],   arr_dmmv_iq2_s_f32_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32", arr_dmmv_iq3_xxs_f32_f32_len[reduc16], arr_dmmv_iq3_xxs_f32_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32",   arr_dmmv_iq3_s_f32_f32_len[reduc16],   arr_dmmv_iq3_s_f32_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32",  arr_dmmv_iq4_xs_f32_f32_len[reduc16],  arr_dmmv_iq4_xs_f32_f32_data[reduc16],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32",  arr_dmmv_iq4_nl_f32_f32_len[reduc16],  arr_dmmv_iq4_nl_f32_f32_data[reduc16],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32",   arr_dmmv_mxfp4_f32_f32_len[reduc16],   arr_dmmv_mxfp4_f32_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
 
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32",  arr_dmmv_f32_f16_f32_len[reduc],  arr_dmmv_f32_f16_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32",  arr_dmmv_f16_f16_f32_len[reduc],  arr_dmmv_f16_f16_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32", arr_dmmv_bf16_f16_f32_len[reduc], arr_dmmv_bf16_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32", arr_dmmv_q4_0_f16_f32_len[reduc], arr_dmmv_q4_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32", arr_dmmv_q4_1_f16_f32_len[reduc], arr_dmmv_q4_1_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32", arr_dmmv_q5_0_f16_f32_len[reduc], arr_dmmv_q5_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[reduc], arr_dmmv_q5_1_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[reduc], arr_dmmv_q8_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[reduc16], arr_dmmv_q2_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[reduc16], arr_dmmv_q3_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[reduc16], arr_dmmv_q4_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[reduc16], arr_dmmv_q5_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32", arr_dmmv_q6_k_f16_f32_len[reduc16], arr_dmmv_q6_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32",   arr_dmmv_iq1_s_f16_f32_len[reduc16],   arr_dmmv_iq1_s_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32",   arr_dmmv_iq1_m_f16_f32_len[reduc16],   arr_dmmv_iq1_m_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32", arr_dmmv_iq2_xxs_f16_f32_len[reduc16], arr_dmmv_iq2_xxs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32",  arr_dmmv_iq2_xs_f16_f32_len[reduc16],  arr_dmmv_iq2_xs_f16_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32",   arr_dmmv_iq2_s_f16_f32_len[reduc16],   arr_dmmv_iq2_s_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32", arr_dmmv_iq3_xxs_f16_f32_len[reduc16], arr_dmmv_iq3_xxs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32",   arr_dmmv_iq3_s_f16_f32_len[reduc16],   arr_dmmv_iq3_s_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32",  arr_dmmv_iq4_xs_f16_f32_len[reduc16],  arr_dmmv_iq4_xs_f16_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32",  arr_dmmv_iq4_nl_f16_f32_len[reduc16],  arr_dmmv_iq4_nl_f16_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32",   arr_dmmv_mxfp4_f16_f32_len[reduc16],   arr_dmmv_mxfp4_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32",  arr_dmmv_f32_f16_f32_len[reduc],  arr_dmmv_f32_f16_f32_data[reduc],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32",  arr_dmmv_f16_f16_f32_len[reduc],  arr_dmmv_f16_f16_f32_data[reduc],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32", arr_dmmv_bf16_f16_f32_len[reduc], arr_dmmv_bf16_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32", arr_dmmv_q4_0_f16_f32_len[reduc], arr_dmmv_q4_0_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32", arr_dmmv_q4_1_f16_f32_len[reduc], arr_dmmv_q4_1_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32", arr_dmmv_q5_0_f16_f32_len[reduc], arr_dmmv_q5_0_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[reduc], arr_dmmv_q5_1_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[reduc], arr_dmmv_q8_0_f16_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[reduc16], arr_dmmv_q2_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[reduc16], arr_dmmv_q3_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[reduc16], arr_dmmv_q4_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[reduc16], arr_dmmv_q5_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32", arr_dmmv_q6_k_f16_f32_len[reduc16], arr_dmmv_q6_k_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32",   arr_dmmv_iq1_s_f16_f32_len[reduc16],   arr_dmmv_iq1_s_f16_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32",   arr_dmmv_iq1_m_f16_f32_len[reduc16],   arr_dmmv_iq1_m_f16_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32", arr_dmmv_iq2_xxs_f16_f32_len[reduc16], arr_dmmv_iq2_xxs_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32",  arr_dmmv_iq2_xs_f16_f32_len[reduc16],  arr_dmmv_iq2_xs_f16_f32_data[reduc16],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32",   arr_dmmv_iq2_s_f16_f32_len[reduc16],   arr_dmmv_iq2_s_f16_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32", arr_dmmv_iq3_xxs_f16_f32_len[reduc16], arr_dmmv_iq3_xxs_f16_f32_data[reduc16], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32",   arr_dmmv_iq3_s_f16_f32_len[reduc16],   arr_dmmv_iq3_s_f16_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32",  arr_dmmv_iq4_xs_f16_f32_len[reduc16],  arr_dmmv_iq4_xs_f16_f32_data[reduc16],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32",  arr_dmmv_iq4_nl_f16_f32_len[reduc16],  arr_dmmv_iq4_nl_f16_f32_data[reduc16],  "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32",   arr_dmmv_mxfp4_f16_f32_len[reduc16],   arr_dmmv_mxfp4_f16_f32_data[reduc16],   "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
             if (device->integer_dot_product) {
                 const uint32_t subgroup_size_int = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
                 const uint32_t wg_size_subgroup_int = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_int : (subgroup_size_int * 4);
 
-                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32", arr_dmmv_q4_0_q8_1_f32_len[reduc], arr_dmmv_q4_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
-                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32", arr_dmmv_q4_1_q8_1_f32_len[reduc], arr_dmmv_q4_1_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
-                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32", arr_dmmv_q5_0_q8_1_f32_len[reduc], arr_dmmv_q5_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
-                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32", arr_dmmv_q5_1_q8_1_f32_len[reduc], arr_dmmv_q5_1_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
-                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32", arr_dmmv_q8_0_q8_1_f32_len[reduc], arr_dmmv_q8_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32", arr_dmmv_q4_0_q8_1_f32_len[reduc], arr_dmmv_q4_0_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32", arr_dmmv_q4_1_q8_1_f32_len[reduc], arr_dmmv_q4_1_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32", arr_dmmv_q5_0_q8_1_f32_len[reduc], arr_dmmv_q5_0_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32", arr_dmmv_q5_1_q8_1_f32_len[reduc], arr_dmmv_q5_1_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32", arr_dmmv_q8_0_q8_1_f32_len[reduc], arr_dmmv_q8_0_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_MXFP4][i], "mul_mat_vec_mxfp4_q8_1_f32", arr_dmmv_mxfp4_q8_1_f32_len[reduc], arr_dmmv_mxfp4_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_q8_1_f32", arr_dmmv_q2_k_q8_1_f32_len[reduc], arr_dmmv_q2_k_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_kq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_q8_1_f32", arr_dmmv_q3_k_q8_1_f32_len[reduc], arr_dmmv_q3_k_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_q8_1_f32", arr_dmmv_q4_k_q8_1_f32_len[reduc], arr_dmmv_q4_k_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_q8_1_f32", arr_dmmv_q5_k_q8_1_f32_len[reduc], arr_dmmv_q5_k_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_q8_1_f32", arr_dmmv_q6_k_q8_1_f32_len[reduc], arr_dmmv_q6_k_q8_1_f32_data[reduc], "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int, i+1}, 1, true, use_subgroups, subgroup_size_int);
             }
 #endif // GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT
         }
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",        arr_dmmv_id_f32_f32_f32_len[reduc],     arr_dmmv_id_f32_f32_f32_data[reduc],     "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {wg_size_subgroup, 2}, 1, false, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",        arr_dmmv_id_f16_f32_f32_len[reduc],     arr_dmmv_id_f16_f32_f32_data[reduc],     "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {wg_size_subgroup, 2}, 1, false, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32",       arr_dmmv_id_bf16_f32_f32_len[reduc],    arr_dmmv_id_bf16_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {wg_size_subgroup, 2}, 1, false, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32",       arr_dmmv_id_q4_0_f32_f32_len[reduc],    arr_dmmv_id_q4_0_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32",       arr_dmmv_id_q4_1_f32_f32_len[reduc],    arr_dmmv_id_q4_1_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32",       arr_dmmv_id_q5_0_f32_f32_len[reduc],    arr_dmmv_id_q5_0_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32",       arr_dmmv_id_q5_1_f32_f32_len[reduc],    arr_dmmv_id_q5_1_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32",       arr_dmmv_id_q8_0_f32_f32_len[reduc],    arr_dmmv_id_q8_0_f32_f32_data[reduc],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq}, 1, true, use_subgroups, force_subgroup_size);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32",       arr_dmmv_id_q2_k_f32_f32_len[reduc16],    arr_dmmv_id_q2_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32",       arr_dmmv_id_q3_k_f32_f32_len[reduc16],    arr_dmmv_id_q3_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32",       arr_dmmv_id_q4_k_f32_f32_len[reduc16],    arr_dmmv_id_q4_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32",       arr_dmmv_id_q5_k_f32_f32_len[reduc16],    arr_dmmv_id_q5_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32",       arr_dmmv_id_q6_k_f32_f32_len[reduc16],    arr_dmmv_id_q6_k_f32_f32_data[reduc16],    "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ1_S],   "mul_mat_vec_id_iq1_s_f32",   arr_dmmv_id_iq1_s_f32_f32_len[reduc16],   arr_dmmv_id_iq1_s_f32_f32_data[reduc16],   "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ1_M],   "mul_mat_vec_id_iq1_m_f32",   arr_dmmv_id_iq1_m_f32_f32_len[reduc16],   arr_dmmv_id_iq1_m_f32_f32_data[reduc16],   "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", arr_dmmv_id_iq2_xxs_f32_f32_len[reduc16], arr_dmmv_id_iq2_xxs_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ2_XS],  "mul_mat_vec_id_iq2_xs_f32",  arr_dmmv_id_iq2_xs_f32_f32_len[reduc16],  arr_dmmv_id_iq2_xs_f32_f32_data[reduc16],  "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ2_S],   "mul_mat_vec_id_iq2_s_f32",   arr_dmmv_id_iq2_s_f32_f32_len[reduc16],   arr_dmmv_id_iq2_s_f32_f32_data[reduc16],   "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", arr_dmmv_id_iq3_xxs_f32_f32_len[reduc16], arr_dmmv_id_iq3_xxs_f32_f32_data[reduc16], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   arr_dmmv_id_iq3_s_f32_f32_len[reduc16],   arr_dmmv_id_iq3_s_f32_f32_data[reduc16],   "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ4_XS],  "mul_mat_vec_id_iq4_xs_f32",  arr_dmmv_id_iq4_xs_f32_f32_len[reduc16],  arr_dmmv_id_iq4_xs_f32_f32_data[reduc16],  "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_IQ4_NL],  "mul_mat_vec_id_iq4_nl_f32",  arr_dmmv_id_iq4_nl_f32_f32_len[reduc16],  arr_dmmv_id_iq4_nl_f32_f32_data[reduc16],  "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[w][GGML_TYPE_MXFP4],   "mul_mat_vec_id_mxfp4_f32",   arr_dmmv_id_mxfp4_f32_f32_len[reduc16],   arr_dmmv_id_mxfp4_f32_f32_data[reduc16],   "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq}, 1, true, use_subgroups16, force_subgroup_size16);
+
+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+        if (device->integer_dot_product) {
+            const uint32_t subgroup_size_int = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
+            const uint32_t wg_size_subgroup_int = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_int : (subgroup_size_int * 4);
+
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_q8_1_f32", arr_dmmv_id_q4_0_q8_1_f32_len[reduc], arr_dmmv_id_q4_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_q8_1_f32", arr_dmmv_id_q4_1_q8_1_f32_len[reduc], arr_dmmv_id_q4_1_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_q8_1_f32", arr_dmmv_id_q5_0_q8_1_f32_len[reduc], arr_dmmv_id_q5_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_q8_1_f32", arr_dmmv_id_q5_1_q8_1_f32_len[reduc], arr_dmmv_id_q5_1_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_q8_1_f32", arr_dmmv_id_q8_0_q8_1_f32_len[reduc], arr_dmmv_id_q8_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_MXFP4], "mul_mat_vec_id_mxfp4_q8_1_f32", arr_dmmv_id_mxfp4_q8_1_f32_len[reduc], arr_dmmv_id_mxfp4_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_q8_1_f32", arr_dmmv_id_q2_k_q8_1_f32_len[reduc], arr_dmmv_id_q2_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_q8_1_f32", arr_dmmv_id_q3_k_q8_1_f32_len[reduc], arr_dmmv_id_q3_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_q8_1_f32", arr_dmmv_id_q4_k_q8_1_f32_len[reduc], arr_dmmv_id_q4_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_q8_1_f32", arr_dmmv_id_q5_k_q8_1_f32_len[reduc], arr_dmmv_id_q5_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_q8_1_f32", arr_dmmv_id_q6_k_q8_1_f32_len[reduc], arr_dmmv_id_q6_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+        }
+#endif // GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT
     }
 
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_S],   "mul_mat_vec_id_iq1_s_f32",   mul_mat_vec_id_iq1_s_f32_len,   mul_mat_vec_id_iq1_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_M],   "mul_mat_vec_id_iq1_m_f32",   mul_mat_vec_id_iq1_m_f32_len,   mul_mat_vec_id_iq1_m_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS],  "mul_mat_vec_id_iq2_xs_f32",  mul_mat_vec_id_iq2_xs_f32_len,  mul_mat_vec_id_iq2_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S],   "mul_mat_vec_id_iq2_s_f32",   mul_mat_vec_id_iq2_s_f32_len,   mul_mat_vec_id_iq2_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   mul_mat_vec_id_iq3_s_f32_len,   mul_mat_vec_id_iq3_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS],  "mul_mat_vec_id_iq4_xs_f32",  mul_mat_vec_id_iq4_xs_f32_len,  mul_mat_vec_id_iq4_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL],  "mul_mat_vec_id_iq4_nl_f32",  mul_mat_vec_id_iq4_nl_f32_len,  mul_mat_vec_id_iq4_nl_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_MXFP4],   "mul_mat_vec_id_mxfp4_f32",   mul_mat_vec_id_mxfp4_f32_len,   mul_mat_vec_id_mxfp4_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+#if !defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+    GGML_UNUSED(rm_stdq_int);
+    GGML_UNUSED(rm_kq_int);
+#endif
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -3510,21 +3764,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
 
     if (device->subgroup_clustered && device->subgroup_require_full_support) {
-        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_subgroup_len, quantize_q8_1_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
         ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
         ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_len, quantize_q8_1_x4_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
     }
 
     for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
         if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
-            ggml_vk_create_pipeline2(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true);
+            ggml_vk_create_pipeline2(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_p021_push_constants), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true);
         } else {
-            ggml_vk_create_pipeline2(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
+            ggml_vk_create_pipeline2(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_p021_push_constants), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
         }
     }
-    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 12 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_nc_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@@ -3534,6 +3786,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_partials_f32, "rms_norm_partials_f32", rms_norm_partials_f32_len, rms_norm_partials_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_partials_f32, "rms_norm_mul_partials_f32", rms_norm_partials_f32_len, rms_norm_partials_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 1}, 1, true);
 
+    if (device->float_controls_rte_fp16 &&
+        sizeof(vk_op_rms_norm_mul_rope_push_constants) <= device->properties.limits.maxPushConstantsSize) {
+        ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_rope_f32_f32, "rms_norm_mul_rope_f32_f32", rms_norm_mul_rope_f32_f32_len, rms_norm_mul_rope_f32_f32_data, "main", 7, sizeof(vk_op_rms_norm_mul_rope_push_constants), {1, 1, 1}, {0, 1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_rope_f32_f16, "rms_norm_mul_rope_f32_f16", rms_norm_mul_rope_f32_f16_rte_len, rms_norm_mul_rope_f32_f16_rte_data, "main", 7, sizeof(vk_op_rms_norm_mul_rope_push_constants), {1, 1, 1}, {0, 1}, 1, true);
+    }
+
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
@@ -3553,6 +3811,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_i32_f32, "contig_cpy_i32_f32", contig_cpy_i32_f32_len, contig_cpy_i32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_i32, "contig_cpy_f32_i32", contig_cpy_f32_i32_len, contig_cpy_f32_i32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_32, "cpy_transpose_32", cpy_transpose_32_len, cpy_transpose_32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_16, "cpy_transpose_16", cpy_transpose_16_len, cpy_transpose_16_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
+
     if (device->float_controls_rte_fp16) {
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
@@ -3641,7 +3902,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_upscale_nearest_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_NEAREST}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_ac_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_bicubic_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BICUBIC}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -3650,6 +3911,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
+    if (device->float_controls_rte_fp16) {
+        ggml_vk_create_pipeline(device, device->pipeline_log[0], "log_f32_rte", log_f32_rte_len, log_f32_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_log[1], "log_f16_rte", log_f16_rte_len, log_f16_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_log[0], "log_f32", log_f32_len, log_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_log[1], "log_f16", log_f16_len, log_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    }
+
+    ggml_vk_create_pipeline(device, device->pipeline_tri[0], "tri_f32", tri_f32_len, tri_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_tri[1], "tri_f16", tri_f16_len, tri_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
@@ -3668,10 +3940,18 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_UNARY(gelu_quick)
     CREATE_UNARY(silu)
     CREATE_UNARY(relu)
+    CREATE_UNARY(neg)
     CREATE_UNARY(tanh)
     CREATE_UNARY(sigmoid)
     CREATE_UNARY(hardsigmoid)
     CREATE_UNARY(hardswish)
+    CREATE_UNARY(abs)
+    CREATE_UNARY(softplus)
+    CREATE_UNARY(step)
+    CREATE_UNARY(round)
+    CREATE_UNARY(ceil)
+    CREATE_UNARY(floor)
+    CREATE_UNARY(trunc)
 #undef CREATE_UNARY
 
 #define CREATE_UNARY_RTE(name)  \
@@ -3685,6 +3965,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_UNARY_RTE(exp)
 #undef CREATE_UNARY_RTE
 
+    ggml_vk_create_pipeline(device, device->pipeline_add1_f16_f16, "add1_f16_f16", add1_f16_f16_len, add1_f16_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add1_f16_f32, "add1_f16_f32", add1_f16_f32_len, add1_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add1_f32_f32, "add1_f32_f32", add1_f32_f32_len, add1_f32_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_arange_f32, "arange_f32", arange_f32_len, arange_f32_data, "main", 1, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_fill_f32, "fill_f32", fill_f32_len, fill_f32_data, "main", 1, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
 #define CREATE_GLU(name)  \
     if (device->float_controls_rte_fp16) {  \
         ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);   \
@@ -3737,15 +4025,50 @@ static void ggml_vk_load_shaders(vk_device& device) {
     }
 
     for (uint32_t i = 0; i < num_argsort_pipelines; ++i) {
-        ggml_vk_create_pipeline2(device, device->pipeline_argsort_f32[i], "argsort_f32_"+std::to_string(i), argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<<i, 1, 1}, {1u<<i, i}, 1, true);
+        uint32_t BLOCK_SIZE = 1u << std::min(i, device->max_workgroup_size_log2);
+        if (i <= device->max_workgroup_size_log2 &&
+            2 * sizeof(int) * BLOCK_SIZE <= device->properties.limits.maxComputeSharedMemorySize) {
+            const uint32_t NCOLS_PADDED_LOG2 = i;
+            ggml_vk_create_pipeline2(device, device->pipeline_argsort_f32[i], "argsort_f32_"+std::to_string(i), argsort_f32_len, argsort_f32_data, "main", 3, sizeof(vk_op_argsort_push_constants), {BLOCK_SIZE, 1, 1}, {BLOCK_SIZE, NCOLS_PADDED_LOG2}, 1, true);
+        }
+        const uint32_t WG_UNROLL_FACTOR = BLOCK_SIZE > 1 ? 2 : 1;
+        BLOCK_SIZE /= WG_UNROLL_FACTOR;
+        ggml_vk_create_pipeline2(device, device->pipeline_argsort_large_f32[i], "argsort_large_f32_"+std::to_string(i), argsort_large_f32_len, argsort_large_f32_data, "main", 3, sizeof(vk_op_argsort_push_constants), {BLOCK_SIZE * WG_UNROLL_FACTOR, 1, 1}, {BLOCK_SIZE, WG_UNROLL_FACTOR}, 1, true);
+    }
+
+    for (uint32_t i = 0; i < num_topk_pipelines; ++i) {
+        const uint32_t BLOCK_SIZE = 1u << i;
+        const uint32_t NCOLS_PADDED_LOG2 = i;
+        if (i <= device->max_workgroup_size_log2) {
+            uint32_t nary_shmem = 2 * sizeof(int) * BLOCK_SIZE +
+                                  sizeof(int) * device->subgroup_size +
+                                  2 * sizeof(int) +
+                                  (BLOCK_SIZE / device->subgroup_size) * sizeof(int);
+            if (device->subgroup_arithmetic && device->subgroup_require_full_support && device->subgroup_shuffle && device->subgroup_ballot &&
+                nary_shmem <= device->properties.limits.maxComputeSharedMemorySize) {
+                ggml_vk_create_pipeline2(device, device->pipeline_topk_f32[i], "topk_f32_"+std::to_string(i), topk_nary_search_f32_len, topk_nary_search_f32_data, "main", 2, sizeof(vk_op_topk_push_constants), {BLOCK_SIZE, 1, 1}, {BLOCK_SIZE, device->subgroup_size, device->subgroup_size_log2}, 1, true, true, device->subgroup_size);
+            } else if (2 * sizeof(int) * BLOCK_SIZE <= device->properties.limits.maxComputeSharedMemorySize) {
+                ggml_vk_create_pipeline2(device, device->pipeline_topk_f32[i], "topk_f32_"+std::to_string(i), topk_argsort_f32_len, topk_argsort_f32_data, "main", 2, sizeof(vk_op_topk_push_constants), {BLOCK_SIZE, 1, 1}, {BLOCK_SIZE, NCOLS_PADDED_LOG2}, 1, true);
+            }
+        }
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 128, device->subgroup_size }, 1, true, true, device->subgroup_size);
+
     ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
 
+    for (auto &s : device->pipeline_solve_tri_f32) {
+        const vk_solve_tri_pipeline_state &state = s.first;
+        ggml_vk_create_pipeline(
+            device, s.second, "solve_tri_f32",
+            solve_tri_f32_len, solve_tri_f32_data, "main", 3,
+            sizeof(vk_op_binary_push_constants), {1, 1, 1}, { 0, state.N, state.K }, 1, true);
+    }
+
 #define IM2COL(bda) \
     ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32 ## bda ## _len, im2col_f32 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);   \
     ggml_vk_create_pipeline(device, device->pipeline_im2col_3d_f32, "im2col_3d_f32", im2col_3d_f32 ## bda ## _len, im2col_3d_f32 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_3d_push_constants), {512, 1, 1}, { 512 }, 1, true);      \
@@ -3813,22 +4136,22 @@ static void ggml_vk_load_shaders(vk_device& device) {
         switch (s) {
         default:
         case CONV_SHAPE_128x128:
-            conv2d_BS_K = 128;
-            conv2d_BS_NPQ = 128;
+            conv2d_BS_K = conv_shapes_wg_denoms[CONV_SHAPE_128x128][0];
+            conv2d_BS_NPQ = conv_shapes_wg_denoms[CONV_SHAPE_128x128][1];
             conv2d_BS_CRS = 16;
             if (device->vendor_id == VK_VENDOR_ID_AMD && device->architecture != vk_device_architecture::AMD_GCN) {
                 conv2d_UNROLL = false;
             }
             break;
         case CONV_SHAPE_64x32:
-            conv2d_BS_K = 64;
-            conv2d_BS_NPQ = 32;
+            conv2d_BS_K = conv_shapes_wg_denoms[CONV_SHAPE_64x32][0];
+            conv2d_BS_NPQ = conv_shapes_wg_denoms[CONV_SHAPE_64x32][1];
             conv2d_BS_CRS = 32;
             conv2d_TS_K   = 4;
             break;
         case CONV_SHAPE_32x256:
-            conv2d_BS_K = 32;
-            conv2d_BS_NPQ = 256;
+            conv2d_BS_K = conv_shapes_wg_denoms[CONV_SHAPE_32x256][0];
+            conv2d_BS_NPQ = conv_shapes_wg_denoms[CONV_SHAPE_32x256][1];
             conv2d_BS_CRS = 16;
             break;
         }
@@ -3862,10 +4185,22 @@ static void ggml_vk_load_shaders(vk_device& device) {
         std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
 
 #define CREATE_CONV(name, type_suffix, spv_suffix) \
-        ggml_vk_create_pipeline( \
-            device, device->pipeline_##name##type_suffix[s], #name #type_suffix, \
-            name##type_suffix##spv_suffix##_len, name##type_suffix##spv_suffix##_data, "main", 3, \
-            sizeof(vk_op_##name##_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        for (auto &c : device->pipeline_##name##type_suffix[s]) { \
+            const vk_conv2d_pipeline_state &state = c.first;  \
+            std::vector<uint32_t> spec_constants_cpy = spec_constants; \
+            spec_constants_cpy.push_back(state.s0); \
+            spec_constants_cpy.push_back(state.s1); \
+            spec_constants_cpy.push_back(state.p0); \
+            spec_constants_cpy.push_back(state.p1); \
+            spec_constants_cpy.push_back(state.d0); \
+            spec_constants_cpy.push_back(state.d1); \
+            spec_constants_cpy.push_back(state.KW); \
+            spec_constants_cpy.push_back(state.KH); \
+            ggml_vk_create_pipeline( \
+                device, c.second, #name #type_suffix, \
+                name##type_suffix##spv_suffix##_len, name##type_suffix##spv_suffix##_data, "main", 3, \
+                sizeof(vk_op_##name##_push_constants), wg_denoms, spec_constants_cpy, 1, true, use_collectives);    \
+        }
 #define CREATE_CONVS(spv_suffix) \
         CREATE_CONV(conv2d, _f32, spv_suffix) \
         CREATE_CONV(conv2d, _f16_f32, spv_suffix) \
@@ -3901,7 +4236,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
     for (auto &c : compiles) {
         c.wait();
     }
-    device->need_compiles = false;
 }
 
 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch);
@@ -4057,6 +4391,16 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->vendor_id = device->properties.vendorID;
         device->driver_id = driver_props.driverID;
 
+        // Implementing the async backend interfaces seems broken on older Intel HW,
+        // see https://github.com/ggml-org/llama.cpp/issues/17302.
+        device->support_async = (device->vendor_id != VK_VENDOR_ID_INTEL ||
+                                 std::string(device->properties.deviceName.data()).find("(DG1)") == std::string::npos) &&
+                                getenv("GGML_VK_DISABLE_ASYNC") == nullptr;
+
+        if (!device->support_async) {
+            GGML_LOG_DEBUG("ggml_vulkan: WARNING: Async execution disabled on certain Intel devices.\n");
+        }
+
         const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");
 
         if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) {
@@ -4088,6 +4432,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size);
 
         device->subgroup_size = subgroup_props.subgroupSize;
+        device->subgroup_size_log2 = uint32_t(log2f(float(device->subgroup_size)));
         device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
         if (sm_builtins) {
             device->shader_core_count = sm_props.shaderSMCount;
@@ -4114,6 +4459,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->subgroup_ballot = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                   (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eBallot);
 
+        device->subgroup_vote = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
+                                (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eVote);
+
         const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;
 
         device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
@@ -4124,6 +4472,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->integer_dot_product = device->integer_dot_product && shader_integer_dot_product_props.integerDotProduct4x8BitPackedSignedAccelerated;
 
+        device->max_workgroup_size_log2 = uint32_t(log2f(float(device->properties.limits.maxComputeWorkGroupInvocations)));
+
         std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties();
 
         // Try to find a non-graphics compute queue and transfer-focused queues
@@ -4259,12 +4609,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
                             device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_multi_add_push_constants) &&
-                            vk12_features.runtimeDescriptorArray &&
-                            device->vendor_id != VK_VENDOR_ID_INTEL &&
                             getenv("GGML_VK_DISABLE_MULTI_ADD") == nullptr;
 
         device->shader_int64 = device_features2.features.shaderInt64;
         device->buffer_device_address = vk12_features.bufferDeviceAddress;
+        device->vulkan_memory_model = vk12_features.vulkanMemoryModel;
 
         if (device->subgroup_size_control) {
             device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
@@ -5007,6 +5356,8 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->prealloc_size_x = 0;
     ctx->prealloc_size_y = 0;
     ctx->prealloc_size_split_k = 0;
+    // Fixed size of 1KB, for deterministic behavior
+    ctx->prealloc_size_add_rms_partials = 1024;
 
     ctx->fence = ctx->device->device.createFence({});
     ctx->almost_ready_fence = ctx->device->device.createFence({});
@@ -5144,6 +5495,12 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             case GGML_TYPE_Q5_0:
             case GGML_TYPE_Q5_1:
             case GGML_TYPE_Q8_0:
+            case GGML_TYPE_MXFP4:
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+            case GGML_TYPE_Q6_K:
                 break;
             default:
                 return nullptr;
@@ -5283,9 +5640,28 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     }
 }
 
-static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t m, uint32_t k) {
     VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec_id()");
-    GGML_ASSERT(b_type == GGML_TYPE_F32);
+    GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_Q8_1);
+
+    if (b_type == GGML_TYPE_Q8_1) {
+        switch (a_type) {
+            case GGML_TYPE_Q4_0:
+            case GGML_TYPE_Q4_1:
+            case GGML_TYPE_Q5_0:
+            case GGML_TYPE_Q5_1:
+            case GGML_TYPE_Q8_0:
+            case GGML_TYPE_MXFP4:
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+            case GGML_TYPE_Q6_K:
+                break;
+            default:
+                return nullptr;
+        }
+    }
 
     switch (a_type) {
         case GGML_TYPE_F32:
@@ -5316,7 +5692,31 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
             return nullptr;
     }
 
-    return ctx->device->pipeline_dequant_mul_mat_vec_id_f32[a_type];
+    // heuristic to choose workgroup size
+    uint32_t dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
+    if ((ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA && ctx->device->architecture != vk_device_architecture::NVIDIA_PRE_TURING) || ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
+        // Prefer larger workgroups when M is small, to spread the work out more
+        // and keep more SMs busy.
+        // q6_k seems to prefer small workgroup size even for "medium" values of M.
+        if (a_type == GGML_TYPE_Q6_K) {
+            if (m < 4096 && k >= 1024) {
+                dmmv_wg = DMMV_WG_SIZE_LARGE;
+            }
+        } else {
+            if (m <= 8192 && k >= 1024) {
+                dmmv_wg = DMMV_WG_SIZE_LARGE;
+            }
+        }
+    }
+
+    if (b_type == GGML_TYPE_Q8_1) {
+        if (ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
+            dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
+        }
+        return ctx->device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[dmmv_wg][a_type];
+    }
+
+    return ctx->device->pipeline_dequant_mul_mat_vec_id_f32[dmmv_wg][a_type];
 }
 
 static void * ggml_vk_host_malloc(vk_device& device, size_t size) {
@@ -5367,7 +5767,7 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) {
     device->pinned_memory.erase(device->pinned_memory.begin() + index);
 }
 
-static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
+static void ggml_vk_host_get(const vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
     std::lock_guard<std::recursive_mutex> guard(device->mutex);
     buf = nullptr;
     buf_offset = 0;
@@ -5382,6 +5782,32 @@ static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf
     }
 }
 
+static vk_subbuffer ggml_vk_tensor_subbuffer(
+    const ggml_backend_vk_context * ctx, const ggml_tensor * tensor, bool allow_misalign = false) {
+
+    vk_buffer buffer = nullptr;
+    size_t offset = 0;
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, tensor->data, buffer, offset);
+    }
+    if (!buffer) {
+        auto buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
+        buffer = buf_ctx->dev_buffer;
+        offset = vk_tensor_offset(tensor) + tensor->view_offs;
+    }
+    GGML_ASSERT(buffer != nullptr);
+
+    size_t size = ggml_nbytes(tensor);
+
+    size_t misalign_bytes = offset & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    // The shader must support misaligned offsets when indexing into the buffer
+    GGML_ASSERT(allow_misalign || misalign_bytes == 0);
+    offset &= ~misalign_bytes;
+    size += misalign_bytes;
+
+    return vk_subbuffer{buffer, offset, size};
+}
+
 static vk_submission ggml_vk_begin_submission(vk_device& device, vk_command_pool& p, bool one_time = true) {
     vk_submission s;
     s.buffer = ggml_vk_create_cmd_buffer(device, p);
@@ -5505,6 +5931,16 @@ static void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size) {
     }
 }
 
+static void ggml_vk_ensure_sync_staging_buffer(ggml_backend_vk_context * ctx, size_t size) {
+    if (ctx->sync_staging == nullptr || ctx->sync_staging->size < size) {
+        VK_LOG_MEMORY("ggml_vk_ensure_sync_staging_buffer(" << size << ")");
+        ggml_vk_destroy_buffer(ctx->sync_staging);
+        ctx->sync_staging = ggml_vk_create_buffer_check(ctx->device, size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+    }
+}
+
 static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_context& subctx, vk_buffer& dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) {
     VK_LOG_DEBUG("ggml_vk_buffer_write_nc_async(" << tensor << ")");
     GGML_ASSERT(!ggml_is_contiguous(tensor));
@@ -5706,7 +6142,7 @@ static void ggml_vk_buffer_write(vk_buffer& dst, size_t offset, const void * src
     ggml_vk_buffer_write_2d(dst, offset, src, 0, size, 1);
 }
 
-static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging = false) {
+static bool ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging = false) {
     VK_LOG_DEBUG("ggml_vk_buffer_read_2d_async(offset=" << offset << ", width=" << width << ", height=" << height << ")");
     GGML_ASSERT(width > 0);
     GGML_ASSERT(height > 0);
@@ -5739,12 +6175,13 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
         ggml_vk_sync_buffers(nullptr, subctx);
         subctx->s->buffer.copyBuffer(src->buffer, buf->buffer, slices);
 
-        return;
+        return true;
     }
     VK_LOG_DEBUG("STAGING");
 
     if (!sync_staging) {
-        GGML_ABORT("Asynchronous read from non-pinned memory not supported");
+        // copy was not handled caller needs to fall back
+        return false;
     }
 
     // Fall back to staging buffer
@@ -5757,9 +6194,10 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
     subctx->s->buffer.copyBuffer(src->buffer, staging_buffer->buffer, slices);
 
     deferred_memcpy(dst, staging_buffer->ptr, copy_size, &subctx->out_memcpys);
+    return true;
 }
 
-static void ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t size, bool sync_staging = false) {
+static bool ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t size, bool sync_staging = false) {
     return ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, size, size, size, 1, sync_staging);
 }
 
@@ -5778,7 +6216,8 @@ static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_
 
         vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
-        ggml_vk_buffer_read_async(subctx, src, offset, dst, size, true);
+        bool ret = ggml_vk_buffer_read_async(subctx, src, offset, dst, size, true);
+        GGML_ASSERT(ret);
         ggml_vk_ctx_end(subctx);
 
         ggml_vk_submit(subctx, src->device->fence);
@@ -6041,6 +6480,17 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
     // Choose "contiguous copy" shader if src/dst are contiguous
     bool contig = ggml_is_contiguous(src) && (!dst || ggml_is_contiguous(dst));
 
+    // Use optimized "transpose" shader if src dim1 is the innermost dimension.
+    bool transpose = dst && src->nb[1] == ggml_type_size(to) && ggml_are_same_shape(dst, src);
+
+    if (transpose && src->type == to) {
+        if (ggml_type_size(to) == 4) {
+            return ctx->device->pipeline_cpy_transpose_32;
+        } else if (ggml_type_size(to) == 2) {
+            return ctx->device->pipeline_cpy_transpose_16;
+        }
+    }
+
     if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
         if (contig) {
             return ctx->device->pipeline_contig_cpy_f32_f32;
@@ -6143,7 +6593,7 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
     GGML_ABORT("fatal error");
 }
 
-static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) {
+static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline pipeline, const ggml_tensor * tensor, const vk_subbuffer & in, const vk_subbuffer & out) {
     VK_LOG_DEBUG("ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
     std::cerr << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")");
     const int tensor_type_size = ggml_type_size(tensor->type);
@@ -6172,30 +6622,30 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
     ggml_vk_sync_buffers(ctx, subctx);
 }
 
-static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type, bool use_x4_blocks) {
+static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
     switch(type) {
         case GGML_TYPE_Q8_1:
-            return use_x4_blocks ? ctx->device->pipeline_quantize_q8_1_x4 : ctx->device->pipeline_quantize_q8_1;
+            return ctx->device->pipeline_quantize_q8_1_x4;
         default:
             std::cerr << "Missing quantize pipeline for type: " << ggml_type_name(type) << std::endl;
             GGML_ABORT("fatal error");
     }
 }
 
-static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne, bool use_x4_blocks = false) {
+static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, const vk_subbuffer & in, const vk_subbuffer & out, uint32_t ne) {
     VK_LOG_DEBUG("ggml_vk_quantize_q8_1(" << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ", " << ne << ")");
 
-    vk_pipeline pipeline = use_x4_blocks ? ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true) : ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, false);
+    vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
 
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
     ggml_vk_sync_buffers(ctx, subctx);
 }
 
-static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool disable_split_k, bool dryrun = false) {
+static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool disable_split_k) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << ggml_type_name(src0->type) << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << ggml_type_name(src1->type) << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << ggml_type_name(dst->type) << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
-    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    std::cerr << "))");
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
 
@@ -6247,7 +6697,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
 
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;
+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && !y_non_contig && (ne11 * ne10) % 4 == 0;
 
     // Check for mmq first
     vk_matmul_pipeline mmp = quantize_y ? ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, GGML_TYPE_Q8_1, (ggml_prec)dst->op_params[0]) : nullptr;
@@ -6276,16 +6726,17 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11;
-    const int x_ne = ne01 * ne00;
-    const int y_ne = padded_n * ne10;
-    const int d_ne = ne11 * ne01;
+    const uint64_t x_ne = ggml_nelements(src0);
+    // 128 elements per Q8_1 x4 block
+    const uint64_t y_ne = padded_n * ne10 * ne12 * ne13;
+    const uint64_t d_ne = ggml_nelements(dst);
 
     const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, disable_split_k, pipeline);
 
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
     const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
-    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
+    const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
     const uint64_t d_sz = sizeof(float) * d_ne;
 
     vk_pipeline to_fp16_vk_0 = nullptr;
@@ -6306,30 +6757,28 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
 
     if (quantize_y) {
-        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
     }
 
-    if (dryrun) {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        uint64_t y_sz_upd = y_sz * ne12 * ne13;
-        if (quantize_y) {
-            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
-        }
-        const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0;
+    {
+        const uint64_t split_k_size = split_k > 1 ? d_sz * split_k : 0;
         if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
                 (split_k > 1 && split_k_size > ctx->device->properties.limits.maxStorageBufferRange)) {
             GGML_ABORT("Requested preallocation size is too large");
         }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
-        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
         if (split_k > 1 && ctx->prealloc_size_split_k < split_k_size) {
             ctx->prealloc_size_split_k = split_k_size;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
 
         // Request descriptor sets
@@ -6346,13 +6795,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         if (split_k > 1) {
             ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, 1);
         }
-        return;
     }
 
     vk_buffer d_D = dst_buf_ctx->dev_buffer;
     const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
     GGML_ASSERT(d_D != nullptr);
-    GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03);
+    GGML_ASSERT(d_D->size >= d_buf_offset + d_sz);
     vk_buffer d_X;
     uint64_t x_buf_offset = 0;
     vk_buffer d_Y;
@@ -6369,7 +6817,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     }
     if (qx_needs_dequant) {
         d_X = ctx->prealloc_x;
-        GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
+        GGML_ASSERT(d_X->size >= x_sz);
     } else {
         d_X = d_Qx;
         x_buf_offset = qx_buf_offset;
@@ -6377,10 +6825,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     }
     if (qy_needs_dequant) {
         d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
+        GGML_ASSERT(d_Y->size >= y_sz);
     } else if (quantize_y) {
         d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz, 144) * 144);
     } else {
         d_Y = d_Qy;
         y_buf_offset = qy_buf_offset;
@@ -6397,7 +6845,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, ggml_vk_subbuffer(ctx, d_Qx, qx_buf_offset), ggml_vk_subbuffer(ctx, d_X, 0));
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
-        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_X, 0, x_sz } }, pc, { (uint32_t)(x_ne), 1, 1});
         ggml_vk_sync_buffers(ctx, subctx);
     }
     if (y_non_contig) {
@@ -6417,7 +6865,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
-            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne * ne12 * ne13, true);
+            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
         }
@@ -6434,16 +6882,11 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
     }
 
-    uint32_t y_sz_total = y_sz * ne12 * ne13;
-    if (quantize_y) {
-        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
-    }
-
     // compute
     ggml_vk_matmul(
         ctx, subctx, pipeline,
-        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz_total },
-        ggml_vk_subbuffer(ctx, d_D, d_buf_offset), { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k },
+        { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz },
+        ggml_vk_subbuffer(ctx, d_D, d_buf_offset), { ctx->prealloc_split_k, 0, d_sz * split_k },
         ne01, ne11, ne10,
         ne10, ne10, stride_d, stride_batch_x, stride_batch_y, stride_batch_d,
         split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n
@@ -6465,20 +6908,35 @@ static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_
         return false;
     }
 
+    // General performance issue with q3_k and q6_k due to 2-byte alignment
+    if (src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q6_K) {
+        return false;
+    }
+
     // MMVQ is generally good for batches
     if (n > 1) {
         return true;
     }
 
+    // Quantization overhead is not worth it for small k
     switch (device->vendor_id) {
     case VK_VENDOR_ID_NVIDIA:
+        if (k <= 4096) {
+            return false;
+        }
+
         switch (src0_type) {
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q8_0:
             return device->architecture == vk_device_architecture::NVIDIA_PRE_TURING;
         default:
             return true;
         }
     case VK_VENDOR_ID_AMD:
+        if (k < 2048) {
+            return false;
+        }
+
         switch (src0_type) {
         case GGML_TYPE_Q8_0:
             return device->architecture == vk_device_architecture::AMD_GCN;
@@ -6486,6 +6944,10 @@ static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_
             return true;
         }
     case VK_VENDOR_ID_INTEL:
+        if (k < 2048) {
+            return false;
+        }
+
         switch (src0_type) {
         // From tests on A770 Linux, may need more tuning
         case GGML_TYPE_Q4_0:
@@ -6499,14 +6961,17 @@ static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_
     }
 
     GGML_UNUSED(m);
-    GGML_UNUSED(k);
 }
 
-static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
     VK_LOG_DEBUG("ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
-    std::cerr << "), " << (dryrun ? "dryrun" : "") << "),)");
+    std::cerr << ")),)");
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
 
@@ -6522,8 +6987,8 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
-    const uint64_t ne22 = dst->ne[2];
-    const uint64_t ne23 = dst->ne[3];
+    // const uint64_t ne22 = dst->ne[2];
+    // const uint64_t ne23 = dst->ne[3];
 
     const uint64_t r2 = ne12 / ne02;
     const uint64_t r3 = ne13 / ne03;
@@ -6533,30 +6998,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     GGML_ASSERT(ne11 == 1 || ne12 * ne13 == 1);
     bool batch_n = ne11 > 1;
 
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
-    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
-
-    vk_buffer d_Qx = nullptr;
-    size_t qx_buf_offset = 0;
-    vk_buffer d_Qy = nullptr;
-    size_t qy_buf_offset = 0;
-
-    bool src0_uma = false;
-    bool src1_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, src0->data, d_Qx, qx_buf_offset);
-        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
-        src0_uma = d_Qx != nullptr;
-        src1_uma = d_Qy != nullptr;
-    }
-
     const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
 
     const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
-    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0 && ggml_vk_should_use_mmvq(ctx->device, ne01, ne11, ne10, src0->type);
+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && !y_non_contig && (ne11 * ne10) % 4 == 0 && ggml_vk_should_use_mmvq(ctx->device, ne01, ne11, ne10, src0->type);
 
     vk_pipeline to_fp16_vk_0 = nullptr;
     vk_pipeline to_fp16_vk_1 = nullptr;
@@ -6580,7 +7026,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     }
 
     if (quantize_y) {
-        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
     }
 
     const bool qx_needs_dequant = x_non_contig;
@@ -6593,32 +7039,27 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
 
-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = ne11 * ne10;
-    const uint64_t d_ne = ne11 * ne01;
+    const uint64_t x_ne = ggml_nelements(src0);
+    const uint64_t y_ne = ggml_nelements(src1);
 
     const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
-    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
     const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
-    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
-    const uint64_t d_sz = sizeof(float) * d_ne;
+    const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) :
+                         (f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
 
-    if (dryrun) {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        uint64_t y_sz_upd = y_sz * ne12 * ne13;
-        if (quantize_y) {
-            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
-        }
+    {
         if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange)) {
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange)) {
             GGML_ABORT("Requested preallocation size is too large");
         }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
-        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
 
         // Request descriptor sets
@@ -6632,42 +7073,23 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
             ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
         }
         ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
-        return;
     }
 
-    vk_buffer d_D = dst_buf_ctx->dev_buffer;
-    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
-    GGML_ASSERT(d_D != nullptr);
-    vk_buffer d_X;
-    uint64_t x_buf_offset = 0;
-    vk_buffer d_Y;
-    uint64_t y_buf_offset = 0;
-    if(!src0_uma) {
-        d_Qx = src0_buf_ctx->dev_buffer;
-        qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
-        GGML_ASSERT(d_Qx != nullptr);
-    }
-    if(!src1_uma) {
-        d_Qy = src1_buf_ctx->dev_buffer;
-        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
-        GGML_ASSERT(d_Qy != nullptr);
-    }
+    vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
+    vk_subbuffer d_Qx = ggml_vk_tensor_subbuffer(ctx, src0);
+    vk_subbuffer d_Qy = ggml_vk_tensor_subbuffer(ctx, src1);
+    vk_subbuffer d_X, d_Y;
+
     if (qx_needs_dequant) {
-        d_X = ctx->prealloc_x;
+        d_X = { ctx->prealloc_x, 0, ctx->prealloc_x->size };
     } else {
         d_X = d_Qx;
-        x_buf_offset = qx_buf_offset;
         GGML_ASSERT(qx_sz == x_sz);
     }
-    if (qy_needs_dequant) {
-        d_Y = ctx->prealloc_y;
-    } else if (quantize_y) {
-        d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
+    if (qy_needs_dequant || quantize_y) {
+        d_Y = { ctx->prealloc_y, 0, ctx->prealloc_y->size };
     } else {
         d_Y = d_Qy;
-        y_buf_offset = qy_buf_offset;
-        GGML_ASSERT(qy_sz == y_sz);
     }
 
     if (x_non_contig) {
@@ -6676,7 +7098,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         }
 
         GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
-        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, ggml_vk_subbuffer(ctx, d_Qx, qx_buf_offset), ggml_vk_subbuffer(ctx, d_X, 0));
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, d_Qx, d_X);
     }
     if (y_non_contig) {
         GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
@@ -6685,7 +7107,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
-            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0));
+            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, d_Qy, d_Y);
             ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
         }
@@ -6696,7 +7118,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
-            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne * ne12 * ne13, true);
+            ggml_vk_quantize_q8_1(ctx, subctx, d_Qy, d_Y, y_ne);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
         }
@@ -6725,20 +7147,41 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
-    // TODO: Clean up this whole sz * ne_2 * ne_3 thing, it hasn't been necessary for a long time
-    uint32_t y_sz_total = y_sz * ne12 * ne13;
-    if (quantize_y) {
-        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
+    uint32_t fusion_flags = 0;
+
+    vk_subbuffer d_F0 = d_D;
+    if (ctx->num_additional_fused_ops > 0) {
+        const ggml_tensor * add = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * bias = add->src[0] == dst ? add->src[1] : add->src[0];
+
+        d_F0 = ggml_vk_tensor_subbuffer(ctx, bias);
+        fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS0;
+    }
+
+    vk_subbuffer d_F1 = d_D;
+    if (ctx->num_additional_fused_ops == 2) {
+        const ggml_tensor * add = cgraph->nodes[node_idx + 2];
+        const ggml_tensor * bias = add->src[0] == cgraph->nodes[node_idx + 1] ? add->src[1] : add->src[0];
+
+        d_F1 = ggml_vk_tensor_subbuffer(ctx, bias);
+        fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS1;
     }
 
     // compute
     const vk_mat_vec_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
         stride_batch_x, stride_batch_y, stride_batch_d,
+        fusion_flags,
         (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-                              { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz_total }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
+                              {
+                                d_X,
+                                d_Y,
+                                d_D,
+                                d_F0,
+                                d_F1,
+                              },
                               pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
 
     if (x_non_contig) {
@@ -6749,11 +7192,14 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     }
 }
 
-static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
     VK_LOG_DEBUG("ggml_vk_mul_mat_p021_f16_f32(" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
-    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    std::cerr << "))");
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]);  // NOLINT
     GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]);  // NOLINT
@@ -6765,67 +7211,56 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     const uint64_t ne02 = src0->ne[2];
     // const uint64_t ne03 = src0->ne[3];
 
-    const uint64_t ne10 = src1->ne[0];
+    //const uint64_t ne10 = src1->ne[0];
     const uint64_t ne11 = src1->ne[1];
     const uint64_t ne12 = src1->ne[2];
     // const uint64_t ne13 = src1->ne[3];
 
     GGML_ASSERT(ne11 == 1);
 
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
-    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
-
-    vk_buffer d_Qy = nullptr;
-    size_t qy_buf_offset = 0;
-
-    bool src1_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
-        src1_uma = d_Qy != nullptr;
-    }
-
-    const uint64_t x_ne = ne00 * ne01 * ne02;
-    const uint64_t y_ne = ne10 * ne11 * ne12;
-    const uint64_t d_ne = ne01 * ne11 * ne12;
-
-    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
-    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
-    const uint64_t d_sz = sizeof(float) * d_ne;
-
     // With grouped query attention there are > 1 Q matrices per K, V matrix.
     uint32_t gqa_ratio = (uint32_t)ne12 / (uint32_t)ne02;
     if (gqa_ratio > 8 || gqa_ratio == 0 || ne12 != ne02 * gqa_ratio) {
         gqa_ratio = 1;
     }
 
-    if (dryrun) {
+    {
         // Request descriptor sets
         ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
-        return;
     }
 
-    vk_buffer d_D = dst_buf_ctx->dev_buffer;
-    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
-    GGML_ASSERT(d_D != nullptr);
-    vk_buffer d_Qx = src0_buf_ctx->dev_buffer;
-    const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
-    GGML_ASSERT(d_Qx != nullptr);
-    if (!src1_uma) {
-        d_Qy = src1_buf_ctx->dev_buffer;
-        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
-        GGML_ASSERT(d_Qx != nullptr);
+    vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops], true);
+    vk_subbuffer d_Qx = ggml_vk_tensor_subbuffer(ctx, src0);
+    vk_subbuffer d_Qy = ggml_vk_tensor_subbuffer(ctx, src1, true);
+
+    vk_subbuffer d_F0 = d_D;
+
+    uint32_t fusion_flags = 0;
+
+    if (ctx->num_additional_fused_ops > 0) {
+        const ggml_tensor * add = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * bias = add->src[0] == dst ? add->src[1] : add->src[0];
+
+        d_F0 = ggml_vk_tensor_subbuffer(ctx, bias);
+        fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS0;
     }
 
-    const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-    const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset;
+    vk_subbuffer d_F1 = d_D;
+    if (ctx->num_additional_fused_ops > 1) {
+        const ggml_tensor * bias = cgraph->nodes[node_idx + 2]->src[1];
 
-    const uint64_t d_buffer_offset = (d_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-    const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
+        d_F1 = ggml_vk_tensor_subbuffer(ctx, bias);
+        fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS1;
+    }
 
     // compute
-    const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+
+    vk_mat_vec_p021_push_constants pc = {
+        (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12,
+        0, 0, fusion_flags
+    };
+
+    init_pushconst_tensor_offsets(ctx, pc, src0, src1, nullptr, nullptr, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
 
     uint32_t workgroups_z = (uint32_t)ne12;
     // When gqa_ratio > 1, each invocation does multiple rows and we can launch fewer workgroups
@@ -6833,14 +7268,24 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         workgroups_z /= gqa_ratio;
     }
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, workgroups_z });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1],
+        {
+            d_Qx,
+            d_Qy,
+            d_D,
+            d_F0,
+            d_F1,
+        }, pc, { 1, (uint32_t)ne01, workgroups_z });
 }
 
-static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
     VK_LOG_DEBUG("ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
-    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    std::cerr << "))");
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
@@ -6869,61 +7314,63 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     GGML_ASSERT(ne11 == 1);
     GGML_ASSERT(src0->ne[3] == src1->ne[3]); // checked in supports_op
 
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
-    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
-
-    vk_buffer d_Qy = nullptr;
-    size_t qy_buf_offset = 0;
-
-    bool src1_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
-        src1_uma = d_Qy != nullptr;
-    }
-
-    const uint64_t d_ne = ne01 * ne11 * ne12 * ne03;
-
     const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_y = nb12 / sizeof(float);
 
-    const uint64_t qx_sz = ggml_nbytes(src0);
-    const uint64_t qy_sz = ggml_nbytes(src1);
-    const uint64_t d_sz = sizeof(float) * d_ne;
-
-    if (dryrun) {
+    {
         // Request descriptor sets
         ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
-        return;
     }
 
-    vk_buffer d_D = dst_buf_ctx->dev_buffer;
-    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
-    GGML_ASSERT(d_D != nullptr);
-    vk_buffer d_Qx = src0_buf_ctx->dev_buffer;
-    const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
-    GGML_ASSERT(d_Qx != nullptr);
-    if (!src1_uma) {
-        d_Qy = src1_buf_ctx->dev_buffer;
-        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
-        GGML_ASSERT(d_Qx != nullptr);
+    vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops], true);
+    vk_subbuffer d_Qx = ggml_vk_tensor_subbuffer(ctx, src0);
+    vk_subbuffer d_Qy = ggml_vk_tensor_subbuffer(ctx, src1, true);
+    vk_subbuffer d_F0 = d_D;
+
+    uint32_t fusion_flags = 0;
+
+    if (ctx->num_additional_fused_ops > 0) {
+        const ggml_tensor * add = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * bias = add->src[0] == dst ? add->src[1] : add->src[0];
+
+        d_F0 = ggml_vk_tensor_subbuffer(ctx, bias);
+        fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS0;
     }
 
-    const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-    const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset;
+    vk_subbuffer d_F1 = d_D;
+    if (ctx->num_additional_fused_ops > 1) {
+        const ggml_tensor * bias = cgraph->nodes[node_idx + 2]->src[1];
 
-    const uint64_t d_buffer_offset = (d_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-    const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
+        d_F1 = ggml_vk_tensor_subbuffer(ctx, bias);
+        fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS1;
+    }
 
     // compute
-    const std::array<uint32_t, 12> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)), nb03, nb13, nb23 };
+    vk_mat_vec_nc_push_constants pc = {
+        (uint32_t)ne00, (uint32_t)ne01,
+        row_stride_x, channel_stride_x, channel_stride_y,
+        (uint32_t)(ne12 / ne02), (uint32_t)ne12,
+        0, 0,
+        nb03, nb13, nb23, fusion_flags
+    };
+
+    init_pushconst_tensor_offsets(ctx, pc, src0, src1, nullptr, nullptr, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
+
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
-        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { (uint32_t)ne03, (uint32_t)ne01, (uint32_t)ne12 });
+        {
+            d_Qx,
+            d_Qy,
+            d_D,
+            d_F0,
+            d_F1,
+        }, pc, { (uint32_t)ne03, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
-static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
     VK_LOG_DEBUG("ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")");
 
     // Handle huge A matrix by splitting the M dimensions. This works well for convolution use cases
@@ -6948,7 +7395,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, g
             dst2.ne[0] = cur_M_size;
             src02.ne[1] = cur_M_size;
 
-            ggml_vk_mul_mat_q_f16(ctx, subctx, &src02, src1, &dst2, true, dryrun);
+            ggml_vk_mul_mat_q_f16(ctx, subctx, &src02, src1, &dst2, true);
 
             m_offset += cur_M_size;
         }
@@ -6962,21 +7409,21 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, g
         src1->nb[1] <= src1->nb[3] &&
         src0->ne[3] == 1 &&
         src1->ne[3] == 1) {
-        ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
+        ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, cgraph, node_idx);
     } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1 &&
                !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
-        ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
+        ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, cgraph, node_idx);
     // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
     // when ne12 and ne13 are one.
     } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) &&
                (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || ggml_is_quantized(src0->type))) {
-        ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
+        ggml_vk_mul_mat_vec_q_f16(ctx, subctx, cgraph, node_idx);
     } else {
-        ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, false, dryrun);
+        ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, false);
     }
 }
 
-static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
@@ -6987,7 +7434,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
     const uint64_t ne02 = src0->ne[2];
-    const uint64_t ne03 = src0->ne[3];
+    // const uint64_t ne03 = src0->ne[3];
 
     const uint64_t ne10 = src1->ne[0];
     const uint64_t ne11 = src1->ne[1];
@@ -7002,8 +7449,8 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
-    const uint64_t ne22 = dst->ne[2];
-    const uint64_t ne23 = dst->ne[3];
+    // const uint64_t ne22 = dst->ne[2];
+    // const uint64_t ne23 = dst->ne[3];
 
     const uint64_t n_as = ne02;
 
@@ -7044,7 +7491,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;
+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && !y_non_contig && (ne11 * ne10) % 4 == 0;
 
     // Check for mmq first
     vk_matmul_pipeline mmp = quantize_y ? ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, GGML_TYPE_Q8_1, (ggml_prec)dst->op_params[0]) : nullptr;
@@ -7073,14 +7520,14 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = padded_n * ne10;
-    const uint64_t d_ne = ne21 * ne20;
+    const uint64_t x_ne = ggml_nelements(src0);
+    const uint64_t y_ne = padded_n * ne10 * ne12 * ne13;
+    const uint64_t d_ne = ggml_nelements(dst);
 
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
     const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
-    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
+    const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
     const uint64_t ids_sz = nbi2;
     const uint64_t d_sz = sizeof(float) * d_ne;
 
@@ -7102,25 +7549,22 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
 
     if (quantize_y) {
-        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
     }
 
-    if (dryrun) {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        uint64_t y_sz_upd = y_sz * ne12 * ne13;
-        if (quantize_y) {
-            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
-        }
+    {
         if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange)) {
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange)) {
             GGML_ABORT("Requested preallocation size is too large");
         }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
-        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
 
         // Request descriptor sets
@@ -7134,7 +7578,6 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         if (quantize_y) {
             ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
         }
-        return;
     }
 
     vk_buffer d_D = dst_buf_ctx->dev_buffer;
@@ -7161,7 +7604,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     }
     if (qx_needs_dequant) {
         d_X = ctx->prealloc_x;
-        GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
+        GGML_ASSERT(d_X->size >= x_sz);
     } else {
         d_X = d_Qx;
         x_buf_offset = qx_buf_offset;
@@ -7169,10 +7612,10 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     }
     if (qy_needs_dequant) {
         d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
+        GGML_ASSERT(d_Y->size >= y_sz);
     } else if (quantize_y) {
         d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz, 144) * 144);
     } else {
         d_Y = d_Qy;
         y_buf_offset = qy_buf_offset;
@@ -7190,7 +7633,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
         ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
-            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_X, 0, x_sz } }, pc, { (uint32_t)x_ne, 1, 1});
         ggml_vk_sync_buffers(ctx, subctx);
     }
     if (y_non_contig) {
@@ -7210,7 +7653,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
-            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne * ne12 * ne13, true);
+            ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
         }
@@ -7227,16 +7670,11 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
     }
 
-    uint32_t y_sz_total = y_sz * ne12 * ne13;
-    if (quantize_y) {
-        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
-    }
-
     // compute
     ggml_vk_matmul_id(
         ctx, subctx, pipeline,
-        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz_total },
-        { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz },
+        { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz },
+        { d_D, d_buf_offset, d_sz }, { d_ids, ids_buf_offset, ids_sz },
         ne01, ne21, ne10, ne10, ne10, ne01,
         stride_batch_x, stride_batch_y, ne20*ne21,
         n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n
@@ -7245,89 +7683,50 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     if (x_non_contig || qx_needs_dequant) {
         ctx->prealloc_x_need_sync = true;
     }
-    if (y_non_contig) {
+    if (y_non_contig || quantize_y) {
         ctx->prealloc_y_need_sync = true;
     }
 }
 
-static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
+    ggml_tensor * ids = dst->src[2];
     VK_LOG_DEBUG("ggml_vk_mul_mat_vec_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
-    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    std::cerr << "))");
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
     GGML_ASSERT(ids->type == GGML_TYPE_I32);
 
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
-    const uint64_t ne02 = src0->ne[2];
-    const uint64_t ne03 = src0->ne[3];
+    // const uint64_t ne02 = src0->ne[2];
+    // const uint64_t ne03 = src0->ne[3];
 
     const uint64_t ne10 = src1->ne[0];
     const uint64_t ne11 = src1->ne[1];
     const uint64_t ne12 = src1->ne[2];
-    const uint64_t ne13 = src1->ne[3];
+    // const uint64_t ne13 = src1->ne[3];
 
     const uint64_t nei0 = ids->ne[0];
     const uint64_t nei1 = ids->ne[1];
 
-    const uint64_t nbi2 = ids->nb[2];
-
     GGML_ASSERT(nei1 == 1);
 
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
-    const uint64_t ne22 = dst->ne[2];
-    const uint64_t ne23 = dst->ne[3];
-
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
-    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
-    ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
-
-    vk_buffer d_Qx = nullptr;
-    size_t qx_buf_offset = 0;
-    vk_buffer d_Qy = nullptr;
-    size_t qy_buf_offset = 0;
-    vk_buffer d_ids = nullptr;
-    size_t ids_buf_offset = 0;
-
-    bool src0_uma = false;
-    bool src1_uma = false;
-    bool ids_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, src0->data, d_Qx, qx_buf_offset);
-        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
-        ggml_vk_host_get(ctx->device, ids->data, d_ids, ids_buf_offset);
-        src0_uma = d_Qx != nullptr;
-        src1_uma = d_Qy != nullptr;
-        ids_uma = d_ids != nullptr;
-    }
+    // const uint64_t ne22 = dst->ne[2];
+    // const uint64_t ne23 = dst->ne[3];
 
     const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
 
     const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
-
-    const bool qx_needs_dequant = x_non_contig;
-    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
-
-    // Not implemented
-    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
-
-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = ne11 * ne10;
-    const uint64_t d_ne = ne21 * ne20;
-
-    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
-    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
-    const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
-    const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
-    const uint64_t ids_sz = nbi2;
-    const uint64_t d_sz = sizeof(float) * d_ne;
+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && !y_non_contig && (ne11 * ne10) % 4 == 0 && ggml_vk_should_use_mmvq(ctx->device, ne01, ne12, ne10, src0->type);
 
     vk_pipeline to_fp16_vk_0 = nullptr;
     vk_pipeline to_fp16_vk_1 = nullptr;
@@ -7339,24 +7738,51 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec_id(ctx, src0->type, src1->type);
+
+    // Check for mmq first
+    vk_pipeline dmmv = quantize_y ? ggml_vk_get_dequantize_mul_mat_vec_id(ctx, src0->type, GGML_TYPE_Q8_1, ne20, ne00) : nullptr;
+    vk_pipeline to_q8_1 = nullptr;
+
+    if (dmmv == nullptr) {
+        // Fall back to f16 dequant mul mat
+        dmmv = ggml_vk_get_dequantize_mul_mat_vec_id(ctx, src0->type, src1->type, ne20, ne00);
+        quantize_y = false;
+    }
+
+    if (quantize_y) {
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
+    }
+
+    const bool qx_needs_dequant = x_non_contig;
+    const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);
+
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
 
-    if (dryrun) {
-        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
+    const uint64_t x_ne = ggml_nelements(src0);
+    const uint64_t y_ne = ggml_nelements(src1);
+
+    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
+    const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
+    const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) :
+                                       (f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
+
+    {
         if (
-                (qx_needs_dequant && x_sz_upd > ctx->device->properties.limits.maxStorageBufferRange) ||
-                (qy_needs_dequant && y_sz_upd > ctx->device->properties.limits.maxStorageBufferRange)) {
+                (qx_needs_dequant && x_sz > ctx->device->properties.limits.maxStorageBufferRange) ||
+                (qy_needs_dequant && y_sz > ctx->device->properties.limits.maxStorageBufferRange)) {
             GGML_ABORT("Requested preallocation size is too large");
         }
-        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
-            ctx->prealloc_size_x = x_sz_upd;
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
-        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
-            ctx->prealloc_size_y = y_sz_upd;
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz) {
+            ctx->prealloc_size_y = y_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
         }
 
         // Request descriptor sets
@@ -7366,45 +7792,28 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         if (qy_needs_dequant) {
             ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
+        if (quantize_y) {
+            ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
+        }
         ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
-        return;
     }
 
-    vk_buffer d_D = dst_buf_ctx->dev_buffer;
-    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
-    GGML_ASSERT(d_D != nullptr);
-    vk_buffer d_X;
-    uint64_t x_buf_offset = 0;
-    vk_buffer d_Y;
-    uint64_t y_buf_offset = 0;
-    if(!src0_uma) {
-        d_Qx = src0_buf_ctx->dev_buffer;
-        qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
-        GGML_ASSERT(d_Qx != nullptr);
-    }
-    if(!src1_uma) {
-        d_Qy = src1_buf_ctx->dev_buffer;
-        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
-        GGML_ASSERT(d_Qy != nullptr);
-    }
-    if(!ids_uma) {
-        d_ids = ids_buf_ctx->dev_buffer;
-        ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs;
-        GGML_ASSERT(d_ids != nullptr);
-    }
+    vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
+    vk_subbuffer d_Qx = ggml_vk_tensor_subbuffer(ctx, src0);
+    vk_subbuffer d_Qy = ggml_vk_tensor_subbuffer(ctx, src1);
+    vk_subbuffer d_ids = ggml_vk_tensor_subbuffer(ctx, ids);
+    vk_subbuffer d_F0 = d_D;
+    vk_subbuffer d_X, d_Y;
+
     if (qx_needs_dequant) {
-        d_X = ctx->prealloc_x;
+        d_X = { ctx->prealloc_x, 0, ctx->prealloc_x->size };
     } else {
         d_X = d_Qx;
-        x_buf_offset = qx_buf_offset;
-        GGML_ASSERT(qx_sz == x_sz);
     }
-    if (qy_needs_dequant) {
-        d_Y = ctx->prealloc_y;
+    if (qy_needs_dequant || quantize_y) {
+        d_Y = { ctx->prealloc_y, 0, ctx->prealloc_y->size };
     } else {
         d_Y = d_Qy;
-        y_buf_offset = qy_buf_offset;
-        GGML_ASSERT(qy_sz == y_sz);
     }
 
     if (x_non_contig) {
@@ -7415,7 +7824,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
     if (x_non_contig) {
         GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
-        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, ggml_vk_subbuffer(ctx, d_Qx, qx_buf_offset), ggml_vk_subbuffer(ctx, d_X, 0));
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, d_Qx, d_X);
     }
     if (y_non_contig) {
         GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
@@ -7424,11 +7833,22 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
-            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0));
+            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, d_Qy, d_Y);
             ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
         }
     }
+    if (quantize_y) {
+        if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
+            ctx->prealloc_y_last_tensor_used != src1) {
+            if (ctx->prealloc_y_need_sync) {
+                ggml_vk_sync_buffers(ctx, subctx);
+            }
+            ggml_vk_quantize_q8_1(ctx, subctx, d_Qy, d_Y, y_ne);
+            ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
+            ctx->prealloc_y_last_tensor_used = src1;
+        }
+    }
 
     uint32_t stride_batch_y = ne10*ne11;
 
@@ -7446,31 +7866,72 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
+    uint32_t fusion_flags = 0;
+
+    if (ctx->num_additional_fused_ops > 0) {
+        const ggml_tensor * bias = cgraph->nodes[node_idx + 1]->src[1];
+
+        d_F0 = ggml_vk_tensor_subbuffer(ctx, bias);
+
+        if (cgraph->nodes[node_idx + 1]->op == GGML_OP_MUL) {
+            fusion_flags |= MAT_VEC_FUSION_FLAGS_SCALE0;
+        } else {
+            GGML_ASSERT(cgraph->nodes[node_idx + 1]->op == GGML_OP_ADD_ID);
+            fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS0;
+        }
+    }
+
+    vk_subbuffer d_F1 = d_D;
+    if (ctx->num_additional_fused_ops > 1) {
+        const ggml_tensor * scale = cgraph->nodes[node_idx + 2]->src[1];
+
+        d_F1 = ggml_vk_tensor_subbuffer(ctx, scale);
+        fusion_flags |= MAT_VEC_FUSION_FLAGS_SCALE1;
+    }
+
     // compute
     const vk_mat_vec_id_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        (uint32_t)x_ne, stride_batch_y, (uint32_t)(ne20*ne21),
+        (uint32_t)(ne00 * ne01), stride_batch_y, (uint32_t)(ne20 * ne21),
+        fusion_flags,
         (uint32_t)nei0, (uint32_t)ne11,
     };
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-        { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
-        vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
+        {
+            d_X,
+            d_Y,
+            d_D,
+            d_F0,
+            d_F1,
+            d_ids,
+        },
         pc, { groups_x, (uint32_t)nei0, groups_z });
 
     if (x_non_contig) {
         ctx->prealloc_x_need_sync = true;
     }
-    if (y_non_contig) {
+    if (y_non_contig || quantize_y) {
         ctx->prealloc_y_need_sync = true;
     }
 }
 
-static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+static bool ggml_vk_use_mul_mat_vec_id(const struct ggml_cgraph * cgraph, int node_idx) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src2 = dst->src[2];
+    return src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type));
+}
+
+static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
+    ggml_tensor * src2 = dst->src[2];
     VK_LOG_DEBUG("ggml_vk_mul_mat_id(" << src0 << ", " << src1 << ", " << src2 << ", " << dst << ")");
-    if (src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
-        ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
+    if (ggml_vk_use_mul_mat_vec_id(cgraph, node_idx)) {
+        ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, cgraph, node_idx);
     } else {
-        ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
+        ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst);
     }
 }
 
@@ -7478,7 +7939,7 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
     // Needs to be kept up to date on shader changes
     GGML_UNUSED(hsv);
     const uint32_t wg_size = scalar_flash_attention_workgroup_size;
-    const uint32_t Br = get_fa_scalar_num_large_rows(hsv);
+    const uint32_t Br = get_fa_scalar_num_large_rows(hsk, hsv);
     const uint32_t Bc = scalar_flash_attention_Bc;
 
     const uint32_t tmpsh = wg_size * sizeof(float);
@@ -7530,7 +7991,7 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     return supported;
 }
 
-static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, const ggml_tensor * sinks, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, const ggml_tensor * sinks, ggml_tensor * dst) {
     VK_LOG_DEBUG("ggml_vk_flash_attn((" << q << ", name=" << q->name << ", type=" << q->type << ", ne0=" << q->ne[0] << ", ne1=" << q->ne[1] << ", ne2=" << q->ne[2] << ", ne3=" << q->ne[3] << ", nb0=" << q->nb[0] << ", nb1=" << q->nb[1] << ", nb2=" << q->nb[2] << ", nb3=" << q->nb[3];
     std::cerr << "), (" << k << ", name=" << k->name << ", type=" << k->type << ", ne0=" << k->ne[0] << ", ne1=" << k->ne[1] << ", ne2=" << k->ne[2] << ", ne3=" << k->ne[3] << ", nb0=" << k->nb[0] << ", nb1=" << k->nb[1] << ", nb2=" << k->nb[2] << ", nb3=" << k->nb[3];
     std::cerr << "), (" << v << ", name=" << v->name << ", type=" << v->type << ", ne0=" << v->ne[0] << ", ne1=" << v->ne[1] << ", ne2=" << v->ne[2] << ", ne3=" << v->ne[3] << ", nb0=" << v->nb[0] << ", nb1=" << v->nb[1] << ", nb2=" << v->nb[2] << ", nb3=" << v->nb[3];
@@ -7538,7 +7999,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     if (sinks) {
         std::cerr << "), (" << sinks << ", name=" << sinks->name << ", type=" << sinks->type << ", ne0=" << sinks->ne[0] << ", ne1=" << sinks->ne[1] << ", ne2=" << sinks->ne[2] << ", ne3=" << sinks->ne[3] << ", nb0=" << sinks->nb[0] << ", nb1=" << sinks->nb[1] << ", nb2=" << sinks->nb[2] << ", nb3=" << sinks->nb[3];
     }
-    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    std::cerr << "))");
 
     GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
     GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
@@ -7609,7 +8070,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     case FA_SCALAR:
     case FA_COOPMAT1:
         // We may switch from coopmat1 to scalar, so use the scalar limit for both
-        max_gqa = get_fa_scalar_num_large_rows(HSV);
+        max_gqa = get_fa_scalar_num_large_rows(HSK, HSV);
         break;
     case FA_COOPMAT2:
         max_gqa = get_fa_num_small_rows(FA_COOPMAT2);
@@ -7675,12 +8136,15 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     vk_pipeline pipeline = nullptr;
 
-    auto &pipelines = ctx->device->pipeline_flash_attn_f32_f16[k->type];
-    auto it = pipelines.find(fa_pipeline_state);
-    if (it != pipelines.end()) {
-        pipeline = it->second;
-    } else {
-        pipelines[fa_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+    {
+        std::lock_guard<std::recursive_mutex> guard(ctx->device->mutex);
+        auto &pipelines = ctx->device->pipeline_flash_attn_f32_f16[k->type];
+        auto it = pipelines.find(fa_pipeline_state);
+        if (it != pipelines.end()) {
+            pipeline = it->second;
+        } else {
+            pipelines[fa_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+        }
     }
 
     assert(pipeline);
@@ -7712,15 +8176,15 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
     if (ctx->prealloc_size_split_k < split_k_size) {
         ctx->prealloc_size_split_k = split_k_size;
+        ggml_vk_preallocate_buffers(ctx, subctx);
     }
 
-    if (dryrun) {
+    {
         // Request descriptor sets
         ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (split_k > 1) {
             ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
         }
-        return;
     }
 
     float scale         = 1.0f;
@@ -7740,72 +8204,12 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr, d_S = nullptr;
-    size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0, s_buf_offset = 0;
-
-    bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false, S_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, q->data, d_Q, q_buf_offset);
-        ggml_vk_host_get(ctx->device, k->data, d_K, k_buf_offset);
-        ggml_vk_host_get(ctx->device, v->data, d_V, v_buf_offset);
-        ggml_vk_host_get(ctx->device, dst->data, d_D, d_buf_offset);
-        Q_uma = d_Q != nullptr;
-        K_uma = d_K != nullptr;
-        V_uma = d_V != nullptr;
-        D_uma = d_D != nullptr;
-        if (mask) {
-            ggml_vk_host_get(ctx->device, mask->data, d_M, m_buf_offset);
-            M_uma = d_M != nullptr;
-        }
-        if (sinks) {
-            ggml_vk_host_get(ctx->device, sinks->data, d_S, s_buf_offset);
-            S_uma = d_S != nullptr;
-        }
-    }
-
-
-    ggml_backend_vk_buffer_context * d_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * q_buf_ctx = (ggml_backend_vk_buffer_context *)q->buffer->context;
-    ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context;
-    ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context;
-
-    if (!Q_uma) {
-        d_Q = q_buf_ctx->dev_buffer;
-        q_buf_offset = vk_tensor_offset(q) + q->view_offs;
-    }
-    if (!K_uma) {
-        d_K = k_buf_ctx->dev_buffer;
-        k_buf_offset = vk_tensor_offset(k) + k->view_offs;
-    }
-    if (!V_uma) {
-        d_V = v_buf_ctx->dev_buffer;
-        v_buf_offset = vk_tensor_offset(v) + v->view_offs;
-    }
-    if (!D_uma) {
-        d_D = d_buf_ctx->dev_buffer;
-        d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
-    }
-
-    if (!M_uma) {
-        d_M = d_Q;
-        m_buf_offset = q_buf_offset;
-        if (mask) {
-            ggml_backend_vk_buffer_context * m_buf_ctx = (ggml_backend_vk_buffer_context*)mask->buffer->context;
-            d_M = m_buf_ctx->dev_buffer;
-            m_buf_offset = vk_tensor_offset(mask) + mask->view_offs;
-        }
-    }
-
-    if (!S_uma) {
-        d_S = d_Q;
-        s_buf_offset = q_buf_offset;
-        if (sinks) {
-            ggml_backend_vk_buffer_context * s_buf_ctx = (ggml_backend_vk_buffer_context*)sinks->buffer->context;
-            d_S = s_buf_ctx->dev_buffer;
-            s_buf_offset = vk_tensor_offset(sinks) + sinks->view_offs;
-        }
-    }
+    vk_subbuffer q_buf = ggml_vk_tensor_subbuffer(ctx, q);
+    vk_subbuffer k_buf = ggml_vk_tensor_subbuffer(ctx, k);
+    vk_subbuffer v_buf = ggml_vk_tensor_subbuffer(ctx, v);
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+    vk_subbuffer mask_buf = mask ? ggml_vk_tensor_subbuffer(ctx, mask) : q_buf;
+    vk_subbuffer sinks_buf = sinks ? ggml_vk_tensor_subbuffer(ctx, sinks) : q_buf;
 
     uint32_t mask_n_head_log2 = ((sinks != nullptr) << 24) | ((mask != nullptr) << 16) | n_head_log2;
 
@@ -7827,15 +8231,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
             ggml_vk_sync_buffers(ctx, subctx);
         }
 
+        vk_subbuffer split_k_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
-                                    {
-                                        ggml_vk_subbuffer(ctx, d_Q, q_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_K, k_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_V, v_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_M, m_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_S, s_buf_offset),
-                                        ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0),
-                                    },
+                                    {q_buf, k_buf, v_buf, mask_buf, sinks_buf, split_k_buf},
                                     // We only use split_k when group query attention is enabled, which means
                                     // there's no more than one tile of rows (i.e. workgroups_x would have been
                                     // one). We reuse workgroups_x to mean the number of splits, so we need to
@@ -7845,23 +8243,12 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         ggml_vk_sync_buffers(ctx, subctx);
         const std::array<uint32_t, 5> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k, (sinks != nullptr) };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
-                                    {
-                                        ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0),
-                                        ggml_vk_subbuffer(ctx, d_S, s_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_D, d_buf_offset),
-                                    },
+                                    {split_k_buf, sinks_buf, dst_buf},
                                     pc2, { (uint32_t)ne1, HSV, (uint32_t)ne3 });
         ctx->prealloc_split_k_need_sync = true;
     } else {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
-                                    {
-                                        ggml_vk_subbuffer(ctx, d_Q, q_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_K, k_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_V, v_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_M, m_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_S, s_buf_offset),
-                                        ggml_vk_subbuffer(ctx, d_D, d_buf_offset),
-                                    },
+                                    {q_buf, k_buf, v_buf, mask_buf, sinks_buf, dst_buf},
                                     pc, { workgroups_x, workgroups_y, workgroups_z });
     }
 }
@@ -8001,14 +8388,16 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_UPSCALE:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            int mode = ggml_get_op_params_i32(dst, 0);
+            ggml_scale_mode mode = (ggml_scale_mode)(ggml_get_op_params_i32(dst, 0) & 0xFF);
             switch (mode) {
                 case GGML_SCALE_MODE_NEAREST:
                     return ctx->device->pipeline_upscale_nearest_f32;
                 case GGML_SCALE_MODE_BILINEAR:
                     return ctx->device->pipeline_upscale_bilinear_f32;
-                case GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS:
-                    return ctx->device->pipeline_upscale_bilinear_ac_f32;
+                case GGML_SCALE_MODE_BICUBIC:
+                    return ctx->device->pipeline_upscale_bicubic_f32;
+                default:
+                    return nullptr;
             }
         }
         return nullptr;
@@ -8037,6 +8426,18 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_cos_f32;
         }
         return nullptr;
+    case GGML_OP_LOG:
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_log[dst->type == GGML_TYPE_F16];
+        }
+        return nullptr;
+    case GGML_OP_TRI:
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_tri[dst->type == GGML_TYPE_F16];
+        }
+        return nullptr;
     case GGML_OP_CLAMP:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_clamp_f32;
@@ -8126,6 +8527,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 return ctx->device->pipeline_gelu_quick[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_RELU:
                 return ctx->device->pipeline_relu[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_NEG:
+                return ctx->device->pipeline_neg[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_TANH:
                 return ctx->device->pipeline_tanh[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_SIGMOID:
@@ -8134,6 +8537,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 return ctx->device->pipeline_hardsigmoid[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_HARDSWISH:
                 return ctx->device->pipeline_hardswish[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_ABS:
+                return ctx->device->pipeline_abs[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_SOFTPLUS:
+                return ctx->device->pipeline_softplus[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_STEP:
+                return ctx->device->pipeline_step[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_ROUND:
+                return ctx->device->pipeline_round[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_CEIL:
+                return ctx->device->pipeline_ceil[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_FLOOR:
+                return ctx->device->pipeline_floor[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_TRUNC:
+                return ctx->device->pipeline_trunc[dst->type == GGML_TYPE_F16];
             default:
                 break;
         }
@@ -8236,19 +8653,6 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             }
             return nullptr;
         }
-    case GGML_OP_ARGSORT:
-        if (ctx->num_additional_fused_ops) {
-            uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
-            GGML_ASSERT(idx < num_topk_moe_pipelines);
-            topk_moe_mode mode = ggml_vk_num_additional_ops_to_topk_moe_mode(ctx->num_additional_fused_ops);
-            return ctx->device->pipeline_topk_moe[idx][mode];
-        }
-
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
-            uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
-            return ctx->device->pipeline_argsort_f32[idx];
-        }
-        return nullptr;
     case GGML_OP_SUM:
     case GGML_OP_SUM_ROWS:
     case GGML_OP_MEAN:
@@ -8256,6 +8660,31 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_sum_rows_f32;
         }
         return nullptr;
+    case GGML_OP_CUMSUM:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_cumsum_f32;
+        }
+        return nullptr;
+    case GGML_OP_SOLVE_TRI:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+
+            vk_solve_tri_pipeline_state solve_tri_pipeline_state(src0->ne[0], src1->ne[0]);
+
+            vk_pipeline pipeline = nullptr;
+
+            {
+                std::lock_guard<std::recursive_mutex> guard(ctx->device->mutex);
+                auto it = ctx->device->pipeline_solve_tri_f32.find(solve_tri_pipeline_state);
+                if (it != ctx->device->pipeline_solve_tri_f32.end()) {
+                    pipeline = it->second;
+                } else {
+                    ctx->device->pipeline_solve_tri_f32[solve_tri_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+                }
+            }
+
+            return pipeline;
+        }
+        return nullptr;
     case GGML_OP_ARGMAX:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
             return ctx->device->pipeline_argmax_f32;
@@ -8341,14 +8770,14 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_CONV_TRANSPOSE_2D:
         if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
             ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-            std::array<uint32_t, 3> elements;
+            std::array<uint32_t, 3> elements{};
             if (op == GGML_OP_CONV_2D) elements = ggml_vk_get_conv_elements(dst);
             else if (op == GGML_OP_CONV_TRANSPOSE_2D) elements = ggml_vk_get_conv_transpose_2d_elements(dst);
             vk_conv_shapes shape;
 
             uint32_t tiles[CONV_SHAPE_COUNT];
             for (uint32_t i = 0; i < CONV_SHAPE_COUNT; ++i) {
-                tiles[i] = CEIL_DIV(elements[0], ctx->device->pipeline_conv2d_f32[i]->wg_denoms[0]) * CEIL_DIV(elements[1], ctx->device->pipeline_conv2d_f32[i]->wg_denoms[1]);
+                tiles[i] = CEIL_DIV(elements[0], conv_shapes_wg_denoms[i][0]) * CEIL_DIV(elements[1], conv_shapes_wg_denoms[i][1]);
             }
 
             // We can't query number of shader cores on Intel, use 32 as a placeholder
@@ -8363,19 +8792,45 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 shape = CONV_SHAPE_64x32;
             }
 
+            uint32_t KW = static_cast<uint32_t>(src0->ne[0]);
+            uint32_t KH = static_cast<uint32_t>(src0->ne[1]);
+            uint32_t s0 = static_cast<uint32_t>(dst->op_params[0]);
+            uint32_t s1 = op == GGML_OP_CONV_2D ? static_cast<uint32_t>(dst->op_params[1]) : static_cast<uint32_t>(dst->op_params[0]);
+            uint32_t p0 = op == GGML_OP_CONV_2D ? static_cast<uint32_t>(dst->op_params[2]) : 0;
+            uint32_t p1 = op == GGML_OP_CONV_2D ? static_cast<uint32_t>(dst->op_params[3]) : 0;
+            uint32_t d0 = op == GGML_OP_CONV_2D ? static_cast<uint32_t>(dst->op_params[4]) : 1;
+            uint32_t d1 = op == GGML_OP_CONV_2D ? static_cast<uint32_t>(dst->op_params[5]) : 1;
+
+            vk_conv2d_pipeline_state conv2d_pipeline_state(s0, s1, p0, p1, d0, d1, KW, KH);
+
+            std::map<vk_conv2d_pipeline_state, vk_pipeline> *pipelines = nullptr;
             if (op == GGML_OP_CONV_2D) {
                 if (src0->type == GGML_TYPE_F32) {
-                    return ctx->device->pipeline_conv2d_f32[shape];
+                    pipelines = &ctx->device->pipeline_conv2d_f32[shape];
                 } else if (src0->type == GGML_TYPE_F16) {
-                    return ctx->device->pipeline_conv2d_f16_f32[shape];
+                    pipelines = &ctx->device->pipeline_conv2d_f16_f32[shape];
                 }
             } else if (op == GGML_OP_CONV_TRANSPOSE_2D) {
                 if (src0->type == GGML_TYPE_F32) {
-                    return ctx->device->pipeline_conv_transpose_2d_f32[shape];
+                    pipelines = &ctx->device->pipeline_conv_transpose_2d_f32[shape];
                 } else if (src0->type == GGML_TYPE_F16) {
-                    return ctx->device->pipeline_conv_transpose_2d_f16_f32[shape];
+                    pipelines = &ctx->device->pipeline_conv_transpose_2d_f16_f32[shape];
                 }
             }
+
+            vk_pipeline pipeline = nullptr;
+
+            {
+                std::lock_guard<std::recursive_mutex> guard(ctx->device->mutex);
+                auto it = pipelines->find(conv2d_pipeline_state);
+                if (it != pipelines->end()) {
+                    pipeline = it->second;
+                } else {
+                    (*pipelines)[conv2d_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+                }
+            }
+
+            return pipeline;
         }
         return nullptr;
     case GGML_OP_CONV_2D_DW:
@@ -8393,6 +8848,27 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             }
         }
         return nullptr;
+    case GGML_OP_ADD1:
+        if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_add1_f16_f16;
+        }
+        if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_add1_f16_f32;
+        }
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_add1_f32_f32;
+        }
+        return nullptr;
+    case GGML_OP_ARANGE:
+        if (dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_arange_f32;
+        }
+        return nullptr;
+    case GGML_OP_FILL:
+        if (dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_fill_f32;
+        }
+        return nullptr;
     default:
         return nullptr;
     }
@@ -8400,60 +8876,6 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     GGML_UNUSED(src2);
 }
 
-static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
-    switch (op) {
-    case GGML_OP_CPY:
-    case GGML_OP_GET_ROWS:
-    case GGML_OP_ADD:
-    case GGML_OP_SUB:
-    case GGML_OP_MUL:
-    case GGML_OP_DIV:
-    case GGML_OP_ADD_ID:
-    case GGML_OP_CONCAT:
-    case GGML_OP_UPSCALE:
-    case GGML_OP_SQR:
-    case GGML_OP_SQRT:
-    case GGML_OP_SIN:
-    case GGML_OP_COS:
-    case GGML_OP_CLAMP:
-    case GGML_OP_PAD:
-    case GGML_OP_REPEAT:
-    case GGML_OP_REPEAT_BACK:
-    case GGML_OP_ROPE:
-    case GGML_OP_RMS_NORM:
-    case GGML_OP_CONV_2D_DW:
-    case GGML_OP_IM2COL:
-    case GGML_OP_IM2COL_3D:
-    case GGML_OP_SET_ROWS:
-    case GGML_OP_SUM:
-    case GGML_OP_SUM_ROWS:
-    case GGML_OP_MEAN:
-        return true;
-    default:
-        return false;
-    }
-}
-
-static uint32_t get_misalign_bytes(ggml_backend_vk_context * ctx, const ggml_tensor * t)
-{
-    return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
-}
-
-template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
-    GGML_UNUSED(p);
-    GGML_UNUSED(src0);
-    GGML_UNUSED(src1);
-    GGML_UNUSED(src2);
-    GGML_UNUSED(src3);
-    GGML_UNUSED(dst);
-    static_assert(!std::is_const<T>::value, "unexpected type");
-    GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0);
-    GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0);
-    GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0);
-    GGML_ASSERT(!src3 || get_misalign_bytes(ctx, src3) == 0);
-    GGML_ASSERT(!dst  || get_misalign_bytes(ctx, dst) == 0);
-}
-
 template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
@@ -8524,7 +8946,7 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
 }
 
 template<typename PC>
-static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst, ggml_op op, PC&& pc) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     if (src1 != nullptr) {
         std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
@@ -8536,43 +8958,22 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         std::cerr << "), (" << src3 << ", name=" << src3->name << ", type=" << src3->type << ", ne0=" << src3->ne[0] << ", ne1=" << src3->ne[1] << ", ne2=" << src3->ne[2] << ", ne3=" << src3->ne[3] << ", nb0=" << src3->nb[0] << ", nb1=" << src3->nb[1] << ", nb2=" << src3->nb[2] << ", nb3=" << src3->nb[3];
     }
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
-    std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")");
+    std::cerr << "), " << ggml_op_name(op) << ")");
     GGML_ASSERT(op == GGML_OP_GET_ROWS || op == GGML_OP_CPY || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
-    GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0));  // NOLINT
     GGML_ASSERT(dst->buffer != nullptr);
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
     const uint64_t ne02 = src0->ne[2];
     const uint64_t ne03 = src0->ne[3];
-    const uint64_t ne0 = ne00 * ne01;
 
     const bool use_src1 = src1 != nullptr;
     const uint64_t ne10 = use_src1 ? src1->ne[0] : 0;
     const uint64_t ne11 = use_src1 ? src1->ne[1] : 0;
     const uint64_t ne12 = use_src1 ? src1->ne[2] : 0;
     const uint64_t ne13 = use_src1 ? src1->ne[3] : 0;
-    const uint64_t ne1 = ne10 * ne11;
-    // const uint64_t nb10 = use_src1 ? src1->nb[0] : 0;
 
     const bool use_src2 = src2 != nullptr;
-    const uint64_t ne20 = use_src2 ? src2->ne[0] : 0;
-    const uint64_t ne21 = use_src2 ? src2->ne[1] : 0;
-    const uint64_t ne22 = use_src2 ? src2->ne[2] : 0;
-    const uint64_t ne23 = use_src2 ? src2->ne[3] : 0;
-    const uint64_t ne2 = ne20 * ne21;
-
     const bool use_src3 = src3 != nullptr;
-    const uint64_t ne30 = use_src3 ? src3->ne[0] : 0;
-    const uint64_t ne31 = use_src3 ? src3->ne[1] : 0;
-    const uint64_t ne32 = use_src3 ? src3->ne[2] : 0;
-    const uint64_t ne33 = use_src3 ? src3->ne[3] : 0;
-    const uint64_t ne3 = ne30 * ne31;
-
-    const uint64_t ned0 = dst->ne[0];
-    const uint64_t ned1 = dst->ne[1];
-    const uint64_t ned2 = dst->ne[2];
-    const uint64_t ned3 = dst->ne[3];
-    const uint64_t ned = ned0 * ned1;
 
     init_pushconst_fastdiv(pc);
 
@@ -8587,87 +8988,19 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         GGML_ABORT("fatal error");
     }
 
-    if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
-        return;
-    }
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
-    const bool op_supports_incontiguous = ggml_vk_op_supports_incontiguous(op);
+    vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer(ctx, src0, true);
+    vk_subbuffer src1_buf = use_src1 ? ggml_vk_tensor_subbuffer(ctx, src1, true) : vk_subbuffer{};
+    vk_subbuffer src2_buf = use_src2 ? ggml_vk_tensor_subbuffer(ctx, src2, true) : vk_subbuffer{};
+    vk_subbuffer src3_buf = use_src3 ? ggml_vk_tensor_subbuffer(ctx, src3, true) : vk_subbuffer{};
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, true);
 
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
-    ggml_backend_vk_buffer_context * src1_buf_ctx = use_src1 ? (ggml_backend_vk_buffer_context *)src1->buffer->context : nullptr;
-    ggml_backend_vk_buffer_context * src2_buf_ctx = use_src2 ? (ggml_backend_vk_buffer_context *)src2->buffer->context : nullptr;
-    ggml_backend_vk_buffer_context * src3_buf_ctx = use_src3 ? (ggml_backend_vk_buffer_context *)src3->buffer->context : nullptr;
-
-    vk_buffer d_X = nullptr;
-    size_t x_buf_offset = 0;
-    vk_buffer d_Y = nullptr;
-    size_t y_buf_offset = 0;
-    vk_buffer d_Z = nullptr;
-    size_t z_buf_offset = 0;
-    vk_buffer d_W = nullptr;
-    size_t w_buf_offset = 0;
-
-    bool src0_uma = false;
-    bool src1_uma = false;
-    bool src2_uma = false;
-    bool src3_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, src0->data, d_X, x_buf_offset);
-        src0_uma = d_X != nullptr;
-        if (use_src1) {
-            ggml_vk_host_get(ctx->device, src1->data, d_Y, y_buf_offset);
-            src1_uma = d_Y != nullptr;
-        }
-        if (use_src2) {
-            ggml_vk_host_get(ctx->device, src2->data, d_Z, z_buf_offset);
-            src2_uma = d_Z != nullptr;
-        }
-        if (use_src3) {
-            ggml_vk_host_get(ctx->device, src3->data, d_W, w_buf_offset);
-            src3_uma = d_W != nullptr;
-        }
-    }
-
-    vk_buffer d_D = dst_buf_ctx->dev_buffer;
-
-    GGML_ASSERT(d_D != nullptr);
-    uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
-    if(!src0_uma) {
-        d_X = src0_buf_ctx->dev_buffer;
-        x_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
-        GGML_ASSERT(d_X != nullptr);
-    }
-    if (use_src1 && !src1_uma) {
-        d_Y = src1_buf_ctx->dev_buffer;
-        y_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
-        GGML_ASSERT(d_Y != nullptr);
-    }
-    if (use_src2 && !src2_uma) {
-        d_Z = src2_buf_ctx->dev_buffer;
-        z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
-        GGML_ASSERT(d_Z != nullptr);
-    }
-    if (use_src3 && !src3_uma) {
-        d_W = src3_buf_ctx->dev_buffer;
-        w_buf_offset = vk_tensor_offset(src3) + src3->view_offs;
-        GGML_ASSERT(d_W != nullptr);
-    }
-    // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets.
+    // Compute misalignment offset for descriptors and store it in in push constants.
     init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, src3, dst);
-    x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-    y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-    z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-    w_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
-    d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
 
     std::array<uint32_t, 3> elements;
 
-    // Single call if dimension 2 is contiguous
-    GGML_ASSERT(op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1))));
-
     switch (op) {
     case GGML_OP_NORM:
     case GGML_OP_RMS_NORM_BACK:
@@ -8675,6 +9008,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_SOFT_MAX:
     case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_CUMSUM:
     case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
         {
@@ -8687,6 +9021,18 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                 elements = { nr, 1, 1 };
             }
         } break;
+    case GGML_OP_SOLVE_TRI:
+        {
+            uint32_t nr = (uint32_t)(ne02 * ne03);
+            if (nr > 262144) {
+                elements = { 512, 512, CEIL_DIV(nr, 262144) };
+            } else if (nr > 512) {
+                elements = { 512, CEIL_DIV(nr, 512), 1 };
+            } else {
+                elements = { nr, 1, 1 };
+            }
+        }
+        break;
     case GGML_OP_RMS_NORM:
         if (ctx->do_add_rms_partials) {
             // Run one element per thread, 128 threads per workgroup
@@ -8716,8 +9062,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
         break;
     case GGML_OP_ARGSORT:
-        elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 };
-        elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+        GGML_ASSERT(0);
         break;
     case GGML_OP_IM2COL:
         {
@@ -8745,9 +9090,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             const uint32_t KH = ne01;
             const uint32_t KW = ne00;
 
-            const uint32_t OD = ned3 / N;
-            const uint32_t OH = ned2;
-            const uint32_t OW = ned1;
+            const uint32_t OD = dst->ne[3] / N;
+            const uint32_t OH = dst->ne[2];
+            const uint32_t OW = dst->ne[1];
 
             const uint32_t IC_KD_KH_KW = IC*KD*KH*KW;
             const uint32_t N_OD_OH = N*OD*OH;
@@ -8785,11 +9130,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_SUB:
     case GGML_OP_DIV:
     case GGML_OP_MUL:
+    case GGML_OP_ADD1:
+    case GGML_OP_ARANGE:
+    case GGML_OP_FILL:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
+    case GGML_OP_LOG:
+    case GGML_OP_TRI:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_ROLL:
@@ -8825,6 +9175,17 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             } else {
                 elements = { ne, 1, 1 };
             }
+
+            if (pipeline == ctx->device->pipeline_cpy_transpose_32 ||
+                pipeline == ctx->device->pipeline_cpy_transpose_16) {
+                // 32x32 tiles
+                elements[0] = (uint32_t)CEIL_DIV(dst->ne[0], 32);
+                elements[1] = (uint32_t)CEIL_DIV(dst->ne[1], 32);
+                elements[2] = (uint32_t)(dst->ne[2]*dst->ne[3]);
+                elements[0] = std::min(elements[0], ctx->device->properties.limits.maxComputeWorkGroupCount[0]);
+                elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+                elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
+            }
         } break;
     case GGML_OP_ADD_ID:
         {
@@ -8862,116 +9223,54 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         break;
     }
 
-    uint64_t x_sz, y_sz, z_sz, w_sz, d_sz;
-
-    if (op_supports_incontiguous) {
-        x_sz = ggml_nbytes(src0) + get_misalign_bytes(ctx, src0);
-        y_sz = use_src1 ? ggml_nbytes(src1) + get_misalign_bytes(ctx, src1) : 0;
-        z_sz = use_src2 ? ggml_nbytes(src2) + get_misalign_bytes(ctx, src2) : 0;
-        w_sz = use_src3 ? ggml_nbytes(src3) + get_misalign_bytes(ctx, src3) : 0;
-        d_sz = ggml_nbytes(dst) + get_misalign_bytes(ctx, dst);
-
-        if (x_buf_offset + x_sz >= d_X->size) {
-            x_sz = ggml_vk_get_max_buffer_range(ctx, d_X, x_buf_offset);
-        }
-        if (use_src1 && y_buf_offset + y_sz >= d_Y->size) {
-            y_sz = ggml_vk_get_max_buffer_range(ctx, d_Y, y_buf_offset);
-        }
-        if (use_src2 && z_buf_offset + z_sz >= d_Z->size) {
-            z_sz = ggml_vk_get_max_buffer_range(ctx, d_Z, z_buf_offset);
-        }
-        if (use_src3 && w_buf_offset + w_sz >= d_W->size) {
-            w_sz = ggml_vk_get_max_buffer_range(ctx, d_W, w_buf_offset);
-        }
-        if (d_buf_offset + d_sz >= d_D->size) {
-            d_sz = ggml_vk_get_max_buffer_range(ctx, d_D, d_buf_offset);
-        }
-    } else {
-        x_sz = ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0 * ne02 * ne03;
-        y_sz = use_src1 ? ggml_type_size(src1->type) * ne1 * ne12 * ne13 : 0;
-        z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 * ne22 * ne23 : 0;
-        w_sz = use_src3 ? ggml_type_size(src3->type) * ne3 * ne32 * ne33 : 0;
-        d_sz = ggml_type_size(dst->type) * ned * ned2 * ned3;
-    }
-
     if (op == GGML_OP_ADD || op == GGML_OP_RMS_NORM) {
-        vk_buffer d_A = ctx->do_add_rms_partials ? ctx->prealloc_add_rms_partials : d_X;
-        size_t a_buf_offset = ctx->do_add_rms_partials ? ctx->prealloc_size_add_rms_partials_offset : 0;
+        vk_subbuffer a_buf = src0_buf;
+        if (ctx->do_add_rms_partials) {
+            a_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_add_rms_partials, ctx->prealloc_size_add_rms_partials_offset);
+        }
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
-            { vk_subbuffer{ d_X, x_buf_offset, x_sz },
-              vk_subbuffer{ d_Y, y_buf_offset, y_sz },
-              vk_subbuffer{ d_D, d_buf_offset, d_sz },
-              ggml_vk_subbuffer(ctx, d_A, a_buf_offset),
-            }, pc, elements);
+            { src0_buf, src1_buf, dst_buf, a_buf }, pc, elements);
     } else if (op == GGML_OP_GLU) {
         // Empty src1 is possible in glu, but the shader needs a buffer
-        vk_subbuffer subbuf_y;
-        if (use_src1) {
-            subbuf_y = { d_Y, y_buf_offset, y_sz };
-        } else {
-            subbuf_y = { d_X, 0, x_sz };
-        }
-
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        vk_subbuffer subbuf1 = use_src1 ? src1_buf : src0_buf;
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, subbuf1, dst_buf }, pc, elements);
     } else if (op == GGML_OP_SOFT_MAX) {
         // Empty src1 and src2 is possible in soft_max, but the shader needs a buffer
-        vk_subbuffer subbuf_y;
-        if (use_src1) {
-            subbuf_y = { d_Y, y_buf_offset, y_sz };
-        } else {
-            subbuf_y = { d_X, 0, x_sz };
-        }
-
-        vk_subbuffer subbuf_z;
-        if (use_src2) {
-            subbuf_z = { d_Z, z_buf_offset, z_sz };
-        } else {
-            subbuf_z = { d_X, 0, x_sz };
-        }
-
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        vk_subbuffer subbuf1 = use_src1 ? src1_buf : src0_buf;
+        vk_subbuffer subbuf2 = use_src2 ? src2_buf : src0_buf;
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, subbuf1, subbuf2, dst_buf }, pc, elements);
     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
-        // Empty src2 is possible in rope, but the shader needs a buffer
-        vk_subbuffer subbuf_z, subbuf_w;
-        if (use_src2) {
-            subbuf_z = { d_Z, z_buf_offset, z_sz };
-        } else {
-            subbuf_z = { d_X, 0, x_sz };
-        }
-        if (use_src3) {
-            subbuf_w = { d_W, w_buf_offset, w_sz };
-        } else {
-            subbuf_w = { d_X, 0, x_sz };
-        }
-
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz }, subbuf_w }, pc, elements);
+        // Empty src2 and src3 is possible in rope, but the shader needs a buffer
+        vk_subbuffer subbuf2 = use_src2 ? src2_buf : src0_buf;
+        vk_subbuffer subbuf3 = use_src3 ? src3_buf : src0_buf;
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, subbuf2, dst_buf, subbuf3 }, pc, elements);
     } else if (op == GGML_OP_IM2COL || op == GGML_OP_IM2COL_3D) {
         if (ctx->device->shader_int64 && ctx->device->buffer_device_address) {
             // buffer device address path doesn't use dst buffer
-            d_sz = 1;
+            dst_buf.size = 1;
         }
         // im2col uses only src1 and dst buffers
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src1_buf, dst_buf }, pc, elements);
     } else if (op == GGML_OP_COUNT_EQUAL) {
         // count_equal assumes that destination buffer is initialized with zeroes
-        ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz);
+        ggml_vk_buffer_memset_async(subctx, dst_buf.buffer, dst_buf.offset, 0, dst_buf.size);
         ggml_vk_sync_buffers(ctx, subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, dst_buf }, pc, elements);
     } else if (op == GGML_OP_OPT_STEP_SGD) {
         // OPT_STEP_SGD works on src0, it does not need dst
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, src2_buf }, pc, elements);
     } else if (use_src3) {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_W, w_buf_offset, w_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, src2_buf, src3_buf, dst_buf }, pc, elements);
     } else if (use_src2) {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, src2_buf, dst_buf }, pc, elements);
     } else if (use_src1) {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, src1_buf, dst_buf }, pc, elements);
     } else {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, dst_buf }, pc, elements);
     }
 }
 
-static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -8983,10 +9282,10 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, 0,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -9003,10 +9302,10 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, offset,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
+static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx) {
     const ggml_tensor *first_node = cgraph->nodes[node_idx];
     const ggml_tensor *dst = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
 
@@ -9051,10 +9350,7 @@ static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx,
         GGML_ABORT("fatal error");
     }
 
-    if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
-        return;
-    }
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
     ggml_backend_vk_buffer_context * buf_ctx[MAX_PARAMETER_COUNT];
     vk_buffer buf[MAX_PARAMETER_COUNT];
@@ -9116,7 +9412,7 @@ static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx,
         }, pc, elements);
 }
 
-static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -9128,10 +9424,10 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, ctx->do_add_rms_partials,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -9143,10 +9439,10 @@ static void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, 0,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -9158,10 +9454,10 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, 0,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -9173,10 +9469,10 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, 0,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_add_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_add_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t src2_type_size = ggml_type_size(src2->type);
@@ -9188,10 +9484,10 @@ static void ggml_vk_add_id(ggml_backend_vk_context * ctx, vk_context& subctx, co
         (uint32_t)src0->nb[2] / src0_type_size,
         (uint32_t)src1->nb[1] / src1_type_size,
         (uint32_t)src2->nb[1] / src2_type_size,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, int version, bool dryrun = false) {
+static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, int version) {
     GGML_ASSERT(version == 6 || version == 7);
     int num_srcs = version == 6 ? 6 : 7;
 
@@ -9204,44 +9500,12 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
     vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, dst->src[0], dst->src[1], dst->src[2], dst, dst->op);
     GGML_ASSERT(pipeline != nullptr);
 
-    if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
-        return;
-    }
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src_buf_ctxs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr };
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+    vk_subbuffer src_buf[7] = {};
     for (int i = 0; i < num_srcs; i++) {
-        src_buf_ctxs[i] = (ggml_backend_vk_buffer_context *)dst->src[i]->buffer->context;
-    }
-
-    vk_buffer d_D = nullptr, d_srcs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr };
-    size_t dst_offset = 0, src_offsets[7] = { 0, 0, 0, 0, 0, 0, 0 };
-    bool dst_uma = false, srcs_uma[7] = { false, false, false, false, false, false, false };
-
-    if (ctx->device->uma) {
-        for (int i = 0; i < num_srcs; i++) {
-            ggml_vk_host_get(ctx->device, dst->src[i]->data, d_srcs[i], src_offsets[i]);
-            srcs_uma[i] = d_srcs[i] != nullptr;
-        }
-
-        ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset);
-        dst_uma = d_D != nullptr;
-    }
-
-    uint64_t src_sizes[7] = { 0, 0, 0, 0, 0, 0, 0 };
-    for (int i = 0; i < num_srcs; i++) {
-        src_sizes[i] = ggml_nbytes(dst->src[i]);
-        if (!srcs_uma[i]) {
-            d_srcs[i] = src_buf_ctxs[i]->dev_buffer;
-            src_offsets[i] = vk_tensor_offset(dst->src[i]) + dst->src[i]->view_offs;
-        }
-    }
-
-    const uint64_t dst_size = ggml_nbytes(dst);
-    if (!dst_uma) {
-        d_D = dst_buf_ctx->dev_buffer;
-        dst_offset = vk_tensor_offset(dst) + dst->view_offs;
+        src_buf[i] = ggml_vk_tensor_subbuffer(ctx, dst->src[i]);
     }
 
     std::array<uint32_t, 3> elements = {
@@ -9251,33 +9515,20 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
     };
 
     if (version == 6) {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
-            vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
-            vk_subbuffer{ d_srcs[1], src_offsets[1], src_sizes[1] },
-            vk_subbuffer{ d_srcs[2], src_offsets[2], src_sizes[2] },
-            vk_subbuffer{ d_srcs[3], src_offsets[3], src_sizes[3] },
-            vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] },
-            vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
-            vk_subbuffer{ d_D, dst_offset, dst_size }
-        }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+            {src_buf[0], src_buf[1], src_buf[2], src_buf[3], src_buf[4], src_buf[5], dst_buf},
+            pc, elements);
     } else if (version == 7) {
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
-            vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
-            vk_subbuffer{ d_srcs[1], src_offsets[1], src_sizes[1] },
-            vk_subbuffer{ d_srcs[2], src_offsets[2], src_sizes[2] },
-            vk_subbuffer{ d_srcs[3], src_offsets[3], src_sizes[3] },
-            vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] },
-            vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
-            vk_subbuffer{ d_srcs[6], src_offsets[6], src_sizes[6] },
-            vk_subbuffer{ d_D, dst_offset, dst_size }
-        }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+            {src_buf[0], src_buf[1], src_buf[2], src_buf[3], src_buf[4], src_buf[5], src_buf[6], dst_buf},
+            pc, elements);
     } else {
         // shouldn't happen
         GGML_ASSERT(false);
     }
 }
 
-static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
     const size_t seq_length = dst->src[0]->ne[2];
     const size_t n_embed = dst->ne[0];
     const size_t n_heads = dst->src[0]->ne[1];
@@ -9291,12 +9542,11 @@ static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx,
             (uint32_t)n_embed,
             (uint32_t)n_heads,
         },
-        6,
-        dryrun
+        6
     );
 }
 
-static void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
     const size_t seq_length = dst->src[0]->ne[2];
     const size_t n_embed = dst->ne[0];
     const size_t n_heads = dst->src[0]->ne[1];
@@ -9310,12 +9560,11 @@ static void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx,
             (uint32_t)n_embed,
             (uint32_t)n_heads,
         },
-        7,
-        dryrun
+        7
     );
 }
 
-static void ggml_vk_ssm_scan(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_ssm_scan(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * src2 = dst->src[2];
@@ -9337,10 +9586,7 @@ static void ggml_vk_ssm_scan(ggml_backend_vk_context * ctx, vk_context& subctx,
     vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, dst->op);
     GGML_ASSERT(pipeline != nullptr);
 
-    if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
-        return;
-    }
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
     const int64_t s_off = ggml_nelements(src1) * sizeof(float);
 
@@ -9355,40 +9601,10 @@ static void ggml_vk_ssm_scan(ggml_backend_vk_context * ctx, vk_context& subctx,
         n_head, head_dim, n_group, n_tok
     };
 
-    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * src_buf_ctxs[GGML_MAX_SRC];
-    for (int i = 0; i < GGML_MAX_SRC && dst->src[i] != nullptr; i++) {
-        src_buf_ctxs[i] = (ggml_backend_vk_buffer_context *)dst->src[i]->buffer->context;
-    }
-
-    vk_buffer d_D = nullptr, d_srcs[GGML_MAX_SRC] = { nullptr };
-    size_t dst_offset = 0, src_offsets[GGML_MAX_SRC] = { 0 };
-    bool dst_uma = false, srcs_uma[GGML_MAX_SRC] = { false };
-
-    if (ctx->device->uma) {
-        for (int i = 0; i < GGML_MAX_SRC && dst->src[i] != nullptr; i++) {
-            ggml_vk_host_get(ctx->device, dst->src[i]->data, d_srcs[i], src_offsets[i]);
-            srcs_uma[i] = d_srcs[i] != nullptr;
-        }
-        ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset);
-        dst_uma = d_D != nullptr;
-    }
-
-    if (!dst_uma) {
-        d_D = dst_buf_ctx->dev_buffer;
-        dst_offset = vk_tensor_offset(dst) + dst->view_offs;
-    }
-    for (int i = 0; i < GGML_MAX_SRC && dst->src[i] != nullptr; i++) {
-        if (!srcs_uma[i]) {
-            d_srcs[i] = src_buf_ctxs[i]->dev_buffer;
-            src_offsets[i] = vk_tensor_offset(dst->src[i]) + dst->src[i]->view_offs;
-        }
-    }
-
-    size_t dst_size = ggml_nbytes(dst);
-    size_t src_sizes[GGML_MAX_SRC];
-    for (int i = 0; i < GGML_MAX_SRC && dst->src[i] != nullptr; i++) {
-        src_sizes[i] = ggml_nbytes(dst->src[i]);
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+    vk_subbuffer src_buf[7] = {};
+    for (int i = 0; i < 7 && dst->src[i] != nullptr; i++) {
+        src_buf[i] = ggml_vk_tensor_subbuffer(ctx, dst->src[i]);
     }
 
     std::array<uint32_t, 3> elements;
@@ -9398,19 +9614,12 @@ static void ggml_vk_ssm_scan(ggml_backend_vk_context * ctx, vk_context& subctx,
     const uint32_t num_workgroups_y = n_seq;
     elements = { num_workgroups_x, num_workgroups_y, 1 };
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
-        vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
-        vk_subbuffer{ d_srcs[1], src_offsets[1], src_sizes[1] },
-        vk_subbuffer{ d_srcs[2], src_offsets[2], src_sizes[2] },
-        vk_subbuffer{ d_srcs[3], src_offsets[3], src_sizes[3] },
-        vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] },
-        vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
-        vk_subbuffer{ d_srcs[6], src_offsets[6], src_sizes[6] },
-        vk_subbuffer{ d_D, dst_offset, dst_size }
-    }, pc, elements);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {src_buf[0], src_buf[1], src_buf[2], src_buf[3], src_buf[4], src_buf[5], src_buf[6], dst_buf},
+        pc, elements);
 }
 
-static void ggml_vk_ssm_conv(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_ssm_conv(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
@@ -9423,10 +9632,10 @@ static void ggml_vk_ssm_conv(ggml_backend_vk_context * ctx, vk_context& subctx,
         (uint32_t)src0->ne[1],
         (uint32_t)dst->ne[1],
         (uint32_t)dst->ne[2],
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_push_constants&& pc, bool dryrun = false) {
+static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_push_constants&& pc) {
     const ggml_tensor * x = dst->src[0];
     const ggml_tensor * g = dst->src[1];
     const ggml_tensor * gm = dst->src[2];
@@ -9452,90 +9661,37 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
     vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, g, gm, gv, dst, GGML_OP_OPT_STEP_ADAMW);
     GGML_ASSERT(pipeline != nullptr);
 
-    if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
-        return;
-    }
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
-    ggml_backend_vk_buffer_context * x_buf_ctx = (ggml_backend_vk_buffer_context *)x->buffer->context;
-    ggml_backend_vk_buffer_context * g_buf_ctx = (ggml_backend_vk_buffer_context *)g->buffer->context;
-    ggml_backend_vk_buffer_context * gm_buf_ctx = (ggml_backend_vk_buffer_context *)gm->buffer->context;
-    ggml_backend_vk_buffer_context * gv_buf_ctx = (ggml_backend_vk_buffer_context *)gv->buffer->context;
-    ggml_backend_vk_buffer_context * p_buf_ctx = (ggml_backend_vk_buffer_context *)p->buffer->context;
-
-    vk_buffer d_X = nullptr, d_G = nullptr, d_GM = nullptr, d_GV = nullptr, d_P = nullptr;
-    size_t x_offset = 0, g_offset = 0, gm_offset = 0, gv_offset = 0, p_offset = 0;
-    bool X_uma = false, G_uma = false, GM_uma = false, GV_uma = false, P_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, x->data, d_X, x_offset);
-        ggml_vk_host_get(ctx->device, g->data, d_G, g_offset);
-        ggml_vk_host_get(ctx->device, gm->data, d_GM, gm_offset);
-        ggml_vk_host_get(ctx->device, gv->data, d_GV, gv_offset);
-        ggml_vk_host_get(ctx->device, p->data, d_P, p_offset);
-
-        X_uma = d_X != nullptr;
-        G_uma = d_G != nullptr;
-        GM_uma = d_GM != nullptr;
-        GV_uma = d_GV != nullptr;
-        P_uma = d_P != nullptr;
-    }
-
-    if (!X_uma) {
-        d_X = x_buf_ctx->dev_buffer;
-        x_offset = vk_tensor_offset(x) + x->view_offs;
-    }
-    if (!G_uma) {
-        d_G = g_buf_ctx->dev_buffer;
-        g_offset = vk_tensor_offset(g) + g->view_offs;
-    }
-    if (!GM_uma) {
-        d_GM = gm_buf_ctx->dev_buffer;
-        gm_offset = vk_tensor_offset(gm) + gm->view_offs;
-    }
-    if (!GV_uma) {
-        d_GV = gv_buf_ctx->dev_buffer;
-        gv_offset = vk_tensor_offset(gv) + gv->view_offs;
-    }
-    if (!P_uma) {
-        d_P = p_buf_ctx->dev_buffer;
-        p_offset = vk_tensor_offset(p) + p->view_offs;
-    }
-
-    const uint64_t x_size = ggml_nbytes(x);
-    const uint64_t g_size = ggml_nbytes(g);
-    const uint64_t gm_size = ggml_nbytes(gm);
-    const uint64_t gv_size = ggml_nbytes(gv);
-    const uint64_t p_size = ggml_nbytes(p);
+    vk_subbuffer x_buf = ggml_vk_tensor_subbuffer(ctx, x);
+    vk_subbuffer g_buf = ggml_vk_tensor_subbuffer(ctx, g);
+    vk_subbuffer gm_buf = ggml_vk_tensor_subbuffer(ctx, gm);
+    vk_subbuffer gv_buf = ggml_vk_tensor_subbuffer(ctx, gv);
+    vk_subbuffer p_buf = ggml_vk_tensor_subbuffer(ctx, p);
 
     std::array<uint32_t, 3> elements = { (uint32_t)ggml_nelements(x), 1, 1 };
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
-        vk_subbuffer{ d_X, x_offset, x_size },
-        vk_subbuffer{ d_G, g_offset, g_size },
-        vk_subbuffer{ d_GM, gm_offset, gm_size },
-        vk_subbuffer{ d_GV, gv_offset, gv_size },
-        vk_subbuffer{ d_P, p_offset, p_size },
-    }, pc, elements);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {x_buf, g_buf, gm_buf, gv_buf, p_buf},
+        pc, elements);
 }
 
-static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
     const size_t n = ggml_nelements(dst->src[0]);
 
     ggml_vk_op_f32_opt_step_adamw(
         ctx, subctx, dst,
-        { (uint32_t)n, 0, 0.0f, 0.0f },
-        dryrun
+        { (uint32_t)n, 0, 0.0f, 0.0f }
     );
 }
 
-static void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     const size_t n = ggml_nelements(dst->src[0]);
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f });
 }
 
-static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     int * op_params = (int *)dst->op_params;
 
     const uint32_t src0_type_size = ggml_type_size(src0->type);
@@ -9549,70 +9705,142 @@ static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, co
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, op_params[0],
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t mode = (uint32_t)ggml_get_op_params_i32(dst, 0);
 
-    float sf0 = (float)dst->ne[0] / src0->ne[0];
-    float sf1 = (float)dst->ne[1] / src0->ne[1];
-    float sf2 = (float)dst->ne[2] / src0->ne[2];
-    float sf3 = (float)dst->ne[3] / src0->ne[3];
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float sf0 = (float)ne0 / ne00;
+    float sf1 = (float)ne1 / ne01;
+    float sf2 = (float)ne2 / ne02;
+    float sf3 = (float)ne3 / ne03;
+    float pixel_offset = 0.5f;
 
     if (mode & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-        sf0 = (float)(dst->ne[0] - 1) / (src0->ne[0] - 1);
-        sf1 = (float)(dst->ne[1] - 1) / (src0->ne[1] - 1);
+        sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+        sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+        pixel_offset = 0.0f;
     }
 
     ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
         (uint32_t)ggml_nelements(dst), 0, 0,
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1],
-        (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3],
-        sf0, sf1, sf2, sf3,
-    }, dryrun);
+        (uint32_t)ne00, (uint32_t)ne01,
+        (uint32_t)nb00 / src0_type_size, (uint32_t)nb01 / src0_type_size, (uint32_t)nb02 / src0_type_size, (uint32_t)nb03 / src0_type_size,
+        (uint32_t)ne0, (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3,
+        sf0, sf1, sf2, sf3, pixel_offset
+    });
 }
 
-static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
     p.param1 = ggml_get_op_params_f32(dst, 0);
     p.param2 = ggml_get_op_params_f32(dst, 1);
 
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SCALE, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SCALE, std::move(p));
 }
 
-static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
+static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst));
 }
 
-static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
+static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst));
 }
 
-static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
+static void ggml_vk_add1(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ADD1, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    });
 }
 
-static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_COS, vk_op_unary_push_constants_init(src0, dst), dryrun);
+static void ggml_vk_arange(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
+    VK_LOG_DEBUG("ggml_vk_arange(dst=" << dst << ", ne=" << ggml_nelements(dst) << ")");
+
+    vk_op_push_constants pc = {
+        (uint32_t)ggml_nelements(dst),
+        1,
+        ggml_get_op_params_f32(dst, 0),
+        ggml_get_op_params_f32(dst, 2),
+    };
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, nullptr, nullptr, nullptr, dst, GGML_OP_ARANGE);
+    GGML_ASSERT(pipeline != nullptr);
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, false);
+
+    std::array<uint32_t, 3> elements = { (uint32_t)ggml_nelements(dst), 1, 1 };
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { dst_buf }, pc, elements);
 }
 
-static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_fill(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
+    VK_LOG_DEBUG("ggml_vk_fill(dst=" << dst << ", ne=" << ggml_nelements(dst) << ")");
+
+    vk_op_push_constants pc = {
+        (uint32_t)ggml_nelements(dst),
+        1,
+        ggml_get_op_params_f32(dst, 0),
+        0.0f,
+    };
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, nullptr, nullptr, nullptr, dst, GGML_OP_FILL);
+    GGML_ASSERT(pipeline != nullptr);
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, false);
+
+    std::array<uint32_t, 3> elements = { (uint32_t)ggml_nelements(dst), 1, 1 };
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { dst_buf }, pc, elements);
+}
+
+static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst));
+}
+
+static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_COS, vk_op_unary_push_constants_init(src0, dst));
+}
+
+static void ggml_vk_log(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LOG, vk_op_unary_push_constants_init(src0, dst));
+}
+
+static void ggml_vk_tri(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = ggml_get_op_params_f32(dst, 0);
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_TRI, std::move(p));
+}
+
+static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
     p.param1 = ggml_get_op_params_f32(dst, 0);
     p.param2 = ggml_get_op_params_f32(dst, 1);
 
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CLAMP, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CLAMP, std::move(p));
 }
 
-static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_pad_push_constants p = vk_op_pad_push_constants_init(src0, dst);
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p));
 }
 
-static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const int32_t s0 = ggml_get_op_params_i32(dst, 0);
     const int32_t s1 = ggml_get_op_params_i32(dst, 1);
     const int32_t s2 = ggml_get_op_params_i32(dst, 2);
@@ -9624,20 +9852,20 @@ static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     memcpy(&p.param1, &s01_packed, sizeof(float));
     memcpy(&p.param2, &s23_packed, sizeof(float));
 
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ROLL, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ROLL, std::move(p));
 }
 
-static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT, std::move(p));
 }
 
-static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, std::move(p));
 }
 
-static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     uint32_t ne = (uint32_t)ggml_nelements(src0);
     if (ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
         // Convert from number of logical elements to 2- or 4-byte units.
@@ -9650,10 +9878,10 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }
 
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ne);
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CPY, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CPY, std::move(p));
 }
 
-static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -9672,20 +9900,20 @@ static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f, 0,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
 }
 
-static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
-static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const int * int_op_params = (const int *)dst->op_params;
     const float * float_op_params = (const float *)dst->op_params;
 
@@ -9693,7 +9921,7 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
     const float eps = float_op_params[1];
     const uint32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f });
 }
 
 static uint32_t ggml_vk_rms_num_partials(ggml_backend_vk_context * ctx, const ggml_tensor *node) {
@@ -9709,43 +9937,172 @@ static uint32_t ggml_vk_rms_partials_size(ggml_backend_vk_context * ctx, const g
     return num_bytes;
 }
 
-static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, float * op_params, bool dryrun = false) {
+static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *dst, const ggml_tensor *src0, const bool has_ff, bool backprop, const uint32_t set_rows_stride) {
+    const int n_dims        = ((const int32_t *) dst->op_params)[1];
+    const int mode          = ((const int32_t *) dst->op_params)[2];
+    // const int n_ctx         = ((const int32_t *) dst->op_params)[3];
+    const int n_ctx_orig    = ((const int32_t *) dst->op_params)[4];
+    const float freq_base   = ((const float *)   dst->op_params)[5];
+    const float freq_scale  = ((const float *)   dst->op_params)[6];
+    const float ext_factor  = ((const float *)   dst->op_params)[7];
+    const float attn_factor = ((const float *)   dst->op_params)[8];
+    const float beta_fast   = ((const float *)   dst->op_params)[9];
+    const float beta_slow   = ((const float *)   dst->op_params)[10];
+    int sections[4] {};
+    if (mode & GGML_ROPE_TYPE_MROPE) {
+        memcpy(sections, (const int32_t *) dst->op_params + 11, sizeof(int)*4);
+    }
+
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
+
+    float corr_dims[2];
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    uint32_t nb01 = src0->nb[1] / ggml_type_size(src0->type);
+    uint32_t nb02 = src0->nb[2] / ggml_type_size(src0->type);
+
+    vk_op_rope_push_constants rope {
+        (uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
+        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
+        has_ff, (uint32_t)src0->ne[2], nb01, nb02,
+        { sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,
+    };
+
+    return rope;
+}
+
+static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx, float * op_params) {
+    ggml_tensor * dst;
+    const ggml_tensor * src0;
+    const ggml_tensor * src1;
+
+    if (ctx->num_additional_fused_ops > 0) {
+        // fused rms_norm + mul
+        ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+        ggml_tensor *other_src = mul->src[0] == cgraph->nodes[node_idx + 0] ? mul->src[1] : mul->src[0];
+        dst = mul;
+        src0 = cgraph->nodes[node_idx]->src[0];
+        src1 = other_src;
+    } else {
+        dst = cgraph->nodes[node_idx];
+        src0 = src1 = dst->src[0];
+    }
+
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
     uint32_t param3 = ctx->do_add_rms_partials ? ggml_vk_rms_num_partials(ctx, dst) : 0;
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM, {
+    vk_op_binary_push_constants bin {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         op_params[0], 0.0f, (int32_t)param3,
-    }, dryrun);
+    };
 
-    if (ctx->do_add_rms_partials) {
+    // more than one fused op means rms_norm+mul+rope
+    if (ctx->num_additional_fused_ops > 1) {
+        static constexpr uint32_t max_tensors = 7;
+        const ggml_tensor *tensors[max_tensors] {};
+
+        ggml_tensor *rms = cgraph->nodes[node_idx + 0];
+        ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+        ggml_tensor *rope = cgraph->nodes[node_idx + 2];
+
+        ggml_tensor *other_src = mul->src[0] == rms ? mul->src[1] : mul->src[0];
+
+        bool do_set_rows = ctx->num_additional_fused_ops == 4;
+
+        tensors[0] = rms->src[0];
+        tensors[1] = other_src;
+        tensors[2] = mul;
+        tensors[3] = rope->src[1]; // pos
+        tensors[4] = rope->src[2]; // ff
+        tensors[5] = cgraph->nodes[node_idx + ctx->num_additional_fused_ops]; // dst
+        tensors[6] = do_set_rows ? tensors[5]->src[1] : nullptr;
+        const uint32_t set_rows_stride = do_set_rows ? tensors[5]->nb[1] / ggml_type_size(tensors[5]->type) : 0;
+
+        vk_op_rms_norm_mul_rope_push_constants pc;
+        pc.bin = bin;
+        pc.rope = ggml_vk_make_rope_constants(rope, rope->src[0], tensors[4] != nullptr, false, set_rows_stride);
+
+        vk_pipeline pipeline = tensors[5]->type == GGML_TYPE_F16 ? ctx->device->pipeline_rms_norm_mul_rope_f32_f16 : ctx->device->pipeline_rms_norm_mul_rope_f32_f32;
+
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+        ggml_backend_vk_buffer_context * buf_ctx[max_tensors];
+        vk_buffer buf[max_tensors];
+        size_t offset[max_tensors];
+        bool uma[max_tensors];
+
+        for (uint32_t i = 0; i < max_tensors; ++i) {
+            if (!tensors[i]) {
+                // If any remaining descriptors are unused, just point them at src[0]
+                buf[i] = buf[0];
+                offset[i] = 0;
+                continue;
+            }
+            buf_ctx[i] = (ggml_backend_vk_buffer_context *)tensors[i]->buffer->context;
+            buf[i] = nullptr;
+            offset[i] = 0;
+            uma[i] = false;
+
+            if (ctx->device->uma) {
+                ggml_vk_host_get(ctx->device, tensors[i]->data, buf[i], offset[i]);
+                uma[i] = buf[i] != nullptr;
+            }
+            if (!uma[i]) {
+                buf[i] = buf_ctx[i]->dev_buffer;
+                offset[i] = vk_tensor_offset(tensors[i]) + tensors[i]->view_offs;
+            }
+            GGML_ASSERT(buf[i] != nullptr);
+        }
+
+        std::array<uint32_t, 3> elements;
+        elements = { (uint32_t)rms->src[0]->ne[1], (uint32_t)rms->src[0]->ne[2], (uint32_t)rms->src[0]->ne[3] };
+
+        static_assert(max_tensors == 7);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+            {
+                ggml_vk_subbuffer(ctx, buf[0], offset[0]),
+                ggml_vk_subbuffer(ctx, buf[1], offset[1]),
+                ggml_vk_subbuffer(ctx, buf[2], offset[2]),
+                ggml_vk_subbuffer(ctx, buf[3], offset[3]),
+                ggml_vk_subbuffer(ctx, buf[4], offset[4]),
+                ggml_vk_subbuffer(ctx, buf[5], offset[5]),
+                ggml_vk_subbuffer(ctx, buf[6], offset[6]),
+            }, pc, elements);
+    } else {
+        ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM, std::move(bin));
+    }
+
+    if (ctx->do_add_rms_partials_offset_calculation) {
         ctx->prealloc_size_add_rms_partials_offset += ggml_vk_rms_partials_size(ctx, src0);
         ctx->do_add_rms_partials = false;
+        ctx->do_add_rms_partials_offset_calculation = false;
     }
 }
 
-static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
-static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
-static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
 }
 
-static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const float * op_params_f = (const float *)dst->op_params;
 
     const bool swapped = (bool)dst->op_params[1];
@@ -9773,15 +10130,15 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
             mode,
             alpha,
             limit
-        }, dryrun);
+        });
 }
 
-static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     int32_t * op_params = (int32_t *)dst->op_params;
-    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
+    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] });
 }
 
-static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
 
     float scale = op_params[0];
@@ -9814,16 +10171,15 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
         n_head_log2,
         nrows_x,
         src2 != nullptr
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), op_params[0], op_params[1] }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), op_params[0], op_params[1] });
 }
 
-static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
-
+static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx) {
     topk_moe_mode mode = ggml_vk_num_additional_ops_to_topk_moe_mode(ctx->num_additional_fused_ops);
     ggml_tensor * logits = cgraph->nodes[node_idx + 0]->src[0];
     ggml_tensor * weights = (mode == TOPK_MOE_EARLY_SOFTMAX_NORM) ? cgraph->nodes[node_idx + 9] :
@@ -9843,50 +10199,11 @@ static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx,
 
     vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, nullptr, nullptr, nullptr, cgraph->nodes[node_idx], GGML_OP_SOFT_MAX);
 
-    if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
-        return;
-    }
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
-    ggml_backend_vk_buffer_context * logits_buf_ctx = (ggml_backend_vk_buffer_context *)logits->buffer->context;
-    ggml_backend_vk_buffer_context * weights_buf_ctx = (ggml_backend_vk_buffer_context *)weights->buffer->context;
-    ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
-
-    vk_buffer d_logits = nullptr;
-    size_t logits_buf_offset = 0;
-    vk_buffer d_weights = nullptr;
-    size_t weights_buf_offset = 0;
-    vk_buffer d_ids = nullptr;
-    size_t ids_buf_offset = 0;
-
-    bool logits_uma = false;
-    bool weights_uma = false;
-    bool ids_uma = false;
-
-    if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, logits->data, d_logits, logits_buf_offset);
-        ggml_vk_host_get(ctx->device, weights->data, d_weights, weights_buf_offset);
-        ggml_vk_host_get(ctx->device, ids->data, d_ids, ids_buf_offset);
-        logits_uma = d_logits != nullptr;
-        weights_uma = d_weights != nullptr;
-        ids_uma = d_ids != nullptr;
-    }
-
-    if (!logits_uma) {
-        d_logits = logits_buf_ctx->dev_buffer;
-        logits_buf_offset = vk_tensor_offset(logits) + logits->view_offs;
-        GGML_ASSERT(d_logits != nullptr);
-    }
-    if (!weights_uma) {
-        d_weights = weights_buf_ctx->dev_buffer;
-        weights_buf_offset = vk_tensor_offset(weights) + weights->view_offs;
-        GGML_ASSERT(d_weights != nullptr);
-    }
-    if (!ids_uma) {
-        d_ids = ids_buf_ctx->dev_buffer;
-        ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs;
-        GGML_ASSERT(d_ids != nullptr);
-    }
+    vk_subbuffer logits_buf = ggml_vk_tensor_subbuffer(ctx, logits);
+    vk_subbuffer weights_buf = ggml_vk_tensor_subbuffer(ctx, weights);
+    vk_subbuffer ids_buf = ggml_vk_tensor_subbuffer(ctx, ids);
 
     vk_op_topk_moe_push_constants pc {};
     pc.n_rows = n_rows;
@@ -9902,15 +10219,10 @@ static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx,
     const uint32_t rows_per_block = 4;
     std::array<uint32_t, 3> elements = { CEIL_DIV(n_rows, rows_per_block), 1, 1 };
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
-        {
-            ggml_vk_subbuffer(ctx, d_logits, logits_buf_offset),
-            ggml_vk_subbuffer(ctx, d_weights, weights_buf_offset),
-            ggml_vk_subbuffer(ctx, d_ids, ids_buf_offset),
-        }, pc, elements);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {logits_buf, weights_buf, ids_buf}, pc, elements);
 }
 
-static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_cgraph * cgraph, int node_idx, bool backprop, bool dryrun = false) {
+static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_cgraph * cgraph, int node_idx, bool backprop) {
     ggml_tensor * dst = cgraph->nodes[node_idx];
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
@@ -9921,9 +10233,6 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     // const int n_ctx         = ((int32_t *) dst->op_params)[3];
     const int n_ctx_orig    = ((int32_t *) dst->op_params)[4];
     const float freq_base   = ((float *)   dst->op_params)[5];
-    const float freq_scale  = ((float *)   dst->op_params)[6];
-    const float ext_factor  = ((float *)   dst->op_params)[7];
-    const float attn_factor = ((float *)   dst->op_params)[8];
     const float beta_fast   = ((float *)   dst->op_params)[9];
     const float beta_slow   = ((float *)   dst->op_params)[10];
     int sections[4] {};
@@ -9934,11 +10243,6 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
-
-    uint32_t s1 = src0->nb[1] / ggml_type_size(src0->type);
-    uint32_t s2 = src0->nb[2] / ggml_type_size(src0->type);
-
     uint32_t set_rows_stride = 0;
     // Fused rope + view + set_rows passes the set_rows destination stride in set_rows_stride
     // and overrides the dst and sets src3=row_indices
@@ -9948,52 +10252,241 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         dst = cgraph->nodes[node_idx + 2];
     }
 
-    ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, src3, dst, GGML_OP_ROPE, {
-        (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
-        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
-        src2 != nullptr, (uint32_t)src0->ne[2], s1, s2,
-        { sections[0], sections[1], sections[2], sections[3] }, backprop, set_rows_stride,
-    }, dryrun);
+    ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, src3, dst, GGML_OP_ROPE,
+        ggml_vk_make_rope_constants(cgraph->nodes[node_idx], src0, src2 != nullptr, backprop, set_rows_stride));
 }
 
-static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    int32_t * op_params = (int32_t *)dst->op_params;
+static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t * op_params = (const uint32_t *)dst->op_params;
 
     uint32_t ncols = src0->ne[0];
     uint32_t nrows = ggml_nrows(src0);
 
-    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
-        ncols,
-        nrows,
-        op_params[0],
-    }, dryrun);
+    uint32_t ncols_pad_log2 = (uint32_t)ceilf(log2f(float(ncols)));
+    uint32_t ncolsp2 = 1 << ncols_pad_log2;
+
+    vk_op_argsort_push_constants pc { ncols, ncolsp2, ncols_pad_log2, nrows, op_params[0], 0, 0, 0, 0, };
+
+    // Pick the largest workgroup size <= ncolsp2
+    uint32_t pipeline_idx = std::min(ncols_pad_log2, num_argsort_pipelines - 1);
+
+    // Use the "small" argsort shader if the whole sort can be done by a single workgroup.
+    bool use_small = ncols_pad_log2 <= ctx->device->max_workgroup_size_log2 &&
+                     ctx->device->pipeline_argsort_f32[pipeline_idx] != nullptr;
+
+    vk_pipeline pipeline = use_small ? ctx->device->pipeline_argsort_f32[pipeline_idx]
+                                     : ctx->device->pipeline_argsort_large_f32[pipeline_idx];
+
+    vk_subbuffer src0_buf = ggml_vk_tensor_subbuffer(ctx, src0);
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+    vk_subbuffer subbuf1 = dst_buf;
+
+    // Reserve space for ivec2 per element, with rows padded to a power of two
+    if (!use_small) {
+        const size_t x_sz = size_t{ncolsp2} * nrows * 2 * sizeof(int);
+
+        if (ctx->prealloc_size_x < x_sz) {
+            ctx->prealloc_size_x = x_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
+        }
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+        subbuf1 = { ctx->prealloc_x, 0, ctx->prealloc_x->size };
+    }
+
+    std::array<uint32_t, 3> elements;
+
+    elements[0] = ncolsp2;
+    elements[1] = std::min((uint32_t)ggml_nrows(src0), ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+    elements[2] = 1;
+
+    // First dispatch initializes tmp_idx and does the first N passes where
+    // there is only communication between threads in the same workgroup.
+    {
+        vk_op_argsort_push_constants pc2 = pc;
+        pc2.outer_start = 0;
+        pc2.outer_end = std::min(ncols_pad_log2, ctx->device->max_workgroup_size_log2);
+        pc2.inner_start = 0;
+        pc2.inner_end = 100;
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, subbuf1, dst_buf }, pc2, elements);
+    }
+    if (!use_small) {
+        ggml_vk_sync_buffers(ctx, subctx);
+        // Loop over outer/inner passes, synchronizing between each pass.
+        for (uint32_t outer = ctx->device->max_workgroup_size_log2; outer < ncols_pad_log2; ++outer) {
+            for (uint32_t inner = 0; inner < outer + 1; ++inner) {
+                vk_op_argsort_push_constants pc2 = pc;
+                pc2.outer_start = outer;
+                pc2.outer_end = outer + 1;
+                pc2.inner_start = inner;
+                pc2.inner_end = inner + 1;
+                // When the inner idx is large enough, there's only communication
+                // within a workgroup. So the remaining inner iterations can all
+                // run in the same dispatch.
+                if (outer - inner < pipeline_idx) {
+                    pc2.inner_end = 100;
+                    inner = outer;
+                    pipeline = ctx->device->pipeline_argsort_large_f32[pipeline_idx];
+                } else {
+                    // Smaller workgroup empirically seems to perform better
+                    pipeline = ctx->device->pipeline_argsort_large_f32[pipeline_idx - 2];
+                }
+                ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+                ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src0_buf, subbuf1, dst_buf }, pc2, elements);
+                ggml_vk_sync_buffers(ctx, subctx);
+            }
+        }
+        ctx->prealloc_x_need_sync = true;
+    }
 }
 
-static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_topk(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    uint32_t ncols = src0->ne[0];
+    uint32_t nrows = ggml_nrows(src0);
+    uint32_t k = dst->ne[0];
+
+    vk_op_topk_push_constants pc { ncols, ncols, k, nrows, 0, 0 };
+
+    // Reserve space for ivec2 per element, double buffered
+    const size_t dbl_buf_size = size_t{ncols} * nrows * 2 * sizeof(int);
+    const size_t x_sz = dbl_buf_size * 2;
+    uint32_t dbl_buf_index = 0;
+
+    if (ctx->prealloc_size_x < x_sz) {
+        ctx->prealloc_size_x = x_sz;
+        ggml_vk_preallocate_buffers(ctx, subctx);
+    }
+    if (ctx->prealloc_x_need_sync) {
+        ggml_vk_sync_buffers(ctx, subctx);
+    }
+
+    std::array<uint32_t, 3> elements;
+    elements[1] = std::min(nrows, ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+    elements[2] = 1;
+
+    uint32_t num_elements = ncols;
+
+    // Each iteration reduces a workgroup's worth of elements down to the K
+    // largest elements. Repeat until we have the top K elements.
+    // Need to do at least one iteration to write out the results.
+    bool done_one_iter = false;
+    while (num_elements > k || !done_one_iter) {
+        done_one_iter = true;
+
+        // Prefer going as small as num_topk_pipelines - 3 for perf reasons.
+        // But if K is larger, then we need a larger workgroup
+        uint32_t max_pipeline = num_topk_pipelines - 1;
+        uint32_t preferred_pipeline = std::max(num_topk_pipelines - 3, (uint32_t)log2f(float(k)) + 2);
+        max_pipeline = std::min(preferred_pipeline, max_pipeline);
+        uint32_t min_pipeline = (uint32_t)log2f(float(k)) + 1;
+        // require full subgroup
+        min_pipeline = std::max(min_pipeline, ctx->device->subgroup_size_log2);
+
+        uint32_t pipeline_idx = (uint32_t)ceilf(log2f(float(num_elements)));
+        pipeline_idx = std::min(pipeline_idx, max_pipeline);
+        pipeline_idx = std::max(pipeline_idx, min_pipeline);
+
+        if (num_elements > (1u << pipeline_idx)) {
+            // If we could finish on this loop iteration (i.e. a single workgroup)
+            // then do so. It's better than the overhead of another pass.
+            for (uint32_t i = pipeline_idx; i < num_topk_pipelines; ++i) {
+                if (num_elements <= (1u << i)) {
+                    pipeline_idx = i;
+                    break;
+                }
+            }
+        }
+
+        vk_pipeline pipeline = ctx->device->pipeline_topk_f32[pipeline_idx];
+        // If the device doesn't support a pipeline this large, use smaller
+        while (!pipeline) {
+            pipeline_idx--;
+            GGML_ASSERT(pipeline_idx >= min_pipeline);
+            pipeline = ctx->device->pipeline_topk_f32[pipeline_idx];
+        }
+
+        vk_op_topk_push_constants pc2 = pc;
+        pc2.ncols_input = num_elements;
+
+        // Number of elements remaining after this pass
+        uint32_t num_dst_elements = (num_elements / pipeline->wg_denoms[0]) * k + std::min(k, num_elements % pipeline->wg_denoms[0]);
+
+        vk_subbuffer src_buf;
+        vk_subbuffer dst_buf;
+
+        if (num_elements == ncols) {
+            pc2.first_pass = 1;
+            src_buf = ggml_vk_tensor_subbuffer(ctx, src0);
+        } else {
+            src_buf = { ctx->prealloc_x, dbl_buf_index * dbl_buf_size, dbl_buf_size };
+        }
+        if (num_dst_elements == k) {
+            pc2.last_pass = 1;
+            dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+        } else {
+            dst_buf = { ctx->prealloc_x, (dbl_buf_index ^ 1) * dbl_buf_size, dbl_buf_size };
+        }
+
+        elements[0] = num_elements;
+
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src_buf, dst_buf }, pc2, elements);
+        num_elements = num_dst_elements;
+        dbl_buf_index ^= 1;
+        if (num_elements > k) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    ctx->prealloc_x_need_sync = true;
+}
+
+static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, ggml_nelements(src0));
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM, p, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM, p);
 }
 
-static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p);
 }
 
-static void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
     p.weight = 1.0f / (float)src0->ne[0];
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_MEAN, p, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_MEAN, p);
 }
 
-static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
+static void ggml_vk_cumsum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CUMSUM, p);
 }
 
-static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_COUNT_EQUAL, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f });
 }
 
-static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_COUNT_EQUAL, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
+}
+
+static void ggml_vk_solve_tri(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SOLVE_TRI, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    });
+}
+
+static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int32_t s0 = dst->op_params[0];
     const int32_t s1 = dst->op_params[1];
     const int32_t p0 = dst->op_params[2];
@@ -10030,10 +10523,10 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co
         pelements,
         IC * KH * KW,
         s0, s1, p0, p1, d0, d1,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_im2col_3d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_im2col_3d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
@@ -10096,20 +10589,20 @@ static void ggml_vk_im2col_3d(ggml_backend_vk_context * ctx, vk_context& subctx,
     pc.OH_OW_IC_KD_KH_KW = OH*OW*IC*KD*KH*KW;
     pc.OW_IC_KD_KH_KW = OW*IC*KD*KH*KW;
 
-    ggml_vk_op_f32<vk_op_im2col_3d_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_IM2COL_3D, std::move(pc), dryrun);
+    ggml_vk_op_f32<vk_op_im2col_3d_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_IM2COL_3D, std::move(pc));
 }
 
-static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const uint32_t dim = dst->op_params[0];
     const uint32_t max_period = dst->op_params[1];
     const uint32_t nb1 = dst->nb[1] / ggml_type_size(dst->type);
 
     ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, {
         nb1, dim, max_period,
-    }, dryrun);
+    });
 }
 
-static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     // src0: (K, Cout, Cin, 1) -- kernel
     // src1: (L, Cin, 1, 1) -- input
     // dst: (*, Cout, 1, 1)
@@ -10137,10 +10630,10 @@ static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context&
     p.nb1 = static_cast<uint32_t>(nb1 / nb0);
     p.s0 = static_cast<uint32_t>(s0);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p));
 }
 
-static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
     const int32_t k1 = dst->op_params[1];
     const int32_t k0 = dst->op_params[2];
@@ -10165,11 +10658,11 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
         parallel_elements,
         op,
         k0, k1, s0, s1, p0, p1,
-    }, dryrun);
+    });
 }
 
 static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
-                            const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+                            const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
@@ -10214,11 +10707,11 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
     GGML_ASSERT(ne03 == ne2);
     GGML_ASSERT(ne02 == ne12);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D, std::move(p));
 }
 
 static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
-                                      const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+                                      const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
@@ -10263,10 +10756,10 @@ static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context
     GGML_ASSERT(ne02 == ne2);
     GGML_ASSERT(ne03 == ne12);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p));
 }
 
-static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     vk_op_conv2d_dw_push_constants p{};
     p.ne = ggml_nelements(dst);
     p.channels = dst->ne[2];
@@ -10287,12 +10780,12 @@ static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_ASSERT(src0->ne[3] == p.channels);
     GGML_ASSERT(src1->ne[3] == p.batches);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p));
 }
 
-static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const float * op_params = (const float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f });
 }
 
 #ifdef GGML_VULKAN_RUN_TESTS
@@ -10451,10 +10944,6 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         }
     }
 
-    if (ctx->device->need_compiles) {
-        ggml_vk_load_shaders(ctx->device);
-    }
-
     ggml_pipeline_allocate_descriptor_sets(ctx);
 
     vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, {vk::MemoryPropertyFlagBits::eDeviceLocal});
@@ -10701,10 +11190,6 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 
     ggml_pipeline_request_descriptor_sets(ctx, p, 1);
 
-    if (ctx->device->need_compiles) {
-        ggml_vk_load_shaders(ctx->device);
-    }
-
     ggml_pipeline_allocate_descriptor_sets(ctx);
 
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
@@ -10802,10 +11287,6 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 //
 //     ggml_pipeline_request_descriptor_sets(ctx, p, 1);
 //
-//     if (ctx->device->need_compiles) {
-//         ggml_vk_load_shaders(ctx->device);
-//     }
-//
 //     ggml_pipeline_allocate_descriptor_sets(ctx);
 //
 //     ggml_vk_buffer_write(x_buf, 0, x, x_sz);
@@ -10976,10 +11457,6 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_quantize_q8_1, num_it);
     }
 
-    if (ctx->device->need_compiles) {
-        ggml_vk_load_shaders(ctx->device);
-    }
-
     ggml_pipeline_allocate_descriptor_sets(ctx);
 
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
@@ -11117,7 +11594,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
 }
 #endif
 
-static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
+static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_context subctx) {
 #if defined(GGML_VULKAN_RUN_TESTS)
     const std::vector<size_t> vals {
         512, 512, 128,
@@ -11207,6 +11684,15 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     GGML_ABORT("fatal error");
 #endif
 
+    if (subctx) {
+        // Submit and wait for any pending work before reallocating the buffers
+        ggml_vk_ctx_end(subctx);
+        ggml_vk_submit(subctx, {});
+        ctx->submit_pending = true;
+        ggml_vk_synchronize(ctx);
+        ggml_vk_ctx_begin(ctx->device, subctx);
+    }
+
     if (ctx->prealloc_x == nullptr || (ctx->prealloc_size_x > 0 && ctx->prealloc_x->size < ctx->prealloc_size_x)) {
         VK_LOG_MEMORY("ggml_vk_preallocate_buffers(x_size: " << ctx->prealloc_size_x << ")");
         // Resize buffer
@@ -11241,13 +11727,13 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     }
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool use_fence, bool almost_ready);
+static void ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool almost_ready);
 
 // Returns true if node has enqueued work into the queue, false otherwise
 // If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution.
-static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool almost_ready, bool submit){
+static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool last_node, bool almost_ready, bool submit){
     ggml_tensor * node = cgraph->nodes[node_idx];
-    if (ggml_is_empty(node) || !node->buffer) {
+    if (ggml_is_empty(node) || ggml_op_is_empty(node->op) || !node->buffer) {
         return false;
     }
 
@@ -11259,198 +11745,32 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     ggml_tensor * src2 = node->src[2];
     ggml_tensor * src3 = node->src[3];
 
-    switch (node->op) {
-    // Return on empty ops to avoid generating a compute_ctx and setting exit_tensor
-    case GGML_OP_RESHAPE:
-    case GGML_OP_VIEW:
-    case GGML_OP_PERMUTE:
-    case GGML_OP_TRANSPOSE:
-    case GGML_OP_NONE:
-        return false;
-    case GGML_OP_UNARY:
-        switch (ggml_get_unary_op(node)) {
-        case GGML_UNARY_OP_EXP:
-        case GGML_UNARY_OP_SILU:
-        case GGML_UNARY_OP_GELU:
-        case GGML_UNARY_OP_GELU_ERF:
-        case GGML_UNARY_OP_GELU_QUICK:
-        case GGML_UNARY_OP_RELU:
-        case GGML_UNARY_OP_TANH:
-        case GGML_UNARY_OP_SIGMOID:
-        case GGML_UNARY_OP_HARDSIGMOID:
-        case GGML_UNARY_OP_HARDSWISH:
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_GLU:
-        switch (ggml_get_glu_op(node)) {
-        case GGML_GLU_OP_GEGLU:
-        case GGML_GLU_OP_REGLU:
-        case GGML_GLU_OP_SWIGLU:
-        case GGML_GLU_OP_SWIGLU_OAI:
-        case GGML_GLU_OP_GEGLU_ERF:
-        case GGML_GLU_OP_GEGLU_QUICK:
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_ADD:
-        {
-            int next_node_idx = node_idx + 1 + ctx->num_additional_fused_ops;
-            if (next_node_idx < cgraph->n_nodes &&
-                cgraph->nodes[next_node_idx]->op == GGML_OP_RMS_NORM &&
-                cgraph->nodes[next_node_idx]->src[0] == cgraph->nodes[next_node_idx - 1] &&
-                ggml_nrows(cgraph->nodes[next_node_idx]) == 1 &&
-                ctx->device->add_rms_fusion) {
-                if (dryrun) {
-                    ctx->prealloc_size_add_rms_partials += ggml_vk_rms_partials_size(ctx, cgraph->nodes[node_idx]);
-                }
+    if (node->op == GGML_OP_ADD) {
+        int next_node_idx = node_idx + 1 + ctx->num_additional_fused_ops;
+        if (next_node_idx < cgraph->n_nodes &&
+            cgraph->nodes[next_node_idx]->op == GGML_OP_RMS_NORM &&
+            cgraph->nodes[next_node_idx]->src[0] == cgraph->nodes[next_node_idx - 1] &&
+            ggml_nrows(cgraph->nodes[next_node_idx]) == 1 &&
+            ctx->device->add_rms_fusion) {
+            uint32_t size = ggml_vk_rms_partials_size(ctx, cgraph->nodes[node_idx]);
+            ctx->do_add_rms_partials_offset_calculation = true;
+            if (ctx->prealloc_size_add_rms_partials_offset + size <= ctx->prealloc_size_add_rms_partials) {
                 ctx->do_add_rms_partials = true;
             }
-        } break;
-    case GGML_OP_REPEAT:
-    case GGML_OP_REPEAT_BACK:
-    case GGML_OP_GET_ROWS:
-    case GGML_OP_ADD_ID:
-    case GGML_OP_ACC:
-    case GGML_OP_SUB:
-    case GGML_OP_MUL:
-    case GGML_OP_DIV:
-    case GGML_OP_CONCAT:
-    case GGML_OP_UPSCALE:
-    case GGML_OP_SCALE:
-    case GGML_OP_SQR:
-    case GGML_OP_SQRT:
-    case GGML_OP_SIN:
-    case GGML_OP_COS:
-    case GGML_OP_CLAMP:
-    case GGML_OP_PAD:
-    case GGML_OP_ROLL:
-    case GGML_OP_CPY:
-    case GGML_OP_SET_ROWS:
-    case GGML_OP_CONT:
-    case GGML_OP_DUP:
-    case GGML_OP_SILU_BACK:
-    case GGML_OP_NORM:
-    case GGML_OP_GROUP_NORM:
-    case GGML_OP_RMS_NORM:
-    case GGML_OP_RMS_NORM_BACK:
-    case GGML_OP_L2_NORM:
-    case GGML_OP_DIAG_MASK_INF:
-    case GGML_OP_SOFT_MAX:
-    case GGML_OP_SOFT_MAX_BACK:
-    case GGML_OP_ROPE:
-    case GGML_OP_ROPE_BACK:
-    case GGML_OP_MUL_MAT:
-    case GGML_OP_MUL_MAT_ID:
-    case GGML_OP_ARGSORT:
-    case GGML_OP_SUM:
-    case GGML_OP_SUM_ROWS:
-    case GGML_OP_MEAN:
-    case GGML_OP_ARGMAX:
-    case GGML_OP_COUNT_EQUAL:
-    case GGML_OP_IM2COL:
-    case GGML_OP_IM2COL_3D:
-    case GGML_OP_TIMESTEP_EMBEDDING:
-    case GGML_OP_CONV_TRANSPOSE_1D:
-    case GGML_OP_POOL_2D:
-    case GGML_OP_CONV_2D:
-    case GGML_OP_CONV_TRANSPOSE_2D:
-    case GGML_OP_CONV_2D_DW:
-    case GGML_OP_RWKV_WKV6:
-    case GGML_OP_RWKV_WKV7:
-    case GGML_OP_SSM_SCAN:
-    case GGML_OP_SSM_CONV:
-    case GGML_OP_LEAKY_RELU:
-    case GGML_OP_FLASH_ATTN_EXT:
-    case GGML_OP_OPT_STEP_ADAMW:
-    case GGML_OP_OPT_STEP_SGD:
-        break;
-    default:
-        std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
-        GGML_ABORT("fatal error");
+        }
     }
 
     vk_context compute_ctx;
 
-    if (!dryrun) {
-        if (ctx->compute_ctx.expired()) {
-            compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-            ctx->compute_ctx = compute_ctx;
-            ggml_vk_ctx_begin(ctx->device, compute_ctx);
-        } else {
-            compute_ctx = ctx->compute_ctx.lock();
-        }
+    if (ctx->compute_ctx.expired()) {
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        switch (node->op) {
-        case GGML_OP_REPEAT:
-        case GGML_OP_REPEAT_BACK:
-        case GGML_OP_ACC:
-        case GGML_OP_GET_ROWS:
-        case GGML_OP_ADD:
-        case GGML_OP_SUB:
-        case GGML_OP_MUL:
-        case GGML_OP_DIV:
-        case GGML_OP_CONCAT:
-        case GGML_OP_UPSCALE:
-        case GGML_OP_SCALE:
-        case GGML_OP_SQR:
-        case GGML_OP_SQRT:
-        case GGML_OP_SIN:
-        case GGML_OP_COS:
-        case GGML_OP_CLAMP:
-        case GGML_OP_PAD:
-        case GGML_OP_CPY:
-        case GGML_OP_SET_ROWS:
-        case GGML_OP_CONT:
-        case GGML_OP_DUP:
-        case GGML_OP_SILU_BACK:
-        case GGML_OP_NORM:
-        case GGML_OP_GROUP_NORM:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_RMS_NORM_BACK:
-        case GGML_OP_L2_NORM:
-        case GGML_OP_UNARY:
-        case GGML_OP_GLU:
-        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_SOFT_MAX_BACK:
-        case GGML_OP_ROPE_BACK:
-        case GGML_OP_ARGSORT:
-        case GGML_OP_SUM:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_MEAN:
-        case GGML_OP_ARGMAX:
-        case GGML_OP_COUNT_EQUAL:
-        case GGML_OP_IM2COL:
-        case GGML_OP_IM2COL_3D:
-        case GGML_OP_TIMESTEP_EMBEDDING:
-        case GGML_OP_CONV_TRANSPOSE_1D:
-        case GGML_OP_POOL_2D:
-        case GGML_OP_CONV_2D:
-        case GGML_OP_CONV_TRANSPOSE_2D:
-        case GGML_OP_CONV_2D_DW:
-        case GGML_OP_LEAKY_RELU:
-        case GGML_OP_OPT_STEP_SGD:
-            {
-                // These operations all go through ggml_vk_op_f32, so short-circuit and
-                // do the only thing needed for the dryrun.
-                vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op);
-                ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
-                if (node->op == GGML_OP_RMS_NORM) {
-                    ctx->do_add_rms_partials = false;
-                }
-                return false;
-            }
-        default:
-            break;
-        }
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
-    if (!dryrun) {
+    {
         // This logic detects dependencies between modes in the graph and calls ggml_vk_sync_buffers
         // to synchronize them. This handles most "normal" synchronization when computing the graph, and when
         // there is no auxiliary memory use, it shouldn't be necessary to call ggml_vk_sync_buffers
@@ -11535,136 +11855,151 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         }
     }
 #if ENABLE_SYNC_LOGGING
-    if (!dryrun) {
-        for (int i = 0; i < ctx->num_additional_fused_ops + 1; ++i) {
-            auto *n = cgraph->nodes[node_idx + i];
-            std::cerr << node_idx + i << " " << ggml_op_name(n->op) << " " <<  n->name;
-            if (n->op == GGML_OP_GLU) {
-                std::cerr << " " << ggml_glu_op_name(ggml_get_glu_op(n)) << " " << (n->src[1] ? "split" : "single") << " ";
-            }
-            std::cerr << std::endl;
+    for (int i = 0; i < ctx->num_additional_fused_ops + 1; ++i) {
+        auto *n = cgraph->nodes[node_idx + i];
+        std::cerr << node_idx + i << " " << ggml_op_name(n->op) << " " <<  n->name;
+        if (n->op == GGML_OP_GLU) {
+            std::cerr << " " << ggml_glu_op_name(ggml_get_glu_op(n)) << " " << (n->src[1] ? "split" : "single") << " ";
         }
+        if (n->op == GGML_OP_ROPE) {
+            const int mode = ((const int32_t *) n->op_params)[2];
+            std::cerr << " rope mode: " << mode;
+        }
+        std::cerr << std::endl;
     }
 #endif
 
     switch (node->op) {
     case GGML_OP_REPEAT:
-        ggml_vk_repeat(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_repeat(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_REPEAT_BACK:
-        ggml_vk_repeat_back(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_repeat_back(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_ACC:
-        ggml_vk_acc(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_acc(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_GET_ROWS:
-        ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_ADD:
         if (ctx->num_additional_fused_ops) {
-            ggml_vk_multi_add(ctx, compute_ctx, cgraph, node_idx, dryrun);
+            ggml_vk_multi_add(ctx, compute_ctx, cgraph, node_idx);
         } else {
-            ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
+            ggml_vk_add(ctx, compute_ctx, src0, src1, node);
         }
         break;
     case GGML_OP_SUB:
-        ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_sub(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_MUL:
-        ggml_vk_mul(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_mul(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_DIV:
-        ggml_vk_div(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_div(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_ADD_ID:
-        ggml_vk_add_id(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+        ggml_vk_add_id(ctx, compute_ctx, src0, src1, src2, node);
 
         break;
     case GGML_OP_CONCAT:
-        ggml_vk_concat(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_concat(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_UPSCALE:
-        ggml_vk_upscale(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_upscale(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_ADD1:
+        ggml_vk_add1(ctx, compute_ctx, src0, src1, node);
+
+        break;
+    case GGML_OP_ARANGE:
+        ggml_vk_arange(ctx, compute_ctx, node);
+
+        break;
+    case GGML_OP_FILL:
+        ggml_vk_fill(ctx, compute_ctx, node);
 
         break;
     case GGML_OP_SCALE:
-        ggml_vk_scale(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_scale(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SQR:
-        ggml_vk_sqr(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_sqr(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SQRT:
-        ggml_vk_sqrt(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_sqrt(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SIN:
-        ggml_vk_sin(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_sin(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_COS:
-        ggml_vk_cos(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_cos(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_LOG:
+        ggml_vk_log(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_TRI:
+        ggml_vk_tri(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_CLAMP:
-        ggml_vk_clamp(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_clamp(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_PAD:
-        ggml_vk_pad(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_pad(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_ROLL:
-        ggml_vk_roll(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_roll(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
-        ggml_vk_cpy(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_cpy(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SET_ROWS:
-        ggml_vk_set_rows(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_set_rows(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_SILU_BACK:
-        ggml_vk_silu_back(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_silu_back(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_NORM:
-        ggml_vk_norm(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_norm(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_GROUP_NORM:
-        ggml_vk_group_norm(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_group_norm(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_RMS_NORM:
-        if (ctx->num_additional_fused_ops > 0) {
-            // fused rms_norm + mul
-            ggml_tensor *mul = cgraph->nodes[node_idx + 1];
-            ggml_tensor *other_src = mul->src[0] == node ? mul->src[1] : mul->src[0];
-            ggml_vk_rms_norm(ctx, compute_ctx, src0, other_src, mul, (float *)node->op_params, dryrun);
-        } else {
-            ggml_vk_rms_norm(ctx, compute_ctx, src0, src0, node, (float *)node->op_params, dryrun);
-        }
+        ggml_vk_rms_norm(ctx, compute_ctx, cgraph, node_idx, (float *)node->op_params);
         break;
     case GGML_OP_RMS_NORM_BACK:
-        ggml_vk_rms_norm_back(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_rms_norm_back(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_L2_NORM:
-        ggml_vk_l2_norm(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_l2_norm(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_UNARY:
@@ -11675,11 +12010,19 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_UNARY_OP_GELU_ERF:
         case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
+        case GGML_UNARY_OP_NEG:
         case GGML_UNARY_OP_TANH:
         case GGML_UNARY_OP_SIGMOID:
         case GGML_UNARY_OP_HARDSIGMOID:
         case GGML_UNARY_OP_HARDSWISH:
-            ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun);
+        case GGML_UNARY_OP_ABS:
+        case GGML_UNARY_OP_SOFTPLUS:
+        case GGML_UNARY_OP_STEP:
+        case GGML_UNARY_OP_ROUND:
+        case GGML_UNARY_OP_CEIL:
+        case GGML_UNARY_OP_FLOOR:
+        case GGML_UNARY_OP_TRUNC:
+            ggml_vk_unary(ctx, compute_ctx, src0, node);
             break;
         default:
             return false;
@@ -11693,151 +12036,159 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_GLU_OP_SWIGLU_OAI:
         case GGML_GLU_OP_GEGLU_ERF:
         case GGML_GLU_OP_GEGLU_QUICK:
-            ggml_vk_glu(ctx, compute_ctx, src0, src1, node, dryrun);
+            ggml_vk_glu(ctx, compute_ctx, src0, src1, node);
             break;
         default:
             return false;
         }
         break;
     case GGML_OP_DIAG_MASK_INF:
-        ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SOFT_MAX:
         if (ctx->num_additional_fused_ops) {
-            ggml_vk_topk_moe(ctx, compute_ctx, cgraph, node_idx, dryrun);
+            ggml_vk_topk_moe(ctx, compute_ctx, cgraph, node_idx);
         } else {
-            ggml_vk_soft_max(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+            ggml_vk_soft_max(ctx, compute_ctx, src0, src1, src2, node);
         }
 
         break;
     case GGML_OP_SOFT_MAX_BACK:
-        ggml_vk_soft_max_back(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_soft_max_back(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_ROPE:
-        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, false, dryrun);
+        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, false);
 
         break;
     case GGML_OP_ROPE_BACK:
-        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, true, dryrun);
+        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, true);
 
         break;
     case GGML_OP_ARGSORT:
         if (ctx->num_additional_fused_ops) {
-            ggml_vk_topk_moe(ctx, compute_ctx, cgraph, node_idx, dryrun);
+            ggml_vk_topk_moe(ctx, compute_ctx, cgraph, node_idx);
         } else {
-            ggml_vk_argsort(ctx, compute_ctx, src0, node, dryrun);
+            ggml_vk_argsort(ctx, compute_ctx, src0, node);
         }
 
+        break;
+    case GGML_OP_TOP_K:
+        ggml_vk_topk(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_SUM:
-        ggml_vk_sum(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_sum(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_SUM_ROWS:
-        ggml_vk_sum_rows(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_sum_rows(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_CUMSUM:
+        ggml_vk_cumsum(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_MEAN:
-        ggml_vk_mean(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_mean(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_ARGMAX:
-        ggml_vk_argmax(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_argmax(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_COUNT_EQUAL:
-        ggml_vk_count_equal(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_count_equal(ctx, compute_ctx, src0, src1, node);
+
+        break;
+    case GGML_OP_SOLVE_TRI:
+        ggml_vk_solve_tri(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_IM2COL:
-        ggml_vk_im2col(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_im2col(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_IM2COL_3D:
-        ggml_vk_im2col_3d(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_im2col_3d(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_TIMESTEP_EMBEDDING:
-        ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_CONV_TRANSPOSE_1D:
-        ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_POOL_2D:
-        ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_pool_2d(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_CONV_2D:
-        ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_CONV_TRANSPOSE_2D:
-        ggml_vk_conv_transpose_2d(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_conv_transpose_2d(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_CONV_2D_DW:
-        ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_LEAKY_RELU:
-        ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun);
+        ggml_vk_leaky_relu(ctx, compute_ctx, src0, node);
 
         break;
     case GGML_OP_MUL_MAT:
-        ggml_vk_mul_mat(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_mul_mat(ctx, compute_ctx, cgraph, node_idx);
 
         break;
     case GGML_OP_MUL_MAT_ID:
-        ggml_vk_mul_mat_id(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+        ggml_vk_mul_mat_id(ctx, compute_ctx, cgraph, node_idx);
 
         break;
 
     case GGML_OP_FLASH_ATTN_EXT:
-        ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node->src[4], node, dryrun);
+        ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node->src[4], node);
 
         break;
 
     case GGML_OP_RWKV_WKV6:
-        ggml_vk_rwkv_wkv6(ctx, compute_ctx, node, dryrun);
+        ggml_vk_rwkv_wkv6(ctx, compute_ctx, node);
 
         break;
 
     case GGML_OP_RWKV_WKV7:
-        ggml_vk_rwkv_wkv7(ctx, compute_ctx, node, dryrun);
+        ggml_vk_rwkv_wkv7(ctx, compute_ctx, node);
 
         break;
 
     case GGML_OP_SSM_SCAN:
-        ggml_vk_ssm_scan(ctx, compute_ctx, node, dryrun);
+        ggml_vk_ssm_scan(ctx, compute_ctx, node);
 
         break;
 
     case GGML_OP_SSM_CONV:
-        ggml_vk_ssm_conv(ctx, compute_ctx, node, dryrun);
+        ggml_vk_ssm_conv(ctx, compute_ctx, node);
 
         break;
 
     case GGML_OP_OPT_STEP_ADAMW:
-        ggml_vk_opt_step_adamw(ctx, compute_ctx, node, dryrun);
+        ggml_vk_opt_step_adamw(ctx, compute_ctx, node);
 
         break;
 
     case GGML_OP_OPT_STEP_SGD:
-        ggml_vk_opt_step_sgd(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+        ggml_vk_opt_step_sgd(ctx, compute_ctx, src0, src1, src2, node);
 
         break;
     default:
         return false;
     }
 
-    if (dryrun) {
-        return false;
-    }
-
     ctx->tensor_ctxs[node_idx] = compute_ctx;
 
 #if defined(GGML_VULKAN_CHECK_RESULTS)
@@ -11859,148 +12210,23 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         ctx->compute_ctx.reset();
 
-        bool ok = ggml_vk_compute_forward(ctx, cgraph, node_begin, node_idx_begin, false, almost_ready);
-        if (!ok) {
-            if (node->op == GGML_OP_UNARY) {
-                std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
-            } else if (node->op == GGML_OP_GLU) {
-                std::cerr << __func__ << ": error: op not supported GLU " << node->name << " (" << ggml_glu_op_name(static_cast<ggml_glu_op>(node->op_params[0])) << ")" << std::endl;
-            } else {
-                std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl;
-            }
-        }
-
+        ggml_vk_compute_forward(ctx, cgraph, node_begin, node_idx_begin, almost_ready);
     }
     return true;
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, ggml_tensor * tensor, int tensor_idx, bool use_fence = true, bool almost_ready = false) {
+static void ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, ggml_tensor * tensor, int tensor_idx, bool almost_ready = false) {
     GGML_UNUSED(cgraph);
-    ggml_backend_buffer * buf = nullptr;
-
-    switch (tensor->op) {
-    case GGML_OP_ADD:
-    case GGML_OP_ACC:
-    case GGML_OP_GET_ROWS:
-    case GGML_OP_SUB:
-    case GGML_OP_MUL:
-    case GGML_OP_DIV:
-    case GGML_OP_ADD_ID:
-    case GGML_OP_CONCAT:
-    case GGML_OP_UPSCALE:
-    case GGML_OP_SCALE:
-    case GGML_OP_SQR:
-    case GGML_OP_SQRT:
-    case GGML_OP_SIN:
-    case GGML_OP_COS:
-    case GGML_OP_CLAMP:
-    case GGML_OP_PAD:
-    case GGML_OP_ROLL:
-    case GGML_OP_CPY:
-    case GGML_OP_SET_ROWS:
-    case GGML_OP_CONT:
-    case GGML_OP_DUP:
-    case GGML_OP_SILU_BACK:
-    case GGML_OP_NORM:
-    case GGML_OP_GROUP_NORM:
-    case GGML_OP_RMS_NORM:
-    case GGML_OP_RMS_NORM_BACK:
-    case GGML_OP_L2_NORM:
-    case GGML_OP_DIAG_MASK_INF:
-    case GGML_OP_SOFT_MAX:
-    case GGML_OP_SOFT_MAX_BACK:
-    case GGML_OP_ROPE:
-    case GGML_OP_ROPE_BACK:
-    case GGML_OP_RESHAPE:
-    case GGML_OP_VIEW:
-    case GGML_OP_PERMUTE:
-    case GGML_OP_TRANSPOSE:
-    case GGML_OP_NONE:
-    case GGML_OP_ARGSORT:
-    case GGML_OP_SUM:
-    case GGML_OP_SUM_ROWS:
-    case GGML_OP_MEAN:
-    case GGML_OP_ARGMAX:
-    case GGML_OP_COUNT_EQUAL:
-    case GGML_OP_IM2COL:
-    case GGML_OP_IM2COL_3D:
-    case GGML_OP_TIMESTEP_EMBEDDING:
-    case GGML_OP_CONV_TRANSPOSE_1D:
-    case GGML_OP_POOL_2D:
-    case GGML_OP_CONV_2D:
-    case GGML_OP_CONV_TRANSPOSE_2D:
-    case GGML_OP_CONV_2D_DW:
-    case GGML_OP_RWKV_WKV6:
-    case GGML_OP_RWKV_WKV7:
-    case GGML_OP_SSM_SCAN:
-    case GGML_OP_SSM_CONV:
-    case GGML_OP_LEAKY_RELU:
-    case GGML_OP_REPEAT:
-    case GGML_OP_REPEAT_BACK:
-    case GGML_OP_OPT_STEP_ADAMW:
-    case GGML_OP_OPT_STEP_SGD:
-        buf = tensor->buffer;
-        break;
-    case GGML_OP_UNARY:
-        switch (ggml_get_unary_op(tensor)) {
-        case GGML_UNARY_OP_EXP:
-        case GGML_UNARY_OP_SILU:
-        case GGML_UNARY_OP_GELU:
-        case GGML_UNARY_OP_GELU_ERF:
-        case GGML_UNARY_OP_GELU_QUICK:
-        case GGML_UNARY_OP_RELU:
-        case GGML_UNARY_OP_TANH:
-        case GGML_UNARY_OP_SIGMOID:
-        case GGML_UNARY_OP_HARDSIGMOID:
-        case GGML_UNARY_OP_HARDSWISH:
-            buf = tensor->buffer;
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_GLU:
-        switch (ggml_get_glu_op(tensor)) {
-        case GGML_GLU_OP_GEGLU:
-        case GGML_GLU_OP_REGLU:
-        case GGML_GLU_OP_SWIGLU:
-        case GGML_GLU_OP_SWIGLU_OAI:
-        case GGML_GLU_OP_GEGLU_ERF:
-        case GGML_GLU_OP_GEGLU_QUICK:
-            buf = tensor->buffer;
-            break;
-        default:
-            return false;
-        }
-        break;
-    case GGML_OP_MUL_MAT:
-    case GGML_OP_MUL_MAT_ID:
-    case GGML_OP_FLASH_ATTN_EXT:
-        buf = tensor->buffer;
-
-        break;
-    default:
-        return false;
-    }
-
-    if (buf == nullptr) {
-        return false;
-    }
+    GGML_UNUSED(tensor);
 
     VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")");
 
     vk_context subctx = ctx->tensor_ctxs[tensor_idx].lock();
 
-    // always wait for the GPU work to be done for the last submit
-    if (tensor_idx == subctx->exit_tensor_idx) {
-        use_fence = true;
-    }
-
     // Only run if ctx hasn't been submitted yet
     if (!subctx->seqs.empty()) {
 #ifdef GGML_VULKAN_CHECK_RESULTS
         ggml_vk_check_results_0(ctx, cgraph, tensor_idx);
-        use_fence = true;
 #endif
 
         // Do staging buffer copies
@@ -12012,17 +12238,16 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
             memset(mset.dst, mset.val, mset.n);
         }
 
-        if (almost_ready && !ctx->almost_ready_fence_pending && !use_fence) {
+        if (almost_ready && !ctx->almost_ready_fence_pending) {
             ggml_vk_submit(subctx, ctx->almost_ready_fence);
             ctx->almost_ready_fence_pending = true;
         } else {
-            ggml_vk_submit(subctx, use_fence ? ctx->fence : vk::Fence{});
+            ggml_vk_submit(subctx, {});
         }
+        ctx->submit_pending = true;
 
-        if (use_fence) {
-            ggml_vk_wait_for_fence(ctx);
-        }
 #ifdef GGML_VULKAN_CHECK_RESULTS
+        ggml_vk_synchronize(ctx);
         ggml_vk_check_results_1(ctx, cgraph, tensor_idx);
 #endif
     }
@@ -12036,8 +12261,6 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
         subctx->out_memcpys.clear();
         subctx->memsets.clear();
     }
-
-    return true;
 }
 
 // Clean up after graph processing is done
@@ -12078,11 +12301,19 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
 // Clean up on backend free
 static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     VK_LOG_DEBUG("ggml_vk_cleanup(" << ctx->name << ")");
+    // discard any unsubmitted command buffers
+    ctx->transfer_ctx.reset();
+    // wait for any pending command buffers to finish
+    ggml_vk_synchronize(ctx);
+
     ggml_vk_graph_cleanup(ctx);
 
     ggml_vk_destroy_buffer(ctx->prealloc_x);
     ggml_vk_destroy_buffer(ctx->prealloc_y);
     ggml_vk_destroy_buffer(ctx->prealloc_split_k);
+    ggml_vk_destroy_buffer(ctx->prealloc_add_rms_partials);
+    ggml_vk_destroy_buffer(ctx->sync_staging);
+
     ctx->prealloc_y_last_pipeline_used = nullptr;
 
     ctx->prealloc_size_x = 0;
@@ -12402,7 +12633,7 @@ static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor
 
     if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
         ctx->transfer_ctx = transfer_ctx;
         ggml_vk_ctx_begin(ctx->device, transfer_ctx);
     } else {
@@ -12425,7 +12656,7 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
 
     if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
         ctx->transfer_ctx = transfer_ctx;
         ggml_vk_ctx_begin(ctx->device, transfer_ctx);
     } else {
@@ -12434,7 +12665,23 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
 
     vk_buffer buf = buf_ctx->dev_buffer;
 
-    ggml_vk_buffer_read_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
+    auto src_offset = vk_tensor_offset(tensor) + tensor->view_offs + offset;
+    bool ret = ggml_vk_buffer_read_async(transfer_ctx, buf, src_offset, data, size);
+
+    // If that failed, copy synchronously through a staging buffer
+    if (!ret) {
+        ggml_vk_ensure_sync_staging_buffer(ctx, size);
+        ggml_vk_sync_buffers(nullptr, transfer_ctx);
+
+        vk::BufferCopy buffer_cpy;
+        buffer_cpy.srcOffset = src_offset;
+        buffer_cpy.dstOffset = 0;
+        buffer_cpy.size = size;
+
+        transfer_ctx->s->buffer.copyBuffer(buf->buffer, ctx->sync_staging->buffer, { buffer_cpy });
+        deferred_memcpy(data, ctx->sync_staging->ptr, size, &transfer_ctx->out_memcpys);
+        ggml_vk_synchronize(ctx);
+    }
 }
 
 static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
@@ -12448,7 +12695,7 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
 
         if (ctx->transfer_ctx.expired()) {
             // Initialize new transfer context
-            transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
+            transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
             ctx->transfer_ctx = transfer_ctx;
             ggml_vk_ctx_begin(ctx->device, transfer_ctx);
         } else {
@@ -12465,36 +12712,56 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
     return false;
 }
 
+static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
+    VK_LOG_DEBUG("ggml_vk_synchronize()");
+
+    bool do_transfer = !ctx->transfer_ctx.expired();
+
+    vk_context transfer_ctx;
+    if (do_transfer) {
+        transfer_ctx = ctx->transfer_ctx.lock();
+
+        ggml_vk_ctx_end(transfer_ctx);
+
+        for (auto& cpy : transfer_ctx->in_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+
+        ggml_vk_submit(transfer_ctx, {});
+        ctx->submit_pending = true;
+    }
+
+    if (ctx->submit_pending) {
+        {
+            std::lock_guard<std::mutex> guard(queue_mutex);
+            ctx->device->compute_queue.queue.submit({}, ctx->fence);
+        }
+        ggml_vk_wait_for_fence(ctx);
+        ctx->submit_pending = false;
+    }
+
+    if (do_transfer) {
+        for (auto& cpy : transfer_ctx->out_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+        ctx->transfer_ctx.reset();
+    }
+}
+
 static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
     VK_LOG_DEBUG("ggml_backend_vk_synchronize()");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
-    if(ctx->transfer_ctx.expired()) {
-        return;
-    }
 
-    vk_context transfer_ctx = ctx->transfer_ctx.lock();
+    ggml_vk_synchronize(ctx);
 
-    ggml_vk_ctx_end(transfer_ctx);
-
-    for (auto& cpy : transfer_ctx->in_memcpys) {
-        memcpy(cpy.dst, cpy.src, cpy.n);
-    }
-
-    ggml_vk_submit(transfer_ctx, ctx->fence);
-    ggml_vk_wait_for_fence(ctx);
-
-    for (auto& cpy : transfer_ctx->out_memcpys) {
-        memcpy(cpy.dst, cpy.src, cpy.n);
-    }
-
-    ctx->transfer_ctx.reset();
+    ggml_vk_graph_cleanup(ctx);
 }
 
 static bool ggml_vk_is_empty(ggml_tensor * node) {
     return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
 }
 
-static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+static bool ggml_vk_can_fuse(const ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
     if (!ggml_can_fuse(cgraph, node_idx, ops)) {
         return false;
     }
@@ -12522,6 +12789,128 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
             return false;
         }
     }
+    auto const &mm_add_ok = [&](const ggml_tensor *mul, const ggml_tensor *add) {
+        const ggml_tensor *bias = add->src[0] == mul ? add->src[1] : add->src[0];
+
+        // mat-vec only
+        if (ggml_nrows(mul) != 1) {
+            return false;
+        }
+        // shaders assume the types match
+        if (mul->type != bias->type) {
+            return false;
+        }
+        // shaders reuse the D shape for bias
+        if (!ggml_are_same_shape(mul, bias) ||
+            !ggml_are_same_stride(mul, bias)) {
+            return false;
+        }
+        // unaligned bias isn't handled
+        if (get_misalign_bytes(ctx, bias) != 0) {
+            return false;
+        }
+        return true;
+    };
+
+    if ((ops.size() == 2 || ops.size() == 3) && ops.begin()[0] == GGML_OP_MUL_MAT && ops.begin()[1] == GGML_OP_ADD) {
+        // additional constraints specific to this fusion
+        const ggml_tensor *mul = cgraph->nodes[node_idx];
+        const ggml_tensor *add = cgraph->nodes[node_idx + 1];
+
+        if (!mm_add_ok(mul, add)) {
+            return false;
+        }
+        if (ops.size() == 3) {
+            if (ops.begin()[2] != GGML_OP_ADD) {
+                return false;
+            }
+            if (!mm_add_ok(add, cgraph->nodes[node_idx + 2])) {
+                return false;
+            }
+        }
+    }
+
+    auto const &mmid_mul_ok = [&](const ggml_tensor *mmid, const ggml_tensor *mul) {
+        const ggml_tensor *scale = mul->src[1];
+
+        if (mmid != mul->src[0]) {
+            return false;
+        }
+        // mat-vec only
+        if (!ggml_vk_use_mul_mat_vec_id(cgraph, node_idx)) {
+            return false;
+        }
+        // shaders assume the types match
+        if (mmid->type != scale->type) {
+            return false;
+        }
+        // shaders assume the bias is contiguous
+        if (!ggml_is_contiguous(scale)) {
+            return false;
+        }
+        // unaligned bias isn't handled
+        if (get_misalign_bytes(ctx, scale) != 0) {
+            return false;
+        }
+        // shader only indexes by expert index
+        if (scale->ne[0] != 1 ||
+            scale->ne[1] != mul->ne[1] ||
+            scale->ne[2] != 1 ||
+            scale->ne[3] != 1) {
+            return false;
+        }
+        return true;
+    };
+
+    if ((ops.size() == 2 || ops.size() == 3) && ops.begin()[0] == GGML_OP_MUL_MAT_ID && ops.begin()[1] == GGML_OP_ADD_ID) {
+        // additional constraints specific to this fusion
+        const ggml_tensor *mul = cgraph->nodes[node_idx];
+        const ggml_tensor *add = cgraph->nodes[node_idx + 1];
+        const ggml_tensor *bias = add->src[1];
+
+        if (mul != add->src[0]) {
+            return false;
+        }
+        // mat-vec only
+        if (!ggml_vk_use_mul_mat_vec_id(cgraph, node_idx)) {
+            return false;
+        }
+        // shaders assume the types match
+        if (mul->type != bias->type) {
+            return false;
+        }
+        // shaders assume the bias is contiguous
+        if (!ggml_is_contiguous(bias)) {
+            return false;
+        }
+        // the ID tensor must be the same for mul_mat_id and add_id
+        if (mul->src[2] != add->src[2]) {
+            return false;
+        }
+        // unaligned bias isn't handled
+        if (get_misalign_bytes(ctx, bias) != 0) {
+            return false;
+        }
+
+        if (ops.size() == 3) {
+            if (ops.begin()[2] != GGML_OP_MUL) {
+                return false;
+            }
+            const ggml_tensor *mul = cgraph->nodes[node_idx + 2];
+            return mmid_mul_ok(add, mul);
+        }
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_MUL_MAT_ID && ops.begin()[1] == GGML_OP_MUL) {
+        // additional constraints specific to this fusion
+        const ggml_tensor *mmid = cgraph->nodes[node_idx];
+        const ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+
+        if (!mmid_mul_ok(mmid, mul)) {
+            return false;
+        }
+    }
+
     return true;
 }
 
@@ -12617,6 +13006,76 @@ static bool ggml_vk_can_fuse_rope_set_rows(ggml_backend_vk_context * ctx, const
     return true;
 }
 
+// Check whether the tensors overlap in memory but are not equal.
+// Fusions can potenitally overwrite src tensors in ways that are not prevented
+// by ggml-alloc. If the fusion is entirely elementwise, then it's OK for them
+// to overlap if they are exactly equal.
+// XXX TODO this check is probably missing from several fusion optimizations.
+static bool ggml_vk_tensors_overlap_but_not_equal(const ggml_tensor * a, const ggml_tensor * b) {
+    ggml_backend_vk_buffer_context * a_buf_ctx = (ggml_backend_vk_buffer_context *)a->buffer->context;
+    vk_buffer a_buf = a_buf_ctx->dev_buffer;
+    ggml_backend_vk_buffer_context * b_buf_ctx = (ggml_backend_vk_buffer_context *)b->buffer->context;
+    vk_buffer b_buf = b_buf_ctx->dev_buffer;
+    if (a_buf == b_buf) {
+        auto a_base = vk_tensor_offset(a) + a->view_offs;
+        auto a_size = ggml_nbytes(a);
+        auto b_base = vk_tensor_offset(b) + b->view_offs;
+        auto b_size = ggml_nbytes(b);
+
+        if (a_base == b_base && a_size == b_size) {
+            return false;
+        }
+
+        if ((b_base <= a_base && a_base < b_base + b_size) ||
+            (a_base <= b_base && b_base < a_base + a_size)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+static bool ggml_vk_can_fuse_rms_norm_mul_rope(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph,
+                                               int node_idx) {
+    GGML_UNUSED(ctx);
+    const ggml_tensor *rms = cgraph->nodes[node_idx + 0];
+    const ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+    const ggml_tensor *rope = cgraph->nodes[node_idx + 2];
+
+    const int mode = ((const int32_t *) rope->op_params)[2];
+
+    // noncontig tensors aren't tested, and don't seem common in practice
+    if (!ggml_is_contiguous(rms) ||
+        !ggml_is_contiguous(mul) ||
+        !ggml_is_contiguous(rope)) {
+        return false;
+    }
+
+    // only norm/neox are handled in the shader
+    if (mode != GGML_ROPE_TYPE_NEOX && mode != GGML_ROPE_TYPE_NORMAL) {
+        return false;
+    }
+
+    // shared memory size for passing data from mul->rope
+    if (mul->ne[0] > 1024) {
+        return false;
+    }
+
+    // must not overwrite srcs in a way that's not elementwise
+    ggml_tensor *other_src = mul->src[0] == rms ? mul->src[1] : mul->src[0];
+    if (ggml_vk_tensors_overlap_but_not_equal(rms->src[0], rope) ||
+        ggml_vk_tensors_overlap_but_not_equal(other_src, rope)) {
+        return false;
+    }
+
+    // conditions for pipeline creation
+    if (!(ctx->device->float_controls_rte_fp16 &&
+        sizeof(vk_op_rms_norm_mul_rope_push_constants) <= ctx->device->properties.limits.maxPushConstantsSize)) {
+        return false;
+    }
+
+    return true;
+}
+
 static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
 
     const ggml_tensor *first_node = cgraph->nodes[node_idx];
@@ -12680,54 +13139,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul));
     }
 
-    ctx->prealloc_size_add_rms_partials = 0;
     ctx->prealloc_size_add_rms_partials_offset = 0;
     ctx->do_add_rms_partials = false;
-
-    uint64_t total_mat_mul_bytes = 0;
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (!ctx->device->disable_fusion) {
-            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
-            if (num_adds) {
-                ctx->num_additional_fused_ops = num_adds - 1;
-            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-                ctx->num_additional_fused_ops = 1;
-            } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 2 }) &&
-                       ggml_check_edges(cgraph, i, rope_view_set_rows_edges) &&
-                       ggml_vk_can_fuse_rope_set_rows(ctx, cgraph, i)) {
-                ctx->num_additional_fused_ops = 2;
-            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax_norm, { i + 3, i + 9 }) &&
-                       ggml_check_edges(cgraph, i, topk_moe_early_softmax_norm_edges) &&
-                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX_NORM)) {
-                ctx->num_additional_fused_ops = topk_moe_early_softmax_norm.size() - 1;
-            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax, { i + 3, i + 4 }) &&
-                       ggml_check_edges(cgraph, i, topk_moe_early_softmax_edges) &&
-                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX)) {
-                ctx->num_additional_fused_ops = topk_moe_early_softmax.size() - 1;
-            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_late_softmax, { i + 1, i + 5 }) &&
-                       ggml_check_edges(cgraph, i, topk_moe_late_softmax_edges) &&
-                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_LATE_SOFTMAX)) {
-                ctx->num_additional_fused_ops = topk_moe_late_softmax.size() - 1;
-            }
-        }
-        ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
-        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
-            total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
-        } else if (cgraph->nodes[i]->op == GGML_OP_CONV_2D || cgraph->nodes[i]->op == GGML_OP_CONV_TRANSPOSE_2D) {
-            // Return CRSxNPQxsizeof(*) to account as many bytes as mul_mat has in im2col->mul_mat mode.
-            auto CRS_size =
-                cgraph->nodes[i]->src[0]->ne[0] * cgraph->nodes[i]->src[0]->ne[1] * cgraph->nodes[i]->src[1]->ne[2];
-            auto NPQ_size = cgraph->nodes[i]->ne[0] * cgraph->nodes[i]->ne[1] * cgraph->nodes[i]->ne[3];
-            total_mat_mul_bytes += NPQ_size * CRS_size * ggml_type_size(cgraph->nodes[i]->type);
-        }
-        i += ctx->num_additional_fused_ops;
-        ctx->num_additional_fused_ops = 0;
-    }
-    if (ctx->device->need_compiles) {
-        ggml_vk_load_shaders(ctx->device);
-    }
-    ggml_vk_preallocate_buffers(ctx);
-    ggml_pipeline_allocate_descriptor_sets(ctx);
+    ctx->do_add_rms_partials_offset_calculation = false;
 
     int last_node = cgraph->n_nodes - 1;
 
@@ -12769,6 +13183,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     ctx->prealloc_y_last_tensor_used = nullptr;
 
     if (ctx->prealloc_size_add_rms_partials) {
+        ggml_vk_preallocate_buffers(ctx, nullptr);
         if (ctx->compute_ctx.expired()) {
             compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
             ctx->compute_ctx = compute_ctx;
@@ -12789,21 +13204,42 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     int submitted_nodes = 0;
     int submit_count = 0;
     uint64_t mul_mat_bytes = 0;
-    uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), total_mat_mul_bytes / 40u);
+    uint64_t total_mul_mat_bytes = 0;
+    uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), ctx->last_total_mul_mat_bytes / 40u);
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (first_node_in_batch) {
             submit_node_idx = i;
         }
 
         if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
-            mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+            auto bytes = ggml_nbytes(cgraph->nodes[i]->src[0]);
+            mul_mat_bytes += bytes;
+            total_mul_mat_bytes += bytes;
         }
 
         if (!ctx->device->disable_fusion) {
             uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
             if (num_adds) {
                 ctx->num_additional_fused_ops = num_adds - 1;
-            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_MUL_MAT, GGML_OP_ADD, GGML_OP_ADD })) {
+                ctx->num_additional_fused_ops = 2;
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_MUL_MAT, GGML_OP_ADD })) {
+                ctx->num_additional_fused_ops = 1;
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 2;
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID })) {
+                ctx->num_additional_fused_ops = 1;
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_MUL_MAT_ID, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 4 }) &&
+                       ggml_check_edges(cgraph, i, rms_norm_mul_rope_view_set_rows_edges) &&
+                       ggml_vk_can_fuse_rms_norm_mul_rope(ctx, cgraph, i) &&
+                       ggml_vk_can_fuse_rope_set_rows(ctx, cgraph, i + 2)) {
+                ctx->num_additional_fused_ops = 4;
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ROPE })&&
+                       ggml_vk_can_fuse_rms_norm_mul_rope(ctx, cgraph, i)) {
+                ctx->num_additional_fused_ops = 2;
+            } else if (ggml_vk_can_fuse(ctx, cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
                 ctx->num_additional_fused_ops = 1;
             } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 2 }) &&
                        ggml_check_edges(cgraph, i, rope_view_set_rows_edges) &&
@@ -12834,11 +13270,11 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
         bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5;
         bool submit = (submitted_nodes >= nodes_per_submit) ||
-                      (mul_mat_bytes >= mul_mat_bytes_per_submit) ||
+                      (mul_mat_bytes_per_submit != 0 && mul_mat_bytes >= mul_mat_bytes_per_submit) ||
                       (i + ctx->num_additional_fused_ops >= last_node) ||
                       (almost_ready && !ctx->almost_ready_fence_pending);
 
-        bool enqueued = ggml_vk_build_graph(ctx, cgraph, i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i + ctx->num_additional_fused_ops >= last_node, almost_ready, submit);
+        bool enqueued = ggml_vk_build_graph(ctx, cgraph, i, cgraph->nodes[submit_node_idx], submit_node_idx, i + ctx->num_additional_fused_ops >= last_node, almost_ready, submit);
 
         if (vk_perf_logger_enabled) {
             if (ctx->compute_ctx.expired()) {
@@ -12878,6 +13314,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ctx->fused_ops_write_mask = 0;
     }
 
+    ctx->last_total_mul_mat_bytes = total_mul_mat_bytes;
+
     if (vk_perf_logger_enabled) {
         // End the command buffer and submit/wait
         GGML_ASSERT(!ctx->compute_ctx.expired());
@@ -12900,7 +13338,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ctx->device->perf_logger->print_timings();
     }
 
-    ggml_vk_graph_cleanup(ctx);
+    if (!ctx->device->support_async) {
+        ggml_vk_synchronize(ctx);
+    }
 
     return GGML_STATUS_SUCCESS;
 
@@ -12937,26 +13377,10 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
         return false;
     };
 
-    // This function tries to reorder the graph to allow nodes to run in parallel.
-    // This helps with small batches, but for large batches its a slowdown, probably
-    // due to cache contention. So only reorder if the majority of nodes have few rows.
-    int num_small_nodes = 0;
-    int num_counted_nodes = 0;
-    for (int i = 0; i < graph->n_nodes; ++i) {
-        if (!is_empty(graph->nodes[i]) &&
-            graph->nodes[i]->op != GGML_OP_SET_ROWS) {
-            if (ggml_nrows(graph->nodes[i]) <= 8) {
-                num_small_nodes++;
-            }
-            num_counted_nodes++;
-        }
-    }
-    if (num_small_nodes < num_counted_nodes / 2) {
-        return;
-    }
-
     std::vector<ggml_tensor *> new_order;
     std::vector<bool> used(graph->n_nodes, false);
+    std::set<ggml_tensor *> used_node_set;
+
     int first_unused = 0;
     while (first_unused < graph->n_nodes) {
         std::vector<int> current_set;
@@ -12979,6 +13403,7 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
             if (match_pattern(pattern, first_unused)) {
                 for (size_t j = 0; j < pattern.size(); ++j) {
                     new_order.push_back(graph->nodes[first_unused + j]);
+                    used_node_set.insert(graph->nodes[first_unused + j]);
                     used[first_unused + j] = true;
                 }
                 while (first_unused < graph->n_nodes && used[first_unused]) {
@@ -13026,21 +13451,44 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
             for (int c = first_unused; c < j; ++c) {
                 if (!used[c] &&
                     is_src_of(graph->nodes[j], graph->nodes[c]) &&
-                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_RMS_NORM && graph->nodes[j]->op == GGML_OP_MUL)) {
+                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_RMS_NORM && graph->nodes[j]->op == GGML_OP_MUL) &&
+                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_MUL_MAT && graph->nodes[j]->op == GGML_OP_ADD) &&
+                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_MUL_MAT_ID && graph->nodes[j]->op == GGML_OP_ADD_ID) &&
+                    !(j == c+1 && c == current_set.back() && graph->nodes[c]->op == GGML_OP_MUL_MAT_ID && graph->nodes[j]->op == GGML_OP_MUL)) {
                     ok = false;
                     break;
                 }
             }
             if (ok) {
                 current_set.push_back(j);
+
+                int rope_idx = j;
+
+                // When we've found RMS_NORM + MUL, try to find a ROPE that uses it
+                if (j > 0 &&
+                    graph->nodes[j]->op == GGML_OP_MUL &&
+                    graph->nodes[j-1]->op == GGML_OP_RMS_NORM) {
+                    for (int k = j + 1; k < std::min(j + 15, graph->n_nodes); ++k) {
+                        if (graph->nodes[k]->op == GGML_OP_ROPE &&
+                            graph->nodes[k]->src[0] == graph->nodes[j] &&
+                            // Check that other srcs are already valid
+                            graph->nodes[k]->src[1]->op == GGML_OP_NONE &&
+                            (graph->nodes[k]->src[2] == nullptr || graph->nodes[k]->src[2]->op == GGML_OP_NONE)) {
+                            rope_idx = k;
+                            current_set.push_back(rope_idx);
+                            used[rope_idx] = true;
+                            break;
+                        }
+                    }
+                }
                 // Look for ROPE + VIEW + SET_ROWS and make them consecutive
-                if (graph->nodes[j]->op == GGML_OP_ROPE) {
+                if (graph->nodes[rope_idx]->op == GGML_OP_ROPE) {
                     int view_idx = -1;
                     int set_rows_idx = -1;
-                    for (int k = j+1; k < std::min(j + 10, graph->n_nodes); ++k) {
+                    for (int k = rope_idx+1; k < std::min(rope_idx + 10, graph->n_nodes); ++k) {
                         if (view_idx == -1 &&
                             graph->nodes[k]->op == GGML_OP_VIEW &&
-                            graph->nodes[k]->src[0] == graph->nodes[j]) {
+                            graph->nodes[k]->src[0] == graph->nodes[rope_idx]) {
                             view_idx = k;
                             continue;
                         }
@@ -13059,6 +13507,36 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
                         used[set_rows_idx] = true;
                     }
                 }
+                // Look for MUL_MAT_ID + ADD_ID + MUL
+                if (j > 0 &&
+                    graph->nodes[j]->op == GGML_OP_ADD_ID &&
+                    graph->nodes[j-1]->op == GGML_OP_MUL_MAT_ID) {
+                    for (int k = j + 1; k < std::min(j + 15, graph->n_nodes); ++k) {
+                        if (graph->nodes[k]->op == GGML_OP_MUL &&
+                            graph->nodes[k]->src[0] == graph->nodes[j] &&
+                            // src1 must either be weights or already processed
+                            (graph->nodes[k]->src[1]->op == GGML_OP_NONE || used_node_set.find(graph->nodes[k]->src[1]) != used_node_set.end())) {
+                            current_set.push_back(k);
+                            used[k] = true;
+                            break;
+                        }
+                    }
+                }
+                // Look for MUL_MAT + ADD + ADD
+                if (j > 0 &&
+                    graph->nodes[j]->op == GGML_OP_ADD &&
+                    graph->nodes[j-1]->op == GGML_OP_MUL_MAT) {
+                    for (int k = j + 1; k < std::min(j + 15, graph->n_nodes); ++k) {
+                        if (graph->nodes[k]->op == GGML_OP_ADD &&
+                            graph->nodes[k]->src[0] == graph->nodes[j] &&
+                            // src1 must either be weights or already processed
+                            (graph->nodes[k]->src[1]->op == GGML_OP_NONE || used_node_set.find(graph->nodes[k]->src[1]) != used_node_set.end())) {
+                            current_set.push_back(k);
+                            used[k] = true;
+                            break;
+                        }
+                    }
+                }
             }
         }
         // Second pass grabs view nodes.
@@ -13091,6 +13569,7 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
         // Push the current set into new_order
         for (auto c : current_set) {
             new_order.push_back(graph->nodes[c]);
+            used_node_set.insert(graph->nodes[c]);
             used[c] = true;
         }
         while (first_unused < graph->n_nodes && used[first_unused]) {
@@ -13108,9 +13587,9 @@ static ggml_backend_i ggml_backend_vk_interface = {
     /* .get_name                = */ ggml_backend_vk_name,
     /* .free                    = */ ggml_backend_vk_free,
     /* .set_tensor_async        = */ NULL,  // ggml_backend_vk_set_tensor_async,
-    /* .get_tensor_async        = */ NULL,  // ggml_backend_vk_get_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_vk_get_tensor_async,
     /* .cpy_tensor_async        = */ NULL,  // ggml_backend_vk_cpy_tensor_async,
-    /* .synchronize             = */ NULL,  // ggml_backend_vk_synchronize,
+    /* .synchronize             = */ ggml_backend_vk_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_update       = */ NULL,
@@ -13139,6 +13618,10 @@ ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
         /* .context = */ ctx,
     };
 
+    if (!ctx->device->support_async) {
+        vk_backend->iface.get_tensor_async = nullptr;
+    }
+
     return vk_backend;
 }
 
@@ -13185,8 +13668,11 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
     GGML_ASSERT(ctx->device < (int) vk_instance.device_supports_membudget.size());
 
     vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[ctx->device]];
-
-    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+    vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops;
+    vk::PhysicalDeviceMemoryProperties2 memprops = {};
+    const bool membudget_supported = vk_instance.device_supports_membudget[ctx->device];
+    const bool is_integrated_gpu = vkdev.getProperties().deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
+    
     vk::PhysicalDeviceProperties2 props2;
     vkdev.getProperties2(&props2);
     GGML_LOG_DEBUG("ggml_backend_vk_get_device_memory called: uuid %s\n", ctx->uuid.c_str());
@@ -13204,7 +13690,7 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
         ggml_dxgi_pdh_release();
     }
 
-    if (!ctx->is_integrated_gpu) 
+    if (!is_integrated_gpu)
     {
         // Use vendor specific management libraries for best VRAM reporting if available
         switch (props2.properties.vendorID) {
@@ -13232,42 +13718,27 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
             break;
         }
     }
-
     // else fallback to memory budget if supported
+
+    if (membudget_supported) {
+        memprops.pNext = &budgetprops;
+    }
+    vkdev.getMemoryProperties2(&memprops);
+
     *total = 0;
     *free = 0;
-    vk::PhysicalDeviceMemoryBudgetPropertiesEXT mem_budget_props;
-    vk::PhysicalDeviceMemoryProperties2 memprops2;
-    memprops2.pNext = &mem_budget_props;
-    vkdev.getMemoryProperties2(&memprops2);
-    for (int i = 0; i < memprops2.memoryProperties.memoryHeapCount; i++) {
-        if (memprops2.memoryProperties.memoryHeaps[i].flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-            *total += memprops2.memoryProperties.memoryHeaps[i].size;
-        } else if (ctx->is_integrated_gpu) {
-            // Include shared memory on iGPUs
-            *total += memprops2.memoryProperties.memoryHeaps[i].size;
-        }
-    }
-    for (int i = 0; i < memprops2.memoryProperties.memoryHeapCount; i++) {
-        if (memprops2.memoryProperties.memoryHeaps[i].flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-            *free += mem_budget_props.heapBudget[i];
-        } else if (ctx->is_integrated_gpu) {
-            *free += mem_budget_props.heapBudget[i];
-        }
-    }
-    if (*total > 0 && *free > 0) {
-        return;
-    } else if (*total > 0) {
-        *free = *total;
-        return;
-    }
 
-    // else just report the physical memory
-    for (const vk::MemoryHeap& heap : memprops2.memoryProperties.memoryHeaps) {
-        if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-            *total = heap.size;
-            *free = heap.size;
-            break;
+    for (uint32_t i = 0; i < memprops.memoryProperties.memoryHeapCount; ++i) {
+        const vk::MemoryHeap & heap = memprops.memoryProperties.memoryHeaps[i];
+
+        if (is_integrated_gpu || (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal)) {
+            *total += heap.size;
+
+            if (membudget_supported && i < budgetprops.heapUsage.size()) {
+                *free += budgetprops.heapBudget[i] - budgetprops.heapUsage[i];
+            } else {
+                *free += heap.size;
+            }
         }
     }
 }
@@ -13417,10 +13888,18 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_NEG:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_ABS:
+                case GGML_UNARY_OP_SOFTPLUS:
+                case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous(op->src[0]) &&
                            (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
                            (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
@@ -13567,8 +14046,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 default:
                     return false;
                 }
-                if (!coopmat2 && !device->subgroup_shuffle) {
-                    // scalar FA uses subgroupShuffle
+                if (!coopmat2 && !(device->subgroup_shuffle && device->subgroup_vote)) {
+                    // scalar/coopmat1 FA uses subgroupShuffle/subgroupAll
                     return false;
                 }
                 return true;
@@ -13671,10 +14150,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 }
 
                 // We can handle copying from a type to the same type if it's
-                // contiguous (memcpy). We use f16 or f32 shaders to do the copy,
+                // either not quantized or is quantized and contiguous.
+                // We use f16 or f32 shaders to do the copy,
                 // so the type/block size must be a multiple of 4.
                 if (src0_type == src1_type &&
-                    ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op) &&
+                    (!ggml_is_quantized(src0_type) || (ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op))) &&
                     (ggml_type_size(src0_type) % 2) == 0) {
                     return true;
                 }
@@ -13709,41 +14189,130 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                    op->type == GGML_TYPE_F32;
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM_BACK:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
-            return op->src[0]->type == GGML_TYPE_F32;
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_LOG:
+        case GGML_OP_TRI:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                   op->type == op->src[0]->type;
         case GGML_OP_ARGSORT:
-            return op->ne[0] <= max_argsort_cols;
+            {
+                if (!ggml_is_contiguous(op) || !ggml_is_contiguous(op->src[0])) {
+                    return false;
+                }
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                auto device = ggml_vk_get_device(ctx->device);
+                // pipeline_argsort_large_f32 requires vulkan memory model.
+                if (device->vulkan_memory_model) {
+                    return true;
+                } else {
+                    return op->ne[0] <= (1 << device->max_workgroup_size_log2);
+                }
+            }
+        case GGML_OP_TOP_K:
+            {
+                if (!ggml_is_contiguous(op) || !ggml_is_contiguous(op->src[0])) {
+                    return false;
+                }
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                auto device = ggml_vk_get_device(ctx->device);
+                // We could potentially support larger, using argsort to sort the
+                // whole thing. Not clear if this is needed.
+                uint32_t min_pipeline = (uint32_t)log2f(float(op->ne[0])) + 1;
+                if (min_pipeline >= num_topk_pipelines ||
+                    !device->pipeline_topk_f32[min_pipeline]) {
+                    return false;
+                }
+            }
+            return true;
         case GGML_OP_UPSCALE:
         case GGML_OP_ACC:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONCAT:
+            return ggml_type_size(op->src[0]->type) == ggml_type_size(GGML_TYPE_F32);
+        case GGML_OP_ADD1:
+            return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32)
+                || (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32)
+                || (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16);
+        case GGML_OP_ARANGE:
+        case GGML_OP_FILL:
+            return op->type == GGML_TYPE_F32;
         case GGML_OP_SCALE:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
         case GGML_OP_ROLL:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_DIAG_MASK_INF:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32
+                && (!op->src[1] || (op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16));
         case GGML_OP_SOFT_MAX_BACK:
-            return true;
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32
+                && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
+        case GGML_OP_CUMSUM:
+            {
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                auto device = ggml_vk_get_device(ctx->device);
+                if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
+                    return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
+                }
+                return false;
+            }
+        case GGML_OP_SOLVE_TRI:
+            {
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                const vk_device& device = ggml_vk_get_device(ctx->device);
+
+                if (op->type != GGML_TYPE_F32 || op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                const uint32_t N = op->src[0]->ne[0];
+                const uint32_t K = op->src[1]->ne[0];
+                // K dimension limited to workgroup size
+                if (K > 128) {
+                    return false;
+                }
+                if (N * N * sizeof(float) + N * K * sizeof(float) > device->properties.limits.maxComputeSharedMemorySize) {
+                    return false;
+                }
+                return true;
+            }
         case GGML_OP_ARGMAX:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_COUNT_EQUAL:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_I32
+                && ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_I32;
         case GGML_OP_IM2COL:
+            return ggml_is_contiguous(op->src[1])
+                && op->src[1]->type == GGML_TYPE_F32
+                && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
         case GGML_OP_IM2COL_3D:
+            return op->src[1]->type == GGML_TYPE_F32
+                && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
         case GGML_OP_TIMESTEP_EMBEDDING:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_2D_DW:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16)
+                && op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_POOL_2D:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
-            return true;
+            return true; // all inputs are contiguous, see ggml.c
         case GGML_OP_SSM_SCAN:
             {
                 for (int i = 0; i < 6; i++) {
@@ -13784,7 +14353,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 return true;
             }
         case GGML_OP_SSM_CONV:
-            return true;
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_TRANSPOSE_1D:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_2D:
@@ -14135,20 +14704,11 @@ size_t comp_size;
 size_t comp_nb[GGML_MAX_DIMS];
 size_t check_counter = 0;
 static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
-    ggml_tensor * tensor = cgraph->nodes[tensor_idx];
+    ggml_tensor * tensor = cgraph->nodes[tensor_idx + ctx->num_additional_fused_ops];
     if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
 
-    bool fused_rms_norm_mul = false;
-    int rms_norm_idx = -1;
-    if (ctx->num_additional_fused_ops == 1 &&
-        tensor->op == GGML_OP_RMS_NORM &&
-        cgraph->nodes[tensor_idx + 1]->op == GGML_OP_MUL) {
-        fused_rms_norm_mul = true;
-        tensor = cgraph->nodes[tensor_idx + 1];
-    }
-
     check_counter++;
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
         return;
@@ -14156,9 +14716,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
 
     VK_LOG_DEBUG("ggml_vk_check_results_0(" << tensor->name << ")");
 
-    ggml_tensor * src0 = tensor->src[0];
-    ggml_tensor * src1 = tensor->src[1];
-
     struct ggml_init_params iparams = {
         /*.mem_size   =*/ 2ul*1024ul*1024ul*1024ul,
         /*.mem_buffer =*/ NULL,
@@ -14168,328 +14725,383 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
     struct ggml_context * ggml_ctx = ggml_init(iparams);
 
     std::array<struct ggml_tensor *, GGML_MAX_SRC> src_clone = {nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr};
-    std::array<size_t, GGML_MAX_SRC> src_size = {};
-    std::array<void *, GGML_MAX_SRC> src_buffer = {};
     const char * srci_name[GGML_MAX_SRC] = {"src0", "src1", "src2", "src3", "src4", "src5", "src6", "src7", "src8", "src9"};
 
+    std::map<ggml_tensor *, ggml_tensor *> cloned_tensors;
+    std::vector<void *> cloned_mallocs;
+
     struct ggml_tensor * tensor_clone = nullptr;
 
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        ggml_tensor * srci = tensor->src[i];
-        if (fused_rms_norm_mul) {
-            rms_norm_idx = tensor->src[0]->op == GGML_OP_RMS_NORM ? 0 : 1;
-            ggml_tensor *rms_norm = tensor->src[rms_norm_idx];
-            switch (i) {
-            case 0: srci = rms_norm->src[0]; break;
-            case 1: srci = tensor->src[1 - rms_norm_idx]; break;
-            default: continue;
+    for (int f = 0; f < ctx->num_additional_fused_ops + 1; ++f) {
+        tensor = cgraph->nodes[tensor_idx + f];
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            ggml_tensor * srci = tensor->src[i];
+            if (srci == nullptr) {
+                continue;
             }
-        }
-        if (srci == nullptr) {
-            continue;
-        }
-        ggml_tensor * srci_clone = ggml_dup_tensor(ggml_ctx, srci);
-        size_t srci_size = ggml_nbytes(srci);
+            // If a src tensor has been cloned, use that one
+            auto it = cloned_tensors.find(srci);
+            if (it != cloned_tensors.end()) {
+                src_clone[i] = it->second;
+                continue;
+            }
+            ggml_tensor * srci_clone = ggml_dup_tensor(ggml_ctx, srci);
+            size_t srci_size = ggml_nbytes(srci);
 
-        src_clone[i] = srci_clone;
-        src_size[i] = ggml_nbytes(srci);
-        src_buffer[i] = malloc(srci_size);
+            src_clone[i] = srci_clone;
+            void *src_buffer = malloc(srci_size);
+            cloned_mallocs.push_back(src_buffer);
 
-        srci_clone->data = src_buffer[i];
-        if (ggml_backend_buffer_is_host(srci->buffer)) {
-            memcpy(srci_clone->data, srci->data, srci_size);
-            memcpy(srci_clone->nb, srci->nb, sizeof(size_t) * GGML_MAX_DIMS);
-        } else if (ggml_backend_buffer_is_vk(srci->buffer)) {
-            ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)srci->buffer->context;
-            vk_buffer& buffer_gpu = buf_ctx->dev_buffer;
-            uint64_t offset = vk_tensor_offset(srci) + srci->view_offs;
-            if (!ggml_is_contiguous(srci) && ggml_vk_dim01_contiguous(srci)) {
-                for (int i3 = 0; i3 < srci->ne[3]; i3++) {
-                    for (int i2 = 0; i2 < srci->ne[2]; i2++) {
-                        const int idx = i3*srci->ne[2] + i2;
-                        ggml_vk_buffer_read(buffer_gpu, offset + idx * srci->nb[2], ((char *)srci_clone->data + idx * srci_clone->nb[2]), srci->ne[1] * srci->nb[1]);
-                    }
-                }
-
-                srci_clone->nb[0] = srci->nb[0];
-                srci_clone->nb[1] = srci->nb[1];
-                for (int i = 2; i < GGML_MAX_DIMS; i++) {
-                    srci_clone->nb[i] = srci_clone->nb[i - 1]*srci_clone->ne[i - 1];
-                }
-            } else {
-                if (offset + srci_size >= buffer_gpu->size) {
-                    srci_size = buffer_gpu->size - offset;
-                }
-                ggml_vk_buffer_read(buffer_gpu, offset, srci_clone->data, srci_size);
+            srci_clone->data = src_buffer;
+            if (ggml_backend_buffer_is_host(srci->buffer)) {
+                memcpy(srci_clone->data, srci->data, srci_size);
                 memcpy(srci_clone->nb, srci->nb, sizeof(size_t) * GGML_MAX_DIMS);
+            } else if (ggml_backend_buffer_is_vk(srci->buffer)) {
+                ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)srci->buffer->context;
+                vk_buffer& buffer_gpu = buf_ctx->dev_buffer;
+                uint64_t offset = vk_tensor_offset(srci) + srci->view_offs;
+                if (!ggml_is_contiguous(srci) && ggml_vk_dim01_contiguous(srci)) {
+                    for (int i3 = 0; i3 < srci->ne[3]; i3++) {
+                        for (int i2 = 0; i2 < srci->ne[2]; i2++) {
+                            const int idx = i3*srci->ne[2] + i2;
+                            ggml_vk_buffer_read(buffer_gpu, offset + idx * srci->nb[2], ((char *)srci_clone->data + idx * srci_clone->nb[2]), srci->ne[1] * srci->nb[1]);
+                        }
+                    }
+
+                    srci_clone->nb[0] = srci->nb[0];
+                    srci_clone->nb[1] = srci->nb[1];
+                    for (int i = 2; i < GGML_MAX_DIMS; i++) {
+                        srci_clone->nb[i] = srci_clone->nb[i - 1]*srci_clone->ne[i - 1];
+                    }
+                } else {
+                    if (offset + srci_size >= buffer_gpu->size) {
+                        srci_size = buffer_gpu->size - offset;
+                    }
+                    ggml_vk_buffer_read(buffer_gpu, offset, srci_clone->data, srci_size);
+                    memcpy(srci_clone->nb, srci->nb, sizeof(size_t) * GGML_MAX_DIMS);
+                }
+            } else {
+                GGML_ABORT("fatal error");
+            }
+
+            if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+                ggml_vk_print_tensor(srci, srci_name[i]);
             }
-        } else {
-            GGML_ABORT("fatal error");
         }
 
-        if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
-            ggml_vk_print_tensor(srci, srci_name[i]);
-        }
-    }
-
-    if (tensor->op == GGML_OP_FLASH_ATTN_EXT) {
-        const float * params = (const float *)tensor->op_params;
-        tensor_clone = ggml_flash_attn_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3], params[0], params[1], params[2]);
-        if (src_clone[4]) {
-            ggml_flash_attn_ext_add_sinks(tensor_clone, src_clone[4]);
-        }
-    } else if (tensor->op == GGML_OP_MUL_MAT) {
-        tensor_clone = ggml_mul_mat(ggml_ctx, src_clone[0], src_clone[1]);
-    } else if (tensor->op == GGML_OP_MUL_MAT_ID) {
-        tensor_clone = ggml_mul_mat_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
-    } else if (tensor->op == GGML_OP_SUB) {
-        tensor_clone = ggml_sub(ggml_ctx, src_clone[0], src_clone[1]);
-    } else if (tensor->op == GGML_OP_MUL) {
-        if (fused_rms_norm_mul) {
-            tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->src[rms_norm_idx]->op_params);
-            tensor_clone = ggml_mul(ggml_ctx, tensor_clone, src_clone[1 - rms_norm_idx]);
-        } else {
-            tensor_clone = ggml_mul(ggml_ctx, src_clone[0], src_clone[1]);
-        }
-    } else if (tensor->op == GGML_OP_DIV) {
-        tensor_clone = ggml_div(ggml_ctx, src_clone[0], src_clone[1]);
-    } else if (tensor->op == GGML_OP_CONCAT) {
-        tensor_clone = ggml_concat(ggml_ctx, src_clone[0], src_clone[1], *(int *)tensor->op_params);
-    } else if (tensor->op == GGML_OP_UPSCALE) {
-        tensor_clone = ggml_interpolate(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
-    } else if (tensor->op == GGML_OP_SCALE) {
-        const float * params = (const float *)tensor->op_params;
-        tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
-    } else if (tensor->op == GGML_OP_SQR) {
-        tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_SQRT) {
-        tensor_clone = ggml_sqrt(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_SIN) {
-        tensor_clone = ggml_sin(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_COS) {
-        tensor_clone = ggml_cos(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_CLAMP) {
-        const float * params = (const float *)tensor->op_params;
-        tensor_clone = ggml_clamp(ggml_ctx, src_clone[0], params[0], params[1]);
-    } else if (tensor->op == GGML_OP_PAD) {
-        tensor_clone = ggml_pad_ext(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3],
-                                                            tensor->op_params[4], tensor->op_params[5], tensor->op_params[6], tensor->op_params[7]);
-    } else if (tensor->op == GGML_OP_REPEAT) {
-        tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor);
-    } else if (tensor->op == GGML_OP_REPEAT_BACK) {
-        tensor_clone = ggml_repeat_back(ggml_ctx, src_clone[0], tensor);
-    } else if (tensor->op == GGML_OP_ADD) {
-        tensor_clone = ggml_add(ggml_ctx, src_clone[0], src_clone[1]);
-    } else if (tensor->op == GGML_OP_ACC) {
-        tensor_clone = ggml_acc(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]);
-    } else if (tensor->op == GGML_OP_NORM) {
-        tensor_clone = ggml_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params);
-    } else if (tensor->op == GGML_OP_GROUP_NORM) {
-        const float * float_params = (const float *)tensor->op_params;
-        tensor_clone = ggml_group_norm(ggml_ctx, src_clone[0], tensor->op_params[0], float_params[1]);
-    } else if (tensor->op == GGML_OP_RMS_NORM) {
-        tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params);
-    } else if (tensor->op == GGML_OP_RMS_NORM_BACK) {
-        const float eps = ((float *) tensor->op_params)[0];
-        tensor_clone = ggml_rms_norm_back(ggml_ctx, src_clone[0], src_clone[1], eps);
-    } else if (tensor->op == GGML_OP_SILU_BACK) {
-        tensor_clone = ggml_silu_back(ggml_ctx, src_clone[0], src_clone[1]);
-    } else if (tensor->op == GGML_OP_L2_NORM) {
-        const float eps = ((float *) tensor->op_params)[0];
-        tensor_clone = ggml_l2_norm(ggml_ctx, src_clone[0], eps);
-    } else if (tensor->op == GGML_OP_SOFT_MAX) {
-        if (src1 != nullptr) {
+        if (tensor->op == GGML_OP_FLASH_ATTN_EXT) {
             const float * params = (const float *)tensor->op_params;
-            tensor_clone = ggml_soft_max_ext(ggml_ctx, src_clone[0], src_clone[1], params[0], params[1]);
-        } else {
-            tensor_clone = ggml_soft_max(ggml_ctx, src_clone[0]);
-        }
-    } else if (tensor->op == GGML_OP_SOFT_MAX_BACK) {
-        tensor_clone = ggml_soft_max_ext_back(ggml_ctx, src_clone[0], src_clone[1], ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
-    } else if (tensor->op == GGML_OP_DIAG_MASK_INF) {
-        tensor_clone = ggml_diag_mask_inf(ggml_ctx, src_clone[0], tensor->op_params[0]);
-    } else if (tensor->op == GGML_OP_ROPE || tensor->op == GGML_OP_ROPE_BACK) {
-        const int n_dims      = ((int32_t *) tensor->op_params)[1];
-        const int mode        = ((int32_t *) tensor->op_params)[2];
-        //const int n_ctx_ggml       = ((int32_t *) tensor->op_params)[3];
-        const int n_ctx_orig_ggml  = ((int32_t *) tensor->op_params)[4];
-        const float freq_base       = ((float *) tensor->op_params)[5];
-        const float freq_scale      = ((float *) tensor->op_params)[6];
-        const float ext_factor      = ((float *) tensor->op_params)[7];
-        const float attn_factor     = ((float *) tensor->op_params)[8];
-        const float beta_fast       = ((float *) tensor->op_params)[9];
-        const float beta_slow       = ((float *) tensor->op_params)[10];
-        if (mode & GGML_ROPE_TYPE_MROPE) {
-            int32_t *sections = ((int32_t *) tensor->op_params) + 11;
-            if (tensor->op == GGML_OP_ROPE) {
-                tensor_clone = ggml_rope_multi(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
-            } else {
-                tensor_clone = ggml_rope_multi_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            tensor_clone = ggml_flash_attn_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3], params[0], params[1], params[2]);
+            if (src_clone[4]) {
+                ggml_flash_attn_ext_add_sinks(tensor_clone, src_clone[4]);
             }
-        } else {
-            if (tensor->op == GGML_OP_ROPE) {
-                tensor_clone = ggml_rope_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+        } else if (tensor->op == GGML_OP_MUL_MAT) {
+            tensor_clone = ggml_mul_mat(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_MUL_MAT_ID) {
+            tensor_clone = ggml_mul_mat_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
+        } else if (tensor->op == GGML_OP_SUB) {
+            tensor_clone = ggml_sub(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_MUL) {
+            tensor_clone = ggml_mul(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_DIV) {
+            tensor_clone = ggml_div(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_CONCAT) {
+            tensor_clone = ggml_concat(ggml_ctx, src_clone[0], src_clone[1], *(int *)tensor->op_params);
+        } else if (tensor->op == GGML_OP_UPSCALE) {
+            tensor_clone = ggml_interpolate(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
+        } else if (tensor->op == GGML_OP_SCALE) {
+            const float * params = (const float *)tensor->op_params;
+            tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
+        } else if (tensor->op == GGML_OP_ADD1) {
+            tensor_clone = ggml_add1(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_ARANGE) {
+            const float start = ggml_get_op_params_f32(tensor, 0);
+            const float stop = ggml_get_op_params_f32(tensor, 1);
+            const float step = ggml_get_op_params_f32(tensor, 2);
+            tensor_clone = ggml_arange(ggml_ctx, start, stop, step);
+        } else if (tensor->op == GGML_OP_FILL) {
+            const float value = ggml_get_op_params_f32(tensor, 0);
+            tensor_clone = ggml_fill(ggml_ctx, tensor_clone, value);
+        } else if (tensor->op == GGML_OP_SQR) {
+            tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_SQRT) {
+            tensor_clone = ggml_sqrt(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_SIN) {
+            tensor_clone = ggml_sin(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_COS) {
+            tensor_clone = ggml_cos(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_LOG) {
+            tensor_clone = ggml_log(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_TRI) {
+            tensor_clone = ggml_tri(ggml_ctx, src_clone[0], ggml_get_op_params_i32(tensor, 0));
+        } else if (tensor->op == GGML_OP_CLAMP) {
+            const float * params = (const float *)tensor->op_params;
+            tensor_clone = ggml_clamp(ggml_ctx, src_clone[0], params[0], params[1]);
+        } else if (tensor->op == GGML_OP_PAD) {
+            tensor_clone = ggml_pad_ext(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3],
+                                                                tensor->op_params[4], tensor->op_params[5], tensor->op_params[6], tensor->op_params[7]);
+        } else if (tensor->op == GGML_OP_REPEAT) {
+            tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor);
+        } else if (tensor->op == GGML_OP_REPEAT_BACK) {
+            tensor_clone = ggml_repeat_back(ggml_ctx, src_clone[0], tensor);
+        } else if (tensor->op == GGML_OP_ADD) {
+            tensor_clone = ggml_add(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_ACC) {
+            tensor_clone = ggml_acc(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]);
+        } else if (tensor->op == GGML_OP_NORM) {
+            tensor_clone = ggml_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params);
+        } else if (tensor->op == GGML_OP_GROUP_NORM) {
+            const float * float_params = (const float *)tensor->op_params;
+            tensor_clone = ggml_group_norm(ggml_ctx, src_clone[0], tensor->op_params[0], float_params[1]);
+        } else if (tensor->op == GGML_OP_RMS_NORM) {
+            tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params);
+        } else if (tensor->op == GGML_OP_RMS_NORM_BACK) {
+            const float eps = ((float *) tensor->op_params)[0];
+            tensor_clone = ggml_rms_norm_back(ggml_ctx, src_clone[0], src_clone[1], eps);
+        } else if (tensor->op == GGML_OP_SILU_BACK) {
+            tensor_clone = ggml_silu_back(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_L2_NORM) {
+            const float eps = ((float *) tensor->op_params)[0];
+            tensor_clone = ggml_l2_norm(ggml_ctx, src_clone[0], eps);
+        } else if (tensor->op == GGML_OP_SOFT_MAX) {
+            if (tensor->src[1] != nullptr) {
+                const float * params = (const float *)tensor->op_params;
+                tensor_clone = ggml_soft_max_ext(ggml_ctx, src_clone[0], src_clone[1], params[0], params[1]);
             } else {
-                tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+                tensor_clone = ggml_soft_max(ggml_ctx, src_clone[0]);
             }
+        } else if (tensor->op == GGML_OP_SOFT_MAX_BACK) {
+            tensor_clone = ggml_soft_max_ext_back(ggml_ctx, src_clone[0], src_clone[1], ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
+        } else if (tensor->op == GGML_OP_DIAG_MASK_INF) {
+            tensor_clone = ggml_diag_mask_inf(ggml_ctx, src_clone[0], tensor->op_params[0]);
+        } else if (tensor->op == GGML_OP_ROPE || tensor->op == GGML_OP_ROPE_BACK) {
+            const int n_dims      = ((int32_t *) tensor->op_params)[1];
+            const int mode        = ((int32_t *) tensor->op_params)[2];
+            //const int n_ctx_ggml       = ((int32_t *) tensor->op_params)[3];
+            const int n_ctx_orig_ggml  = ((int32_t *) tensor->op_params)[4];
+            const float freq_base       = ((float *) tensor->op_params)[5];
+            const float freq_scale      = ((float *) tensor->op_params)[6];
+            const float ext_factor      = ((float *) tensor->op_params)[7];
+            const float attn_factor     = ((float *) tensor->op_params)[8];
+            const float beta_fast       = ((float *) tensor->op_params)[9];
+            const float beta_slow       = ((float *) tensor->op_params)[10];
+            if (mode & GGML_ROPE_TYPE_MROPE) {
+                int32_t *sections = ((int32_t *) tensor->op_params) + 11;
+                if (tensor->op == GGML_OP_ROPE) {
+                    tensor_clone = ggml_rope_multi(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+                } else {
+                    tensor_clone = ggml_rope_multi_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+                }
+            } else {
+                if (tensor->op == GGML_OP_ROPE) {
+                    tensor_clone = ggml_rope_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+                } else {
+                    tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+                }
+            }
+        } else if (tensor->op == GGML_OP_UNARY) {
+            switch (ggml_get_unary_op(tensor)) {
+            case GGML_UNARY_OP_EXP:
+                tensor_clone = ggml_exp(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_SILU:
+                tensor_clone = ggml_silu(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_GELU:
+                tensor_clone = ggml_gelu(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_GELU_ERF:
+                tensor_clone = ggml_gelu_erf(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_GELU_QUICK:
+                tensor_clone = ggml_gelu_quick(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_RELU:
+                tensor_clone = ggml_relu(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_NEG:
+                tensor_clone = ggml_neg(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_TANH:
+                tensor_clone = ggml_tanh(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_SIGMOID:
+                tensor_clone = ggml_sigmoid(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_HARDSIGMOID:
+                tensor_clone = ggml_hardsigmoid(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_HARDSWISH:
+                tensor_clone = ggml_hardswish(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_ABS:
+                tensor_clone = ggml_abs(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_SOFTPLUS:
+                tensor_clone = ggml_softplus(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_STEP:
+                tensor_clone = ggml_step(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_ROUND:
+                tensor_clone = ggml_round(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_CEIL:
+                tensor_clone = ggml_ceil(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_FLOOR:
+                tensor_clone = ggml_floor(ggml_ctx, src_clone[0]);
+                break;
+            case GGML_UNARY_OP_TRUNC:
+                tensor_clone = ggml_trunc(ggml_ctx, src_clone[0]);
+                break;
+            default:
+                std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
+                GGML_ABORT("fatal error");
+            }
+        } else if (tensor->op == GGML_OP_GLU) {
+            if (src_clone[1] == nullptr) {
+                tensor_clone = ggml_glu(ggml_ctx, src_clone[0], (ggml_glu_op) tensor->op_params[0], tensor->op_params[1]);
+            } else {
+                tensor_clone = ggml_glu_split(ggml_ctx, src_clone[0], src_clone[1], (ggml_glu_op) tensor->op_params[0]);
+            }
+            ggml_set_op_params_i32(tensor_clone, 2, ggml_get_op_params_i32(tensor, 2));
+            ggml_set_op_params_i32(tensor_clone, 3, ggml_get_op_params_i32(tensor, 3));
+        } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
+            if (tensor->src[1] == nullptr) {
+                tensor_clone = ggml_dup(ggml_ctx, src_clone[0]);
+                tensor_clone->type = tensor->type;
+            } else {
+                tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
+            }
+        } else if (tensor->op == GGML_OP_CONT) {
+            tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+        } else if (tensor->op == GGML_OP_RESHAPE) {
+            tensor_clone = ggml_reshape_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+        } else if (tensor->op == GGML_OP_VIEW) {
+            tensor_clone = ggml_view_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]);
+        } else if (tensor->op == GGML_OP_PERMUTE) {
+            int32_t * params = (int32_t *)tensor->op_params;
+            tensor_clone = ggml_permute(ggml_ctx, src_clone[0], params[0], params[1], params[2], params[3]);
+        } else if (tensor->op == GGML_OP_TRANSPOSE) {
+            tensor_clone = ggml_transpose(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_GET_ROWS) {
+            tensor_clone = ggml_get_rows(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_ARGSORT) {
+            tensor_clone = ggml_argsort(ggml_ctx, src_clone[0], (ggml_sort_order) *(int *)tensor->op_params);
+        } else if (tensor->op == GGML_OP_TOP_K) {
+            tensor_clone = ggml_top_k(ggml_ctx, src_clone[0], tensor->ne[0]);
+        } else if (tensor->op == GGML_OP_SUM) {
+            tensor_clone = ggml_sum(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_SUM_ROWS) {
+            tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_CUMSUM) {
+            tensor_clone = ggml_cumsum(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_MEAN) {
+            tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_ARGMAX) {
+            tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_COUNT_EQUAL) {
+            tensor_clone = ggml_count_equal(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_SOLVE_TRI) {
+            tensor_clone = ggml_solve_tri(ggml_ctx, src_clone[0], src_clone[1], true, true, false);
+        } else if (tensor->op == GGML_OP_IM2COL) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t p0 = tensor->op_params[2];
+            const int32_t p1 = tensor->op_params[3];
+            const int32_t d0 = tensor->op_params[4];
+            const int32_t d1 = tensor->op_params[5];
+
+            const bool is_2D = tensor->op_params[6] == 1;
+            tensor_clone = ggml_im2col(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1, is_2D, tensor->type);
+        } else if (tensor->op == GGML_OP_IM2COL_3D) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t s2 = tensor->op_params[2];
+            const int32_t p0 = tensor->op_params[3];
+            const int32_t p1 = tensor->op_params[4];
+            const int32_t p2 = tensor->op_params[5];
+            const int32_t d0 = tensor->op_params[6];
+            const int32_t d1 = tensor->op_params[7];
+            const int32_t d2 = tensor->op_params[8];
+            const int32_t IC = tensor->op_params[9];
+
+            tensor_clone = ggml_im2col_3d(ggml_ctx, src_clone[0], src_clone[1], IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, tensor->type);
+        } else if (tensor->op == GGML_OP_TIMESTEP_EMBEDDING) {
+            const int32_t dim = tensor->op_params[0];
+            const int32_t max_period = tensor->op_params[1];
+            tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period);
+        } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_1D){
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t p0 = tensor->op_params[1];
+            const int32_t d0 = tensor->op_params[2];
+            tensor_clone = ggml_conv_transpose_1d(ggml_ctx, src_clone[0], src_clone[1], s0, p0, d0);
+        } else if (tensor->op == GGML_OP_POOL_2D) {
+            enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
+            const int32_t k0 = tensor->op_params[1];
+            const int32_t k1 = tensor->op_params[2];
+            const int32_t s0 = tensor->op_params[3];
+            const int32_t s1 = tensor->op_params[4];
+            const int32_t p0 = tensor->op_params[5];
+            const int32_t p1 = tensor->op_params[6];
+
+            tensor_clone = ggml_pool_2d(ggml_ctx, src_clone[0], op, k0, k1, s0, s1, p0, p1);
+        } else if (tensor->op == GGML_OP_CONV_2D) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t p0 = tensor->op_params[2];
+            const int32_t p1 = tensor->op_params[3];
+            const int32_t d0 = tensor->op_params[4];
+            const int32_t d1 = tensor->op_params[5];
+            tensor_clone = ggml_conv_2d(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
+        } else if (tensor->op == GGML_OP_CONV_2D_DW) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t p0 = tensor->op_params[2];
+            const int32_t p1 = tensor->op_params[3];
+            const int32_t d0 = tensor->op_params[4];
+            const int32_t d1 = tensor->op_params[5];
+            tensor_clone = ggml_conv_2d_dw_direct(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
+        } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_2D) {
+            const int32_t s = tensor->op_params[0];
+            tensor_clone = ggml_conv_transpose_2d_p0(ggml_ctx, src_clone[0], src_clone[1], s);
+        } else if (tensor->op == GGML_OP_LEAKY_RELU) {
+            const float * op_params = (const float *)tensor->op_params;
+            tensor_clone = ggml_leaky_relu(ggml_ctx, src_clone[0], op_params[0], false);
+        } else if (tensor->op == GGML_OP_RWKV_WKV6) {
+            tensor_clone = ggml_rwkv_wkv6(ggml_ctx, src_clone[0], src_clone[1],
+            src_clone[2], src_clone[3], src_clone[4], src_clone[5]);
+        } else if (tensor->op == GGML_OP_RWKV_WKV7) {
+            tensor_clone = ggml_rwkv_wkv7(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3],
+            src_clone[4], src_clone[5], src_clone[6]);
+        } else if (tensor->op == GGML_OP_OPT_STEP_ADAMW) {
+            src_clone[0]->flags = tensor->src[0]->flags;
+            tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],
+            src_clone[2], src_clone[3], src_clone[4]);
+        } else if (tensor->op == GGML_OP_OPT_STEP_SGD) {
+            src_clone[0]->flags = tensor->src[0]->flags;
+            tensor_clone = ggml_opt_step_sgd(ggml_ctx, src_clone[0], src_clone[1],
+            src_clone[2]);
+        } else if (tensor->op == GGML_OP_ADD_ID) {
+            tensor_clone = ggml_add_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
+        } else if (tensor->op == GGML_OP_SSM_SCAN) {
+            tensor_clone = ggml_ssm_scan(ggml_ctx, src_clone[0], src_clone[1], src_clone[2],
+                                         src_clone[3], src_clone[4], src_clone[5], src_clone[6]);
+        } else if (tensor->op == GGML_OP_SSM_CONV) {
+            tensor_clone = ggml_ssm_conv(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_ROLL) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t s2 = tensor->op_params[2];
+            const int32_t s3 = tensor->op_params[3];
+            tensor_clone = ggml_roll(ggml_ctx, src_clone[0], s0, s1, s2, s3);
         }
-    } else if (tensor->op == GGML_OP_UNARY) {
-        switch (ggml_get_unary_op(tensor)) {
-        case GGML_UNARY_OP_EXP:
-            tensor_clone = ggml_exp(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_SILU:
-            tensor_clone = ggml_silu(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_GELU:
-            tensor_clone = ggml_gelu(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_GELU_ERF:
-            tensor_clone = ggml_gelu_erf(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_GELU_QUICK:
-            tensor_clone = ggml_gelu_quick(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_RELU:
-            tensor_clone = ggml_relu(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_TANH:
-            tensor_clone = ggml_tanh(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_SIGMOID:
-            tensor_clone = ggml_sigmoid(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_HARDSIGMOID:
-            tensor_clone = ggml_hardsigmoid(ggml_ctx, src_clone[0]);
-            break;
-        case GGML_UNARY_OP_HARDSWISH:
-            tensor_clone = ggml_hardswish(ggml_ctx, src_clone[0]);
-            break;
-        default:
+        else {
             std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
             GGML_ABORT("fatal error");
         }
-    } else if (tensor->op == GGML_OP_GLU) {
-        if (src_clone[1] == nullptr) {
-            tensor_clone = ggml_glu(ggml_ctx, src_clone[0], (ggml_glu_op) tensor->op_params[0], tensor->op_params[1]);
-        } else {
-            tensor_clone = ggml_glu_split(ggml_ctx, src_clone[0], src_clone[1], (ggml_glu_op) tensor->op_params[0]);
-        }
-        ggml_set_op_params_i32(tensor_clone, 2, ggml_get_op_params_i32(tensor, 2));
-        ggml_set_op_params_i32(tensor_clone, 3, ggml_get_op_params_i32(tensor, 3));
-    } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
-        if (src1 == nullptr) {
-            tensor_clone = ggml_dup(ggml_ctx, src_clone[0]);
-            tensor_clone->type = tensor->type;
-        } else {
-            tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
-        }
-    } else if (tensor->op == GGML_OP_CONT) {
-        tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
-    } else if (tensor->op == GGML_OP_RESHAPE) {
-        tensor_clone = ggml_reshape_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
-    } else if (tensor->op == GGML_OP_VIEW) {
-        tensor_clone = ggml_view_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]);
-    } else if (tensor->op == GGML_OP_PERMUTE) {
-        int32_t * params = (int32_t *)tensor->op_params;
-        tensor_clone = ggml_permute(ggml_ctx, src_clone[0], params[0], params[1], params[2], params[3]);
-    } else if (tensor->op == GGML_OP_TRANSPOSE) {
-        tensor_clone = ggml_transpose(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_GET_ROWS) {
-        tensor_clone = ggml_get_rows(ggml_ctx, src_clone[0], src_clone[1]);
-    } else if (tensor->op == GGML_OP_ARGSORT) {
-        tensor_clone = ggml_argsort(ggml_ctx, src_clone[0], (ggml_sort_order) *(int *)tensor->op_params);
-    } else if (tensor->op == GGML_OP_SUM) {
-        tensor_clone = ggml_sum(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_SUM_ROWS) {
-        tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_MEAN) {
-        tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_ARGMAX) {
-        tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]);
-    } else if (tensor->op == GGML_OP_COUNT_EQUAL) {
-        tensor_clone = ggml_count_equal(ggml_ctx, src_clone[0], src_clone[1]);
-    } else if (tensor->op == GGML_OP_IM2COL) {
-        const int32_t s0 = tensor->op_params[0];
-        const int32_t s1 = tensor->op_params[1];
-        const int32_t p0 = tensor->op_params[2];
-        const int32_t p1 = tensor->op_params[3];
-        const int32_t d0 = tensor->op_params[4];
-        const int32_t d1 = tensor->op_params[5];
-
-        const bool is_2D = tensor->op_params[6] == 1;
-        tensor_clone = ggml_im2col(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1, is_2D, tensor->type);
-    } else if (tensor->op == GGML_OP_IM2COL_3D) {
-        const int32_t s0 = tensor->op_params[0];
-        const int32_t s1 = tensor->op_params[1];
-        const int32_t s2 = tensor->op_params[2];
-        const int32_t p0 = tensor->op_params[3];
-        const int32_t p1 = tensor->op_params[4];
-        const int32_t p2 = tensor->op_params[5];
-        const int32_t d0 = tensor->op_params[6];
-        const int32_t d1 = tensor->op_params[7];
-        const int32_t d2 = tensor->op_params[8];
-        const int32_t IC = tensor->op_params[9];
-
-        tensor_clone = ggml_im2col_3d(ggml_ctx, src_clone[0], src_clone[1], IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, tensor->type);
-    } else if (tensor->op == GGML_OP_TIMESTEP_EMBEDDING) {
-        const int32_t dim = tensor->op_params[0];
-        const int32_t max_period = tensor->op_params[1];
-        tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period);
-    } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_1D){
-        const int32_t s0 = tensor->op_params[0];
-        const int32_t p0 = tensor->op_params[1];
-        const int32_t d0 = tensor->op_params[2];
-        tensor_clone = ggml_conv_transpose_1d(ggml_ctx, src_clone[0], src_clone[1], s0, p0, d0);
-    } else if (tensor->op == GGML_OP_POOL_2D) {
-        enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
-        const int32_t k0 = tensor->op_params[1];
-        const int32_t k1 = tensor->op_params[2];
-        const int32_t s0 = tensor->op_params[3];
-        const int32_t s1 = tensor->op_params[4];
-        const int32_t p0 = tensor->op_params[5];
-        const int32_t p1 = tensor->op_params[6];
-
-        tensor_clone = ggml_pool_2d(ggml_ctx, src_clone[0], op, k0, k1, s0, s1, p0, p1);
-    } else if (tensor->op == GGML_OP_CONV_2D) {
-        const int32_t s0 = tensor->op_params[0];
-        const int32_t s1 = tensor->op_params[1];
-        const int32_t p0 = tensor->op_params[2];
-        const int32_t p1 = tensor->op_params[3];
-        const int32_t d0 = tensor->op_params[4];
-        const int32_t d1 = tensor->op_params[5];
-        tensor_clone = ggml_conv_2d(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
-    } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_2D) {
-        const int32_t s = tensor->op_params[0];
-        tensor_clone = ggml_conv_transpose_2d_p0(ggml_ctx, src_clone[0], src_clone[1], s);
-    } else if (tensor->op == GGML_OP_LEAKY_RELU) {
-        const float * op_params = (const float *)tensor->op_params;
-        tensor_clone = ggml_leaky_relu(ggml_ctx, src_clone[0], op_params[0], false);
-    } else if (tensor->op == GGML_OP_RWKV_WKV6) {
-        tensor_clone = ggml_rwkv_wkv6(ggml_ctx, src_clone[0], src_clone[1],
-        src_clone[2], src_clone[3], src_clone[4], src_clone[5]);
-    } else if (tensor->op == GGML_OP_RWKV_WKV7) {
-        tensor_clone = ggml_rwkv_wkv7(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3],
-        src_clone[4], src_clone[5], src_clone[6]);
-    } else if (tensor->op == GGML_OP_OPT_STEP_ADAMW) {
-        src_clone[0]->flags = src0->flags;
-        tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],
-        src_clone[2], src_clone[3], src_clone[4]);
-    } else if (tensor->op == GGML_OP_OPT_STEP_SGD) {
-        src_clone[0]->flags = src0->flags;
-        tensor_clone = ggml_opt_step_sgd(ggml_ctx, src_clone[0], src_clone[1],
-        src_clone[2]);
-    } else if (tensor->op == GGML_OP_ADD_ID) {
-        tensor_clone = ggml_add_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
-    } else if (tensor->op == GGML_OP_SSM_SCAN) {
-        tensor_clone = ggml_ssm_scan(ggml_ctx, src_clone[0], src_clone[1], src_clone[2],
-                                     src_clone[3], src_clone[4], src_clone[5], src_clone[6]);
-    } else if (tensor->op == GGML_OP_SSM_CONV) {
-        tensor_clone = ggml_ssm_conv(ggml_ctx, src_clone[0], src_clone[1]);
-    }
-    else {
-        std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
-        GGML_ABORT("fatal error");
+        cloned_tensors[tensor] = tensor_clone;
     }
 
     ggml_cgraph * cgraph_cpu = ggml_new_graph(ggml_ctx);
@@ -14507,10 +15119,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
     memcpy(comp_result, tensor_clone->data, comp_size);
     memcpy(comp_nb, tensor_clone->nb, sizeof(size_t) * GGML_MAX_DIMS);
 
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (src_buffer[i] != nullptr) {
-            free(src_buffer[i]);
-        }
+    for (auto m : cloned_mallocs) {
+        free(m);
     }
 
     ggml_free(ggml_ctx);
@@ -14519,15 +15129,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
 }
 
 static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
-    ggml_tensor * tensor = cgraph->nodes[tensor_idx];
+    ggml_tensor * tensor = cgraph->nodes[tensor_idx + ctx->num_additional_fused_ops];
     if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
-    if (ctx->num_additional_fused_ops == 1 &&
-        tensor->op == GGML_OP_RMS_NORM &&
-        cgraph->nodes[tensor_idx + 1]->op == GGML_OP_MUL) {
-        tensor = cgraph->nodes[tensor_idx + 1];
-    }
 
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
         return;
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp
new file mode 100644
index 00000000..07bd1c18
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp
@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    data_d[i] = D_TYPE(abs(float(data_a[i])));
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp
new file mode 100644
index 00000000..db60725d
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp
@@ -0,0 +1,28 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= p.ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03);
+
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset()]));
+
+        idx += num_threads;
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp
new file mode 100644
index 00000000..f4936eea
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp
@@ -0,0 +1,20 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    // p.param1 = start, p.param2 = step
+    float value = p.param1 + p.param2 * float(i);
+    data_d[i] = D_TYPE(value);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
index c4e68bc0..0fc2b9b7 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
@@ -4,28 +4,27 @@
 #include "types.glsl"
 
 layout(constant_id = 0) const int BLOCK_SIZE = 1024;
-layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
+layout(constant_id = 1) const int NCOLS_PADDED_LOG2 = 10;
 #define ASC 0
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1)          buffer D {int data_d[];};
+layout (binding = 2) writeonly buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
+    uint ncols_padded;
+    uint ncols_padded_log2;
     uint nrows;
     uint order;
+    uint outer_start;
+    uint outer_end;
+    uint inner_start;
+    uint inner_end;
 } p;
 
-shared int dst_row[BLOCK_SIZE];
-shared A_TYPE a_sh[BLOCK_SIZE];
-
-void swap(uint idx0, uint idx1) {
-    int tmp = dst_row[idx0];
-    dst_row[idx0] = dst_row[idx1];
-    dst_row[idx1] = tmp;
-}
+shared ivec2 dst_row[BLOCK_SIZE];
 
 void argsort(bool needs_bounds_check, const uint row) {
     // bitonic sort
@@ -34,11 +33,10 @@ void argsort(bool needs_bounds_check, const uint row) {
     const uint row_offset = row * p.ncols;
 
     // initialize indices
-    dst_row[col] = col;
-    a_sh[col] = data_a[row_offset + col];
+    dst_row[col] = ivec2(col, floatBitsToInt(data_a[row_offset + col]));
     barrier();
 
-    uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
+    uint num_outer_loop_iters = NCOLS_PADDED_LOG2;
     [[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
         uint num_inner_loop_iters = outer_idx + 1;
         [[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
@@ -47,14 +45,15 @@ void argsort(bool needs_bounds_check, const uint row) {
             int idx_0 = (col & k) == 0 ? col : ixj;
             int idx_1 = (col & k) == 0 ? ixj : col;
 
-            int sh_idx_0 = dst_row[idx_0];
-            int sh_idx_1 = dst_row[idx_1];
-            bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
-            bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
+            ivec2 sh_idx_0 = dst_row[idx_0];
+            ivec2 sh_idx_1 = dst_row[idx_1];
+            bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.ncols : false;
+            bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= p.ncols : false;
 
             if ((idx_0_oob ||
-                (!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
-                swap(idx_0, idx_1);
+                (!idx_1_oob && intBitsToFloat(sh_idx_0.y) > intBitsToFloat(sh_idx_1.y))) && (ixj > col)) {
+                dst_row[idx_0] = sh_idx_1;
+                dst_row[idx_1] = sh_idx_0;
             }
 
             barrier();
@@ -63,9 +62,9 @@ void argsort(bool needs_bounds_check, const uint row) {
 
     if (col < p.ncols) {
         if (p.order == ASC) {
-            data_d[row_offset + col] = dst_row[col];
+            data_d[row_offset + col] = dst_row[col].x;
         } else {
-            data_d[row_offset + p.ncols - col - 1] = dst_row[col];
+            data_d[row_offset + p.ncols - col - 1] = dst_row[col].x;
         }
     }
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp
new file mode 100644
index 00000000..920bac6b
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp
@@ -0,0 +1,114 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_memory_scope_semantics : enable
+#pragma use_vulkan_memory_model
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int WG_UNROLL_FACTOR = 2;
+#define ASC 0
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) workgroupcoherent buffer B {ivec2 tmp_idx[];};
+layout (binding = 2) workgroupcoherent buffer D {int data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint ncols;
+    uint ncols_padded;
+    uint ncols_padded_log2;
+    uint nrows;
+    uint order;
+    uint outer_start;
+    uint outer_end;
+    uint inner_start;
+    uint inner_end;
+} p;
+
+void argsort(bool needs_bounds_check, const uint row) {
+    // bitonic sort
+    int col = int(gl_GlobalInvocationID.x);
+    col = (col % BLOCK_SIZE) + (col / BLOCK_SIZE) * BLOCK_SIZE * WG_UNROLL_FACTOR;
+
+    const uint row_offset = row * p.ncols;
+    uint idx_offset = row * p.ncols_padded;
+
+    bool need_barrier = false;
+
+    // initialize indices
+    if (p.outer_start == 0 && p.inner_start == 0) {
+        [[unroll]] for (int u = 0; u < WG_UNROLL_FACTOR; ++u) {
+            uint c = u*BLOCK_SIZE + col;
+            if (c < p.ncols_padded) {
+                ivec2 v = ivec2(c, floatBitsToInt(data_a[row_offset + c]));
+                tmp_idx[idx_offset + c] = v;
+            }
+        }
+        need_barrier = true;
+    }
+
+    [[unroll]] for (uint outer_idx = p.outer_start, k = (2 << outer_idx); outer_idx < p.outer_end; k *= 2, outer_idx++) {
+        uint inner_end = min(p.inner_end, outer_idx + 1);
+        for (uint j = k >> (p.inner_start + 1), inner_idx = p.inner_start; inner_idx < inner_end; j /= 2, inner_idx++) {
+            if (need_barrier) {
+                controlBarrier(gl_ScopeWorkgroup, gl_ScopeWorkgroup, gl_StorageSemanticsBuffer, gl_SemanticsAcquireRelease);
+            }
+            need_barrier = true;
+            [[unroll]] for (int u = 0; u < WG_UNROLL_FACTOR; ++u) {
+                int c = u*BLOCK_SIZE + col;
+                const int ixj = int(c ^ j);
+
+                if (ixj < c) {
+                    continue;
+                }
+
+                int idx_0 = (c & k) == 0 ? c : ixj;
+                int idx_1 = (c & k) == 0 ? ixj : c;
+
+                ivec2 sh_idx_0 = tmp_idx[idx_offset + idx_0];
+                ivec2 sh_idx_1 = tmp_idx[idx_offset + idx_1];
+                bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.ncols : false;
+                bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= p.ncols : false;
+
+                if ((idx_0_oob ||
+                    (!idx_1_oob && intBitsToFloat(sh_idx_0.y) > intBitsToFloat(sh_idx_1.y)))) {
+                    tmp_idx[idx_offset + idx_0] = sh_idx_1;
+                    tmp_idx[idx_offset + idx_1] = sh_idx_0;
+                }
+            }
+        }
+    }
+
+    if (p.outer_end == p.ncols_padded_log2 &&
+        p.inner_end >= p.ncols_padded_log2 + 1) {
+        controlBarrier(gl_ScopeWorkgroup, gl_ScopeWorkgroup, gl_StorageSemanticsBuffer, gl_SemanticsAcquireRelease);
+        [[unroll]] for (int u = 0; u < WG_UNROLL_FACTOR; ++u) {
+            uint c = u*BLOCK_SIZE + col;
+            if (c < p.ncols) {
+                if (p.order == ASC) {
+                    data_d[row_offset + c] = tmp_idx[idx_offset + c].x;
+                } else {
+                    data_d[row_offset + p.ncols - c - 1] = tmp_idx[idx_offset + c].x;
+                }
+            }
+        }
+    }
+}
+
+void main() {
+    if (p.ncols == p.ncols_padded) {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    } else {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp
new file mode 100644
index 00000000..0028d372
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(ceil(x));
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
index 0367e80b..e9bdbf7d 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
@@ -62,14 +62,8 @@ layout(push_constant) uniform parameter {
     uint32_t nb3;
 
     // fastdiv helper values
-    uint32_t KWmp;   uint32_t KWL;
-    uint32_t KWKHmp; uint32_t KWKHL;
     uint32_t OWmp;   uint32_t OWL;
     uint32_t OWOHmp; uint32_t OWOHL;
-#ifdef TRANSPOSE
-    uint32_t s0mp; uint32_t s0L;
-    uint32_t s1mp; uint32_t s1L;
-#endif
 }
 
 p;
@@ -84,6 +78,15 @@ layout(constant_id = 4) const uint TS_K            = 8;
 layout(constant_id = 5) const uint use_collectives = 1;
 layout(constant_id = 6) const uint SHMEM_PAD       = 4;
 
+layout(constant_id = 7)  const uint s0             = 1;
+layout(constant_id = 8)  const uint s1             = 1;
+layout(constant_id = 9)  const uint p0             = 0;
+layout(constant_id = 10) const uint p1             = 0;
+layout(constant_id = 11) const uint d0             = 1;
+layout(constant_id = 12) const uint d1             = 1;
+layout(constant_id = 13) const uint KW             = 1;
+layout(constant_id = 14) const uint KH             = 1;
+
 uint32_t       tid     = gl_LocalInvocationID.x;
 const uint32_t WG_SIZE = gl_WorkGroupSize.x;
 
@@ -92,7 +95,7 @@ uint splitWork(uint work_size, uint block_size) {
 }
 
 uint32_t K   = p.Cout;
-uint32_t CRS = p.Cin * p.KH * p.KW;
+uint32_t CRS = p.Cin * KH * KW;
 uint32_t NPQ = p.N * p.OH * p.OW;
 
 uint32_t n_elems_out = K * NPQ;
@@ -187,7 +190,7 @@ void main() {
     }
 #endif
     /* Advance block in CRS dim */
-    for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
+    [[dont_unroll]] for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
         uint32_t CRS_idx_a;
         uint32_t Cin_idx_a;
         uint32_t KH_idx_a;
@@ -200,10 +203,10 @@ void main() {
         uint32_t cached_KW_idx;
         if (use_collectives == 1) {
             cached_CRS_idx                = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
-            cached_Cin_idx                = fastdiv(cached_CRS_idx, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-            uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
-            cached_KH_idx                 = fastdiv(cached_CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-            cached_KW_idx                 = cached_CRS_remainder - cached_KH_idx * p.KW;
+            cached_Cin_idx                = cached_CRS_idx / (KW * KH);
+            uint32_t cached_CRS_remainder = cached_CRS_idx % (KW * KH);
+            cached_KH_idx                 = cached_CRS_remainder / KW;
+            cached_KW_idx                 = cached_CRS_remainder % KW;
 
             CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
             Cin_idx_a = subgroupShuffle(cached_Cin_idx, Ac);
@@ -211,21 +214,21 @@ void main() {
             KW_idx_a  = subgroupShuffle(cached_KW_idx, Ac);
         } else {
             CRS_idx_a              = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-            Cin_idx_a              = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-            uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-            KH_idx_a               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-            KW_idx_a               = CRS_remainder - KH_idx_a * p.KW;
+            Cin_idx_a              = CRS_idx_a / (KW * KH);
+            uint32_t CRS_remainder = CRS_idx_a % (KW * KH);
+            KH_idx_a               = CRS_remainder / KW;
+            KW_idx_a               = CRS_remainder % KW;
         }
 #else
         CRS_idx_a     = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-        Cin_idx_a     = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH); / (p.KW * p.KH);
-        CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-        KH_idx_a      = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-        KW_idx_a      = CRS_remainder - KH_idx_a * p.KW;
+        Cin_idx_a     = CRS_idx_a / (KW * KH);
+        CRS_remainder = CRS_idx_a % (KW * KH);
+        KH_idx_a      = CRS_remainder / KW;
+        KW_idx_a      = CRS_remainder % KW;
 #endif
 
         /* Load kernel to A_block: (BS_K x BS_CRS)*/
-        for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
             uint32_t B_ly    = r_offset + Ar;
             uint32_t B_lx    = Ac;
             uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
@@ -262,27 +265,27 @@ void main() {
                 KW_idx_b  = subgroupShuffle(cached_KW_idx, r_offset + Br);
             } else {
                 CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-                Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-                uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-                KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-                KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+                Cin_idx_b              = CRS_idx_b / (KW * KH);
+                uint32_t CRS_remainder = CRS_idx_b % (KW * KH);
+                KH_idx_b               = CRS_remainder / KW;
+                KW_idx_b               = CRS_remainder % KW;
             }
 #else
             CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-            Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
-            uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-            KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
-            KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+            Cin_idx_b              = CRS_idx_b / (KW * KH);
+            uint32_t CRS_remainder = CRS_idx_b % (KW * KH);
+            KH_idx_b               = CRS_remainder / KW;
+            KW_idx_b               = CRS_remainder % KW;
 #endif
 
 #ifdef TRANSPOSE
-            uint32_t H_idx_x_s1 = OH_idx - KH_idx_b * p.d1 + p.p1;
-            uint32_t W_idx_x_s0 = OW_idx - KW_idx_b * p.d0 + p.p0;
-            uint32_t H_idx = fastdiv(H_idx_x_s1, p.s1mp, p.s1L);
-            uint32_t W_idx = fastdiv(W_idx_x_s0, p.s0mp, p.s0L);
+            uint32_t H_idx_x_s1 = OH_idx - KH_idx_b * d1 + p1;
+            uint32_t W_idx_x_s0 = OW_idx - KW_idx_b * d0 + p0;
+            uint32_t H_idx = H_idx_x_s1 / s1;
+            uint32_t W_idx = W_idx_x_s0 / s0;
 #else
-            uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
-            uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
+            uint32_t H_idx = OH_idx * s1 + KH_idx_b * d1 - p1;
+            uint32_t W_idx = OW_idx * s0 + KW_idx_b * d0 - p0;
 #endif
             uint32_t src_idx =
                 min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
@@ -290,7 +293,7 @@ void main() {
             if (CRS_idx_b >= CRS || NPQ_idx >= NPQ
                 || H_idx >= p.H || W_idx >= p.W // Lower bound checks aren't necessary. (idx >= 0x80000000 for such case)
 #ifdef TRANSPOSE
-                || (H_idx_x_s1 - H_idx * p.s1 != 0) || (W_idx_x_s0 - W_idx * p.s0 != 0)
+                || (H_idx_x_s1 - H_idx * s1 != 0) || (W_idx_x_s0 - W_idx * s0 != 0)
 #endif
                 ) {
                 val = 0.0;
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp
new file mode 100644
index 00000000..220ccc91
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp
@@ -0,0 +1,67 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+// workgroup does 32x32 tile, but uses 32x8 threads
+#define TILE_DIM 32
+layout(local_size_x = 32, local_size_y = 8, local_size_z = 1) in;
+
+shared uint sh[TILE_DIM][TILE_DIM + 1];
+
+void iter(uvec3 wg_id) {
+    const uint tile_col = wg_id.x;
+    const uint tile_row = wg_id.y;
+
+    const uint tid_col = gl_LocalInvocationID.x;
+    const uint tid_row = gl_LocalInvocationID.y;
+
+    const uint i2 = wg_id.z % p.ne12;
+    const uint i3 = wg_id.z / p.ne12;
+    const uint i02 = i2;
+    const uint i03 = i3;
+
+    // The workgroup does TILE_DIM x TILE_DIM, but swaps the LSBs of the
+    // src coords to make memory accesses contiguous, dst has tid.x in i0,
+    // src has tid.x in i01
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i00 = tile_col * TILE_DIM + tid_row + 8 * y;
+        const uint i01 = tile_row * TILE_DIM + tid_col;
+        if (i00 < p.ne00 && i01 < p.ne01 && i02 < p.ne02 && i03 < p.ne03) {
+            const uint src_idx = i00 * p.nb00 + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+            sh[tid_row + 8 * y][tid_col] = uint(data_a[get_aoffset() + src_idx]);
+        }
+    }
+
+    barrier();
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i0 = tile_col * TILE_DIM + tid_col;
+        const uint i1 = tile_row * TILE_DIM + tid_row + 8 * y;
+        if (i0 < p.ne10 && i1 < p.ne11 && i2 < p.ne12 && i3 < p.ne13) {
+            const uint dst_idx = i0 * p.nb10 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+            // load transposed
+            data_d[get_doffset() + dst_idx] = D_TYPE(sh[tid_col][tid_row + 8 * y]);
+        }
+    }
+}
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+void main() {
+    uint z = gl_WorkGroupID.z;
+    uint y = gl_WorkGroupID.y;
+    bool need_barrier = false;
+    for (uint z = gl_WorkGroupID.z; z < p.ne12 * p.ne13; z += gl_NumWorkGroups.z) {
+        for (uint y = gl_WorkGroupID.y; y < CEIL_DIV(p.ne11, TILE_DIM); y += gl_NumWorkGroups.y) {
+            for (uint x = gl_WorkGroupID.x; x < CEIL_DIV(p.ne10, TILE_DIM); x += gl_NumWorkGroups.x) {
+                if (need_barrier) {
+                    barrier();
+                }
+                need_barrier = true;
+                iter(uvec3(x, y, z));
+            }
+        }
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp
new file mode 100644
index 00000000..a4c8fc35
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp
@@ -0,0 +1,69 @@
+#version 450
+
+#include "types.glsl"
+#include "sum_rows.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 128;
+layout (constant_id = 1) const uint SUBGROUP_SIZE = 32;
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+shared FLOAT_TYPE partial[BLOCK_SIZE / SUBGROUP_SIZE];
+shared FLOAT_TYPE last_sum;
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    const uint i03 = fastdiv(row, p.ne0_12mp, p.ne0_12L);
+    const uint i03_offset = i03 * p.ne01*p.ne02;
+    const uint i02 = fastdiv(row - i03_offset, p.ne0_1mp, p.ne0_1L);
+    const uint i01 = row - i03_offset - i02*p.ne01;
+
+    const uint src_idx = get_aoffset() + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+    const uint dst_idx = get_doffset() + i01 * p.nb11 + i02 * p.nb12 + i03 * p.nb13;
+
+    uint subgroup_id = tid / SUBGROUP_SIZE;
+
+    if (tid == 0) {
+        last_sum = 0;
+    }
+
+    uint col = tid;
+    uint num_iter = CEIL_DIV(p.n_cols, BLOCK_SIZE);
+    for (int i = 0; i < num_iter; ++i) {
+        FLOAT_TYPE v = 0;
+        if (col < p.n_cols) {
+            v = FLOAT_TYPE(data_a[src_idx + col]);
+        }
+        v = subgroupInclusiveAdd(v);
+
+        // Store the largest partial sum for each subgroup, then add the partials for all
+        // lower subgroups and the final partial sum from the previous iteration.
+        if (gl_SubgroupInvocationID == SUBGROUP_SIZE - 1) {
+            partial[subgroup_id] = v;
+        }
+        barrier();
+        for (int j = 0; j < subgroup_id; ++j) {
+            v += partial[j];
+        }
+        v += last_sum;
+        barrier();
+        if (tid == BLOCK_SIZE - 1) {
+            last_sum = v;
+        }
+        if (col < p.n_cols) {
+            data_d[dst_idx + col] = D_TYPE(v);
+        }
+        col += BLOCK_SIZE;
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
index 09676a62..70ee542d 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl
@@ -4,13 +4,6 @@
 
 #include "types.glsl"
 
-#if defined(A_TYPE_PACKED16)
-layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
-#endif
-#if defined(A_TYPE_PACKED32)
-layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
-#endif
-
 #if defined(DATA_A_F32)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp
new file mode 100644
index 00000000..a56be76c
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp
@@ -0,0 +1,19 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    // p.param1 = fill value
+    data_d[i] = D_TYPE(p.param1);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
index 2255f9c1..4bef48b0 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
@@ -7,6 +7,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
 #extension GL_KHR_shader_subgroup_shuffle : enable
+#extension GL_KHR_shader_subgroup_vote : enable
 
 #include "types.glsl"
 #include "flash_attn_base.glsl"
@@ -108,6 +109,38 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
+
+            float max_mask = NEG_FLT_MAX_OVER_2;
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) % Bc;
+                uint32_t r = (idx + tid) / Bc;
+                if (idx + tid < Bc * Br) {
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
+                        float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
+                        masksh[c][r] = m;
+                        max_mask = max(max_mask, m);
+                    } else {
+                        masksh[c][r] = float(0);
+                    }
+                }
+            }
+            // skip the block if the mask is entirely -inf
+            bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+            barrier();
+            if (gl_SubgroupInvocationID == 0) {
+                tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+            }
+            barrier();
+            [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                max_mask = max(max_mask, tmpsh[s]);
+            }
+            if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                continue;
+            }
+        }
+
         float Sf[Br][cols_per_thread];
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
@@ -153,21 +186,6 @@ void main() {
         }
 
         if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
-            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
-
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
-                uint32_t c = (idx + tid) % Bc;
-                uint32_t r = (idx + tid) / Bc;
-                if (idx + tid < Bc * Br) {
-                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
-                        masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
-                    } else {
-                        masksh[c][r] = float(0);
-                    }
-                }
-            }
-            barrier();
-
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
                 [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                     float mvf = masksh[c * cols_per_iter + col_tid][r];
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
index 8699fa6c..cd82e4ab 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
@@ -7,6 +7,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
 #extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_vote : enable
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_cooperative_matrix : enable
 
@@ -148,6 +149,37 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
+        float mask_cache[Bc * Br / WorkGroupSize];
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
+
+            float max_mask = NEG_FLT_MAX_OVER_2;
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) % Bc;
+                uint32_t r = (idx + tid) / Bc;
+                if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
+                        float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
+                        mask_cache[idx / WorkGroupSize] = m;
+                        max_mask = max(max_mask, m);
+                    }
+                }
+            }
+            // skip the block if the mask is entirely -inf
+            bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+            barrier();
+            if (gl_SubgroupInvocationID == 0) {
+                tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+            }
+            barrier();
+            [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                max_mask = max(max_mask, tmpsh[s]);
+            }
+            if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                continue;
+            }
+        }
+
         [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
             uint32_t d = (idx + tid) % (HSK / 4);
             uint32_t c = (idx + tid) / (HSK / 4);
@@ -208,7 +240,8 @@ void main() {
                 uint32_t r = (idx + tid) / Bc;
                 if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
                     if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
-                        sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]));
+                        float f = mask_cache[idx / WorkGroupSize];
+                        sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * f);
                     }
                 }
             }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
index fcfc60a8..9a719963 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -29,6 +29,10 @@ ACC_TYPE maxReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
     return max(x, y);
 }
 
+float16_t maxReduceFp16(const in float16_t x, const in float16_t y) {
+    return max(x, y);
+}
+
 ACC_TYPE smearReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
     return x;
 }
@@ -142,6 +146,44 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
+            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
+
+            if (nem1_bounds_check) {
+                tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+                tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
+                tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
+                tensorLayoutM = setTensorLayoutClampValueNV(tensorLayoutM, 0xfc00); // -inf in float16_t
+
+                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
+
+                coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+
+                // skip the block if the mask is entirely -inf
+                coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
+                if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
+            } else {
+                tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
+                // Don't clamp against nem1 when GQA is enabled
+                uint32_t m_height = p.gqa_ratio > 1 ? ~0 : p.nem1;
+                tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, m_height, KV);
+                tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
+
+                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
+
+                coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+
+                // skip the block if the mask is entirely -inf
+                coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
+                if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
+            }
+        }
+
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
         coopmat<float16_t, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
@@ -158,31 +200,7 @@ void main() {
         }
 
         if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
-            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
-
-            if (nem1_bounds_check) {
-                tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
-                tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
-                tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
-
-                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
-
-                coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
-
-                S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
-            } else {
-                tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
-                // Don't clamp against nem1 when GQA is enabled
-                uint32_t m_height = p.gqa_ratio > 1 ? ~0 : p.nem1;
-                tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, m_height, KV);
-                tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
-
-                coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
-
-                coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
-
-                S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
-            }
+            S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
         }
 
         // Clear padding elements to -inf, so they don't contribute to rowmax
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp
new file mode 100644
index 00000000..20017eb1
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(floor(x));
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl
index 99595fc6..ba7909c4 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl
@@ -3,6 +3,9 @@
 
 #include "rte.glsl"
 #include "utils.glsl"
+#if RMS_NORM_ROPE_FUSION
+#include "rope_params.glsl"
+#endif
 
 layout (push_constant) uniform parameter
 {
@@ -12,11 +15,23 @@ layout (push_constant) uniform parameter
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
     uint misalign_offsets;
     float param1; float param2; int param3;
+#if RMS_NORM_ROPE_FUSION
+    rope_params rope;
+#endif
 } p;
 
+#if !RMS_NORM_ROPE_FUSION
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+#endif
 
 // true if src0/src1 are the same shape and the indices can be reused without additional modulus
 layout(constant_id = 0) const bool norepeat = false;
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl
index 8dc9d360..cc181fda 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl
@@ -18,6 +18,13 @@ layout (push_constant) uniform parameter
 } p;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 uint get_idx() {
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/log.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/log.comp
new file mode 100644
index 00000000..ff2812d3
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/log.comp
@@ -0,0 +1,18 @@
+#version 450
+
+#include "rte.glsl"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const float val = float(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(log(val));
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
index 9a03925c..b3c96576 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
@@ -3,6 +3,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
 #include "mul_mat_vec_base.glsl"
+#include "dequant_funcs.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl
index 450dee04..cfc8b0c7 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl
@@ -11,28 +11,7 @@
 #define EXPERT_COUNT 8
 #endif
 
-#include "types.glsl"
-
-#ifndef MMQ
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-#else
-layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
-#endif
-
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-#ifdef B_TYPE_VEC2
-layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
-#endif
-#ifdef B_TYPE_VEC4
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
-#endif
-
-layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
-#ifdef MUL_MAT_ID
-layout (binding = 3) readonly buffer IDS {int data_ids[];};
-#endif
-
-#include "dequant_funcs.glsl"
+#include "mul_mat_vec_iface.glsl"
 
 layout (push_constant) uniform parameter
 {
@@ -45,6 +24,8 @@ layout (push_constant) uniform parameter
     uint batch_stride_b;
     uint batch_stride_d;
 
+    uint fusion_flags;
+
 #ifdef MUL_MAT_ID
     uint nei0;
     uint ne11;
@@ -56,6 +37,10 @@ layout (push_constant) uniform parameter
 #endif
 } p;
 
+#ifdef MUL_MAT_ID
+uint expert_id;
+#endif
+
 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
     const uint expert_idx = gl_GlobalInvocationID.y;
@@ -75,7 +60,7 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
         batch_idx_a = i03 * p.ne02 + i02;
     }
 #else
-    const uint expert_id = data_ids[expert_idx];
+    expert_id = data_ids[expert_idx];
 #endif
 
     a_offset =
@@ -113,6 +98,26 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+#ifdef MUL_MAT_ID
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                }
+#else
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse1[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+#endif
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -148,6 +153,26 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                 [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
                     temp[j][n] += tmpsh[j][n][s];
                 }
+#ifdef MUL_MAT_ID
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                }
+#else
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse1[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+#endif
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -173,6 +198,26 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+#ifdef MUL_MAT_ID
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                }
+#else
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_fuse0[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_fuse1[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+#endif
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
             }
         }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl
new file mode 100644
index 00000000..337dbd79
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl
@@ -0,0 +1,35 @@
+#include "types.glsl"
+
+#define MAT_VEC_FUSION_FLAGS_BIAS0 0x1
+#define MAT_VEC_FUSION_FLAGS_BIAS1 0x2
+#define MAT_VEC_FUSION_FLAGS_SCALE0 0x4
+#define MAT_VEC_FUSION_FLAGS_SCALE1 0x8
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(A_TYPE_VEC4)
+layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
+#endif
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+#ifdef B_TYPE_VEC2
+layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+#endif
+#ifdef B_TYPE_VEC4
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#endif
+
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+layout (binding = 3) readonly buffer Fuse0 {D_TYPE data_fuse0[];};
+layout (binding = 4) readonly buffer Fuse1 {D_TYPE data_fuse1[];};
+
+#ifdef MUL_MAT_ID
+layout (binding = 5) readonly buffer IDS {int data_ids[];};
+#endif
+
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
index 638878d9..beea5296 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
@@ -8,12 +8,7 @@
 
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
-
-layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#include "mul_mat_vec_iface.glsl"
 
 layout (push_constant) uniform parameter
 {
@@ -29,6 +24,7 @@ layout (push_constant) uniform parameter
     uint nb03;
     uint nb13;
     uint nb23;
+    uint fusion_flags;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -117,6 +113,12 @@ void main() {
     }
 
     if (tid == 0) {
-        dst[idst] = tmp[0];
+        if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+            tmp[0] += FLOAT_TYPE(data_fuse0[idst]);
+        }
+        if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+            tmp[0] += FLOAT_TYPE(data_fuse1[idst]);
+        }
+        data_d[idst] = tmp[0];
     }
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp
index 7aa070ee..32628c6e 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp
@@ -10,12 +10,7 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
-
-layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#include "mul_mat_vec_iface.glsl"
 
 layout(constant_id = 0) const int BLOCK_SIZE = 32;
 // gqa_ratio is in the range [1,8]
@@ -29,6 +24,7 @@ layout (push_constant) uniform parameter
     uint nchannels_y;
     uint b_offset;
     uint d_offset;
+    uint fusion_flags;
 } p;
 
 #if !USE_SUBGROUP_ADD
@@ -148,7 +144,13 @@ void main() {
         [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
             // dst is not transposed and not permuted
             const uint idst = (channel + c)*nrows_dst + row_dst;
-            dst[idst] = temp[c];
+            if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                temp[c] += FLOAT_TYPE(data_fuse0[idst]);
+            }
+            if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                temp[c] += FLOAT_TYPE(data_fuse1[idst]);
+            }
+            data_d[idst] = temp[c];
         }
     }
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp
index 64293f6e..15f005be 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp
@@ -10,60 +10,56 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
+#if defined(DATA_A_QUANT_LEGACY) || defined(DATA_A_MXFP4)
 #define K_PER_ITER 8
-
-#include "mul_mmq_funcs.glsl"
+#elif defined(DATA_A_QUANT_K)
+#define K_PER_ITER 16
+#else
+#error unimplemented
+#endif
 
 uint a_offset, b_offset, d_offset;
 
-int32_t cache_b_qs[2];
+int32_t cache_b_qs[K_PER_ITER / 4];
 vec2 cache_b_ds;
 
+#include "mul_mat_vecq_funcs.glsl"
+
 void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i) {
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         const uint col = i*BLOCK_SIZE + tid*K_PER_ITER;
 
         // Preload data_b block
         const uint b_block_idx = (j*p.batch_stride_b + col) / QUANT_K_Q8_1 + b_offset;
-        const uint b_qs_idx = tid % 4;
+        const uint b_qs_idx = tid % (32 / K_PER_ITER);
         const uint b_block_idx_outer = b_block_idx / 4;
         const uint b_block_idx_inner = b_block_idx % 4;
         cache_b_ds = vec2(data_b[b_block_idx_outer].ds[b_block_idx_inner]);
 
 #if QUANT_R == 2
+        // Assumes K_PER_ITER == 8
         cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx];
         cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx + 4];
 #else
+#if K_PER_ITER == 8
         cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2];
         cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2 + 1];
+#elif K_PER_ITER == 16
+        cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4    ];
+        cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4 + 1];
+        cache_b_qs[2] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4 + 2];
+        cache_b_qs[3] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 4 + 3];
+#else
+#error unimplemented
+#endif
 #endif
 
         uint ibi = first_row*p.ncols;
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            const uint a_block_idx = (ibi + col)/QUANT_K + a_offset;
+            const uint a_block_idx = (ibi + col)/QUANT_K_Q8_1 + a_offset;
             ibi += p.ncols;
 
-            int32_t q_sum = 0;
-#if QUANT_R == 2
-            const i32vec2 data_a_qs = repack(a_block_idx, b_qs_idx);
-            q_sum += dotPacked4x8EXT(data_a_qs.x,
-                                     cache_b_qs[0]);
-            q_sum += dotPacked4x8EXT(data_a_qs.y,
-                                     cache_b_qs[1]);
-#else
-            int32_t data_a_qs = repack(a_block_idx, b_qs_idx * 2);
-            q_sum += dotPacked4x8EXT(data_a_qs,
-                                     cache_b_qs[0]);
-            data_a_qs = repack(a_block_idx, b_qs_idx * 2 + 1);
-            q_sum += dotPacked4x8EXT(data_a_qs,
-                                     cache_b_qs[1]);
-#endif
-
-#if QUANT_AUXF == 1
-            temp[j][n] += mul_q8_1(q_sum,  get_d(a_block_idx), cache_b_ds, 4);
-#else
-            temp[j][n] += mul_q8_1(q_sum, get_dm(a_block_idx), cache_b_ds, 4);
-#endif
+            temp[j][n] += mmvq_dot_product(a_block_idx, b_qs_idx);
         }
     }
 }
@@ -72,7 +68,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
 
     get_offsets(a_offset, b_offset, d_offset);
-    a_offset /= QUANT_K;
+    a_offset /= QUANT_K_Q8_1;
     b_offset /= QUANT_K_Q8_1;
 
     FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
@@ -102,14 +98,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     unroll_count = 2;
     unrolled_iters = num_iters & ~(unroll_count - 1);
 
-#if K_PER_ITER == 2
-    if ((p.ncols & 1) != 0 &&
-        unrolled_iters == num_iters &&
-        unrolled_iters > 0) {
-        unrolled_iters -= unroll_count;
-    }
-#endif
-
     while (i < unrolled_iters) {
         // Manually partially unroll the loop
         [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
@@ -128,6 +116,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 void main() {
     const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
 
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
     // do NUM_ROWS at a time, unless there aren't enough remaining rows
     if (first_row + NUM_ROWS <= p.stride_d) {
         compute_outputs(first_row, NUM_ROWS);
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl
new file mode 100644
index 00000000..2389ea0b
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl
@@ -0,0 +1,379 @@
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+
+#include "types.glsl"
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+FLOAT_TYPE get_dm(uint ib) {
+    return FLOAT_TYPE(data_a[ib].d);
+}
+#endif
+
+#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
+FLOAT_TYPE_VEC2 get_dm(uint ib) {
+    return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+}
+#endif
+
+#if defined(DATA_A_MXFP4)
+FLOAT_TYPE get_dm(uint ib) {
+    return FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e));
+}
+#endif
+
+#if defined(DATA_A_Q2_K)
+FLOAT_TYPE_VEC2 get_dm(uint ib) {
+    const uint ib_k = ib / 8;
+    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
+}
+#endif
+
+// Each iqs value maps to a 32-bit integer
+#if defined(DATA_A_Q4_0)
+// 2-byte loads for Q4_0 blocks (18 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    return i32vec2( vui       & 0x0F0F0F0F,
+                   (vui >> 4) & 0x0F0F0F0F);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
+}
+#endif
+
+#if defined(DATA_A_Q4_1)
+// 4-byte loads for Q4_1 blocks (20 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const uint32_t vui = data_a_packed32[ib].qs[iqs];
+    return i32vec2( vui       & 0x0F0F0F0F,
+                   (vui >> 4) & 0x0F0F0F0F);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
+}
+#endif
+
+#if defined(DATA_A_Q5_0)
+// 2-byte loads for Q5_0 blocks (22 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    const int32_t qh = int32_t((uint32_t(data_a_packed16[ib].qh[1]) << 16 | data_a_packed16[ib].qh[0]) >> (4 * iqs));
+    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
+                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
+
+    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
+                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
+
+    return i32vec2(v0, v1);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
+}
+#endif
+
+#if defined(DATA_A_Q5_1)
+// 4-byte loads for Q5_1 blocks (24 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
+    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
+                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
+
+    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
+                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
+
+    return i32vec2(v0, v1);
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+// 2-byte loads for Q8_0 blocks (34 bytes)
+int32_t repack(uint ib, uint iqs) {
+    return pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2    ],
+                          data_a_packed16[ib].qs[iqs * 2 + 1]));
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(float(q_sum) * da * dsb.x);
+}
+#endif
+
+#if defined(DATA_A_MXFP4)
+// 1-byte loads for mxfp4 blocks (17 bytes)
+i32vec2 repack(uint ib, uint iqs) {
+    const uint32_t qs = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
+                                      data_a[ib].qs[iqs * 4 + 1],
+                                      data_a[ib].qs[iqs * 4 + 2],
+                                      data_a[ib].qs[iqs * 4 + 3]));
+
+    const u8vec4 i_a0 = unpack8( qs       & 0x0F0F0F0F);
+    const u8vec4 i_a1 = unpack8((qs >> 4) & 0x0F0F0F0F);
+
+    return i32vec2(pack32(i8vec4(kvalues_mxfp4[i_a0.x], kvalues_mxfp4[i_a0.y], kvalues_mxfp4[i_a0.z], kvalues_mxfp4[i_a0.w])),
+                   pack32(i8vec4(kvalues_mxfp4[i_a1.x], kvalues_mxfp4[i_a1.y], kvalues_mxfp4[i_a1.z], kvalues_mxfp4[i_a1.w])));
+}
+
+FLOAT_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return FLOAT_TYPE(da * dsb.x * float(q_sum) * 0.5);
+}
+#endif
+
+#if defined(DATA_A_QUANT_LEGACY) || defined(DATA_A_MXFP4)
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+#if QUANT_R == 2
+    const i32vec2 data_a_qs = repack(ib_a, iqs);
+    q_sum += dotPacked4x8EXT(data_a_qs.x,
+                             cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(data_a_qs.y,
+                             cache_b_qs[1]);
+#else
+    int32_t data_a_qs = repack(ib_a, iqs * 2);
+    q_sum += dotPacked4x8EXT(data_a_qs,
+                             cache_b_qs[0]);
+    data_a_qs = repack(ib_a, iqs * 2 + 1);
+    q_sum += dotPacked4x8EXT(data_a_qs,
+                             cache_b_qs[1]);
+#endif
+
+    // 2 quants per call => divide sums by 8/2 = 4
+    return mul_q8_1(q_sum, get_dm(ib_a), cache_b_ds, 4);
+}
+#endif
+
+#if defined(DATA_A_Q2_K)
+// 4-byte loads for Q2_K blocks (84 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
+
+    return i32vec4((data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x03030303,
+                   (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x03030303,
+                   (data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x03030303,
+                   (data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x03030303);
+}
+
+uint8_t get_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    return data_a[ib_k].scales[iqs_k / 4];
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t sum_d = 0;
+    int32_t sum_m = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const uint8_t scale = get_scale(ib_a, iqs * 4);
+    const vec2 dm = vec2(get_dm(ib_a));
+    const int32_t scale_m = int32_t(scale >> 4) * 0x01010101; // Duplicate 8-bit value across 32-bits.
+
+    sum_d += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[0]);
+
+    sum_d += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[1]);
+
+    sum_d += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[2]);
+
+    sum_d += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]) * (scale & 0xF);
+    sum_m += dotPacked4x8EXT(scale_m, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * (float(dm.x) * float(sum_d) - float(dm.y) * float(sum_m)));
+}
+#endif
+
+#if defined(DATA_A_Q3_K)
+// 2-byte loads for Q3_K blocks (110 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
+    const uint hm_shift = iqs_k / 8;
+
+    // bitwise OR to add 4 if hmask is set, subtract later
+    const i8vec2 vals00 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2    ] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2    ] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals01 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 1] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 1] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals10 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 2] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 2] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals11 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 3] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 3] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals20 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 4] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 4] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals21 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 5] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 5] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals30 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 6] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 6] >> hm_shift) & uint16_t(0x0101)) << 2));
+    const i8vec2 vals31 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 7] >> qs_shift) & uint16_t(0x0303))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 7] >> hm_shift) & uint16_t(0x0101)) << 2));
+
+    return i32vec4(pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y) - int8_t(4)),
+                   pack32(i8vec4(vals10.x, vals10.y, vals11.x, vals11.y) - int8_t(4)),
+                   pack32(i8vec4(vals20.x, vals20.y, vals21.x, vals21.y) - int8_t(4)),
+                   pack32(i8vec4(vals30.x, vals30.y, vals31.x, vals31.y) - int8_t(4)));
+}
+
+float get_d_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+    const uint is = iqs_k / 4;
+
+    const int8_t scale = int8_t(((data_a[ib_k].scales[is % 8      ] >> (4 * (is / 8))) & 0x0F0F) |
+                               (((data_a[ib_k].scales[8 + (is % 4)] >> (2 * (is / 4))) & 0x0303) << 4));
+    return float(data_a[ib_k].d) * float(scale - 32);
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const float d_scale = get_d_scale(ib_a, iqs * 4);
+
+    q_sum += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]);
+    q_sum += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]);
+    q_sum += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * d_scale * float(q_sum));
+}
+#endif
+
+#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
+// 4-byte loads for Q4_K blocks (144 bytes) and Q5_K blocks (176 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint qs_idx = (iqs_k / 16) * 8 + (iqs_k % 8);
+    const uint qs_shift = ((iqs_k % 16) / 8) * 4;
+
+#if defined(DATA_A_Q4_K)
+    const uint32_t vals0 = (data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x0F0F0F0F;
+    const uint32_t vals1 = (data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x0F0F0F0F;
+    const uint32_t vals2 = (data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x0F0F0F0F;
+    const uint32_t vals3 = (data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x0F0F0F0F;
+
+    return i32vec4(vals0, vals1, vals2, vals3);
+#else // defined(DATA_A_Q5_K)
+    const uint qh_idx = iqs;
+    const uint qh_shift = iqs_k / 8;
+
+    return i32vec4(((data_a_packed32[ib_k].qs[qs_idx    ] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx    ] >> qh_shift) & 0x01010101) << 4),
+                   ((data_a_packed32[ib_k].qs[qs_idx + 1] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx + 1] >> qh_shift) & 0x01010101) << 4),
+                   ((data_a_packed32[ib_k].qs[qs_idx + 2] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx + 2] >> qh_shift) & 0x01010101) << 4),
+                   ((data_a_packed32[ib_k].qs[qs_idx + 3] >> qs_shift) & 0x0F0F0F0F) |
+                  (((data_a_packed32[ib_k].qh[qh_idx + 3] >> qh_shift) & 0x01010101) << 4));
+#endif
+}
+
+vec2 get_dm_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+    const uint is = iqs_k / 8;
+    u8vec2 scale_dm;
+    if (is < 4) {
+        scale_dm = u8vec2(data_a[ib_k].scales[is] & 0x3F, data_a[ib_k].scales[is + 4] & 0x3F);
+    } else {
+        scale_dm = u8vec2((data_a[ib_k].scales[is+4] & 0xF) | ((data_a[ib_k].scales[is-4] & 0xC0) >> 2),
+                          (data_a[ib_k].scales[is+4] >>  4) | ((data_a[ib_k].scales[is  ] & 0xC0) >> 2));
+    }
+
+    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const vec2 dm_scale = get_dm_scale(ib_a, iqs * 4);
+
+    q_sum += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]);
+    q_sum += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]);
+    q_sum += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * float(dm_scale.x) * float(q_sum) - float(dm_scale.y) * float(cache_b_ds.y / 2));
+}
+#endif
+
+#if defined(DATA_A_Q6_K)
+// 2-byte loads for Q6_K blocks (210 bytes)
+i32vec4 repack4(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+
+    const uint ql_idx = (iqs_k / 32) * 16 + iqs_k % 16;
+    const uint ql_shift = ((iqs_k % 32) / 16) * 4;
+
+    const uint qh_idx = (iqs_k / 32) * 8 + iqs;
+    const uint qh_shift = ((iqs_k % 32) / 8) * 2;
+
+    const i8vec2 vals00 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals01 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals10 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 2] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 2] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals11 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 3] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 3] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals20 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 4] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 4] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals21 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 5] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 5] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals30 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 6] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 6] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals31 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 7] >> ql_shift) & uint16_t(0x0F0F))) |
+                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 7] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+
+    return i32vec4(pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y)),
+                   pack32(i8vec4(vals10.x, vals10.y, vals11.x, vals11.y)),
+                   pack32(i8vec4(vals20.x, vals20.y, vals21.x, vals21.y)),
+                   pack32(i8vec4(vals30.x, vals30.y, vals31.x, vals31.y)));
+}
+
+float get_d_scale(uint ib, uint iqs) {
+    const uint ib_k = ib / 8;
+    const uint iqs_k = (ib % 8) * 8 + iqs;
+    return float(data_a[ib_k].d) * float(data_a[ib_k].scales[iqs_k / 4]);
+}
+
+FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
+    int32_t q_sum = 0;
+
+    const i32vec4 qs_a = repack4(ib_a, iqs * 4);
+    const float d_scale = get_d_scale(ib_a, iqs * 4);
+
+    q_sum += dotPacked4x8EXT(qs_a.x, cache_b_qs[0]);
+    q_sum += dotPacked4x8EXT(qs_a.y, cache_b_qs[1]);
+    q_sum += dotPacked4x8EXT(qs_a.z, cache_b_qs[2]);
+    q_sum += dotPacked4x8EXT(qs_a.w, cache_b_qs[3]);
+
+    return FLOAT_TYPE(float(cache_b_ds.x) * float(d_scale) * float(q_sum));
+}
+#endif
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index d260969f..5c5251da 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -100,7 +100,6 @@ layout (push_constant) uniform parameter
 layout (constant_id = 0) const uint BLOCK_SIZE = 64;
 layout (constant_id = 1) const uint BM = 64;
 layout (constant_id = 2) const uint BN = 64;
-layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
 layout (constant_id = 4) const uint WM = 32;
 layout (constant_id = 5) const uint WN = 32;
 layout (constant_id = 6) const uint WMITER = 2;
@@ -109,6 +108,14 @@ layout (constant_id = 8) const uint TN = 2;
 layout (constant_id = 9) const uint TK = 1;  // Only needed for coopmat
 layout (constant_id = 10) const uint WARP = 32;
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#define BK 32
+#define BK_STEP 4
+#else
+layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
+#define BK_STEP 2
+#endif
+
 #ifdef COOPMAT
 #define SHMEM_STRIDE (BK / 2 + 4)
 #else
@@ -244,8 +251,13 @@ void main() {
     }
 #else
     ACC_TYPE_VEC2 sums[WMITER * TM * WNITER * TN/2];
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+    FLOAT_TYPE_VEC4 cache_a[WMITER * TM];
+    FLOAT_TYPE_VEC4 cache_b;
+#else
     FLOAT_TYPE_VEC2 cache_a[WMITER * TM];
     FLOAT_TYPE_VEC2 cache_b;
+#endif
 
     [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
         sums[i] = ACC_TYPE_VEC2(0.0f, 0.0f);
@@ -283,24 +295,41 @@ void main() {
             }
         }
 #else
-        [[unroll]] for (uint i = 0; i < BK / 2; i++) {
+        [[unroll]] for (uint i = 0; i < BK / BK_STEP; i++) {
             // Load from shared into cache
             [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                 [[unroll]] for (uint j = 0; j < TM; j++) {
+                #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                    cache_a[wsir * TM + j].xy = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i    ];
+                    cache_a[wsir * TM + j].zw = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i + 1];
+                #else
                     cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i];
+                #endif
                 }
             }
 
             [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                 [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                    cache_b.xy = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i    ];
+                    cache_b.zw = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i + 1];
+                #else
                     cache_b = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + i];
+                #endif
 
                     [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                         [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
                             // [WNITER][TN][WMITER][TM / 2] -> [wsic][cc][wsir][cr]
                             const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
+                        #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                            sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y),
+                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].w), ACC_TYPE(cache_b.w), sums[sums_idx].x))));
+                            sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y),
+                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].w), ACC_TYPE(cache_b.w), sums[sums_idx].y))));
+                        #else
                             sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y), sums[sums_idx].x));
                             sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y), sums[sums_idx].y));
+                        #endif
                         }
                     }
                 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
index d955b4fc..dc8b3df4 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
@@ -78,8 +78,6 @@ layout (constant_id = 10) const uint WARP = 32;
 
 #define BK 32
 
-#define MMQ_SHMEM
-
 #include "mul_mmq_shmem_types.glsl"
 
 #ifdef MUL_MAT_ID
@@ -211,7 +209,9 @@ void main() {
             const uint iqs = loadr_a;
 
             [[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
-                block_a_to_shmem(k_step * BM + buf_ib, ib + k_step, iqs);
+                if (block + k_step * BK < end_k) {
+                    block_a_to_shmem(k_step * BM + buf_ib, ib + k_step, iqs);
+                }
             }
         }
         [[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) {
@@ -226,7 +226,7 @@ void main() {
             const uint iqs = loadr_b;
 
             [[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
-                block_b_to_shmem(k_step * BN + buf_ib, ib + k_step, iqs);
+                block_b_to_shmem(k_step * BN + buf_ib, ib + k_step, iqs, block + k_step * BK < end_k);
             }
         }
 
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
index c0c03fed..7f32dadf 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
@@ -9,31 +9,6 @@
 #if defined(DATA_A_Q4_0) || defined(DATA_A_Q4_1)
 // 2-byte loads for Q4_0 blocks (18 bytes)
 // 4-byte loads for Q4_1 blocks (20 bytes)
-i32vec2 repack(uint ib, uint iqs) {
-#ifdef DATA_A_Q4_0
-    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
-    const uint32_t vui = pack32(quants);
-    return i32vec2( vui       & 0x0F0F0F0F,
-                   (vui >> 4) & 0x0F0F0F0F);
-#else // DATA_A_Q4_1
-    const uint32_t vui = data_a_packed32[ib].qs[iqs];
-    return i32vec2( vui       & 0x0F0F0F0F,
-                   (vui >> 4) & 0x0F0F0F0F);
-#endif
-}
-
-#ifdef DATA_A_Q4_0
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
-}
-#else // DATA_A_Q4_1
-ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-}
-#endif
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 #ifdef DATA_A_Q4_0
     buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
@@ -73,42 +48,17 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a.y, qs_b1);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
+#ifdef DATA_A_Q4_0
+    return ACC_TYPE(float(cache_a[ib_a].dm) * (float(q_sum) * float(cache_b.ds.x) - 8.0 * float(cache_b.ds.y)));
+#else // DATA_A_Q4_1
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm.x) * float(cache_b.ds.x) + float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
+#endif
 }
-#endif // MMQ_SHMEM
+#endif
 
-#elif defined(DATA_A_Q5_0) || defined(DATA_A_Q5_1)
+#if defined(DATA_A_Q5_0) || defined(DATA_A_Q5_1)
 // 2-byte loads for Q5_0 blocks (22 bytes)
 // 4-byte loads for Q5_1 blocks (24 bytes)
-i32vec2 repack(uint ib, uint iqs) {
-    const u16vec2 quants = u16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-                                   data_a_packed16[ib].qs[iqs * 2 + 1]);
-    const uint32_t vui = pack32(quants);
-#ifdef DATA_A_Q5_0
-    const int32_t qh = int32_t((uint32_t(data_a_packed16[ib].qh[1]) << 16 | data_a_packed16[ib].qh[0]) >> (4 * iqs));
-#else // DATA_A_Q5_1
-    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
-#endif
-    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
-                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
-
-    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
-                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
-
-    return i32vec2(v0, v1);
-}
-
-#ifdef DATA_A_Q5_0
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
-}
-#else // DATA_A_Q5_1
-ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
-}
-#endif
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 #ifdef DATA_A_Q5_0
     buf_a[buf_ib].qs[iqs] = pack32(u16vec2(data_a_packed16[ib].qs[iqs * 2],
@@ -154,23 +104,16 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a1, qs_b1);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
+#ifdef DATA_A_Q5_0
+    return ACC_TYPE(float(cache_a[ib_a].dm) * (float(q_sum) * float(cache_b.ds.x) - 16.0 * float(cache_b.ds.y)));
+#else // DATA_A_Q5_1
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm.x) * float(cache_b.ds.x) + float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
+#endif
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_Q8_0)
 // 2-byte loads for Q8_0 blocks (34 bytes)
-int32_t repack(uint ib, uint iqs) {
-    return pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2    ],
-                          data_a_packed16[ib].qs[iqs * 2 + 1]));
-}
-
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(float(q_sum) * da * dsb.x);
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     buf_a[buf_ib].qs[iqs] = pack32(i16vec2(data_a_packed16[ib].qs[iqs * 2],
                                            data_a_packed16[ib].qs[iqs * 2 + 1]));
@@ -197,28 +140,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a, qs_b);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
+    return ACC_TYPE(float(q_sum) * float(cache_a[ib_a].dm) * float(cache_b.ds.x));
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_MXFP4)
 // 1-byte loads for mxfp4 blocks (17 bytes)
-i32vec2 repack(uint ib, uint iqs) {
-    const uint32_t quants = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
-                                          data_a[ib].qs[iqs * 4 + 1],
-                                          data_a[ib].qs[iqs * 4 + 2],
-                                          data_a[ib].qs[iqs * 4 + 3]));
-
-    return i32vec2( quants       & 0x0F0F0F0F,
-                   (quants >> 4) & 0x0F0F0F0F);
-}
-
-ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(da * dsb.x * float(q_sum));
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint32_t qs = pack32(u8vec4(data_a[ib].qs[iqs * 4    ],
                                       data_a[ib].qs[iqs * 4 + 1],
@@ -252,37 +179,14 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].d, cache_b.ds, 1);
+    return ACC_TYPE(float(cache_a[ib_a].d) * float(cache_b.ds.x) * float(q_sum));
 }
-#endif // MMQ_SHMEM
 #endif
 
 // For k-quants, ib and iqs still assume 32-wide blocks, but k-quants are 256-wide
 // iqs still refers to a 32-bit integer, meaning 0..7 for 32-wide quants
 #if defined(DATA_A_Q2_K)
 // 4-byte loads for Q2_K blocks (84 bytes)
-int32_t repack(uint ib, uint iqs) {
-    const uint ib_k = ib / 8;
-    const uint iqs_k = (ib % 8) * 8 + iqs;
-
-    const uint qs_idx = (iqs_k / 32) * 8 + (iqs_k % 8);
-    const uint qs_shift = ((iqs_k % 32) / 8) * 2;
-
-    return int32_t((data_a_packed32[ib_k].qs[qs_idx] >> qs_shift) & 0x03030303);
-}
-
-uint8_t get_scale(uint ib, uint iqs) {
-    const uint ib_k = ib / 8;
-    const uint iqs_k = (ib % 8) * 8 + iqs;
-
-    return data_a[ib_k].scales[iqs_k / 4];
-}
-
-ACC_TYPE mul_q8_1(const int32_t sum_d, const int32_t sum_m, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(dsb.x * (dma.x * float(sum_d) - dma.y * float(sum_m)));
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
@@ -300,7 +204,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 
     if (iqs == 0) {
         buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
-        buf_a[buf_ib].scales = unpack8(data_a_packed16[ib_k].scales[iqs_k / 8]);
+        buf_a[buf_ib].scales = unpack8(uint32_t(data_a_packed16[ib_k].scales[iqs_k / 8])).xy; // vec4 used due to #12147
     }
 }
 
@@ -326,14 +230,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         sum_m += dotPacked4x8EXT(scale_m, cache_b.qs[iqs]);
     }
 
-    return mul_q8_1(sum_d, sum_m, cache_a[ib_a].dm, cache_b.ds, 1);
+    return ACC_TYPE(float(cache_b.ds.x) * (float(cache_a[ib_a].dm.x) * float(sum_d) - float(cache_a[ib_a].dm.y) * float(sum_m)));
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_Q3_K)
 // 2-byte loads for Q3_K blocks (110 bytes)
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint hm_idx = iqs * QUANT_R_MMQ;
@@ -345,21 +247,22 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 
     // Repack 2x4 quants into one int
     // Add the 3rd bit instead of subtracting it to allow packing the quants
-    const i8vec2 vals00 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2        ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2    ] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals01 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 1    ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 1] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals10 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 2    ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 2] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals11 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 3    ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 3] >> hm_shift) & uint16_t(0x0101)) << 2));
+    // vec4 for unpack8 used due to #12147
+    const i8vec2 vals00 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2        ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2    ] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals01 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 1    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 1] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals10 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 2    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 2] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals11 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 3    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 3] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
     buf_a[buf_ib].qs[iqs] = pack32(u8vec4(vals00.x, vals00.y, vals01.x, vals01.y)) |
                            (pack32(u8vec4(vals10.x, vals10.y, vals11.x, vals11.y)) << 4);
 
     if (iqs == 0) {
         const uint is = iqs_k / 4;
-        const i8vec2 scales = i8vec2(unpack8(((data_a_packed16[ib_k].scales[(is % 8      ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
-                                            (((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4)));
+        const i8vec2 scales = i8vec2(unpack8(uint32_t(((data_a_packed16[ib_k].scales[(is % 8      ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
+                                                     (((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4))).xy); // vec4 used due to #12147
 
         buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales - 32);
     }
@@ -393,18 +296,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
     }
     result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
 
-    return ACC_TYPE(cache_b.ds.x * result);
+    return ACC_TYPE(float(cache_b.ds.x) * result);
 }
-#endif // MMQ_SHMEM
 #endif
 
 #if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
 // 4-byte loads for Q4_K blocks (144 bytes) and Q5_K blocks (176 bytes)
-ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
-    return ACC_TYPE(dsb.x * dma.x * float(q_sum) - dma.y * dsb.y);
-}
-
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint iqs_k = (ib % 8) * 8 + iqs * QUANT_R_MMQ;
@@ -426,7 +323,6 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
                                    (((data_a_packed32[ib_k].qh[qh_idx] >> qh_shift) & 0x01010101) << 4));
 #endif
 
-
     if (iqs == 0) {
         // Scale index
         const uint is = iqs_k / 8;
@@ -463,38 +359,12 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
         q_sum += dotPacked4x8EXT(qs_a, cache_b.qs[iqs]);
     }
 
-    return mul_q8_1(q_sum, cache_a[ib_a].dm, cache_b.ds, 1);
-}
-#endif // MMQ_SHMEM
-#endif
-
-#ifdef MMQ_SHMEM
-void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
-    const uint ib_outer = ib / 4;
-    const uint ib_inner = ib % 4;
-
-    if (iqs == 0) {
-        buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
-    }
-
-    const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
-    buf_b[buf_ib].qs[iqs * 4    ] = values.x;
-    buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
-    buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
-    buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
-}
-
-void block_b_to_registers(const uint ib) {
-    cache_b.ds = buf_b[ib].ds;
-    [[unroll]] for (uint iqs = 0; iqs < BK / 4; iqs++) {
-        cache_b.qs[iqs] = buf_b[ib].qs[iqs];
-    }
+    return ACC_TYPE(float(cache_b.ds.x) * float(cache_a[ib_a].dm.x) * float(q_sum) - float(cache_a[ib_a].dm.y) * float(cache_b.ds.y));
 }
 #endif
 
 #if defined(DATA_A_Q6_K)
 // 2-byte loads for Q6_K blocks (210 bytes)
-#ifdef MMQ_SHMEM
 void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint ib_k = ib / 8;
     const uint iqs_k = (ib % 8) * 8 + iqs;
@@ -505,15 +375,15 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint qh_idx = (iqs_k / 32) * 8 + iqs;
     const uint qh_shift = ((iqs_k % 32) / 8) * 2;
 
-    const i8vec2 vals00 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals01 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals00 = (unpack8(int32_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))).xy |
+                          unpack8(int32_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4)).xy) - int8_t(32);
+    const i8vec2 vals01 = (unpack8(int32_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))).xy |
+                          unpack8(int32_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4)).xy) - int8_t(32);
     buf_a[buf_ib].qs[iqs] = pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y));
 
     if (iqs == 0) {
         const uint is = iqs_k / 4;
-        const i8vec2 scales = unpack8(data_a_packed16[ib_k].scales[is / 2]);
+        const i8vec2 scales = unpack8(int32_t(data_a_packed16[ib_k].scales[is / 2])).xy;
 
         buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales);
     }
@@ -546,32 +416,39 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
     }
     result += float(cache_a[ib_a].d_scales[1]) * float(q_sum);
 
-    return ACC_TYPE(cache_b.ds.x * result);
-}
-#endif // MMQ_SHMEM
-#endif
-
-#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
-FLOAT_TYPE get_d(uint ib) {
-    return FLOAT_TYPE(data_a[ib].d);
+    return ACC_TYPE(float(cache_b.ds.x) * result);
 }
 #endif
 
-#if defined(DATA_A_MXFP4)
-FLOAT_TYPE get_d(uint ib) {
-    return FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e));
-}
-#endif
+void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bool is_in_bounds) {
+    if (is_in_bounds) {
+        const uint ib_outer = ib / 4;
+        const uint ib_inner = ib % 4;
 
-#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
-FLOAT_TYPE_VEC2 get_dm(uint ib) {
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
-}
-#endif
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
+        }
 
-#if defined(DATA_A_Q2_K)
-FLOAT_TYPE_VEC2 get_dm(uint ib) {
-    const uint ib_k = ib / 8;
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
+        const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
+        buf_b[buf_ib].qs[iqs * 4    ] = values.x;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
+    } else {
+        if (iqs == 0) {
+            buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
+        }
+
+        buf_b[buf_ib].qs[iqs * 4    ] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 1] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 2] = 0;
+        buf_b[buf_ib].qs[iqs * 4 + 3] = 0;
+    }
+}
+
+void block_b_to_registers(const uint ib) {
+    cache_b.ds = buf_b[ib].ds;
+    [[unroll]] for (uint iqs = 0; iqs < BK / 4; iqs++) {
+        cache_b.qs[iqs] = buf_b[ib].qs[iqs];
+    }
 }
-#endif
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
index 1e8f694a..10cf5202 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
@@ -23,16 +23,100 @@ layout (push_constant) uniform parameter2
     uint rms_partials;
 } p;
 
-// Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
-// layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
-// layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
-layout (binding = 0) buffer A {A_TYPE data_a[];} a[];
-layout (binding = 0) buffer D {D_TYPE data_d[];} d[];
-
-layout (binding = 0, std430) buffer PartialBuf {float partial_sums[];} partials[];
+// No readonly/writeonly decorations. Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
+layout (binding = 0)  buffer A0 {A_TYPE data_a[];} a0;
+layout (binding = 1)  buffer A1 {A_TYPE data_a[];} a1;
+layout (binding = 2)  buffer A2 {A_TYPE data_a[];} a2;
+layout (binding = 3)  buffer A3 {A_TYPE data_a[];} a3;
+layout (binding = 4)  buffer A4 {A_TYPE data_a[];} a4;
+layout (binding = 5)  buffer A5 {A_TYPE data_a[];} a5;
+layout (binding = 6)  buffer A6 {A_TYPE data_a[];} a6;
+layout (binding = 7)  buffer A7 {A_TYPE data_a[];} a7;
+layout (binding = 8)  buffer A8 {A_TYPE data_a[];} a8;
+layout (binding = 9)  buffer A9 {A_TYPE data_a[];} a9;
+layout (binding = 10) buffer A10 {A_TYPE data_a[];} a10;
+layout (binding = 11) buffer A11 {A_TYPE data_a[];} a11;
+layout (binding = 0)  buffer D0 {D_TYPE data_d[];} d0;
+layout (binding = 1)  buffer D1 {D_TYPE data_d[];} d1;
+layout (binding = 2)  buffer D2 {D_TYPE data_d[];} d2;
+layout (binding = 3)  buffer D3 {D_TYPE data_d[];} d3;
+layout (binding = 4)  buffer D4 {D_TYPE data_d[];} d4;
+layout (binding = 5)  buffer D5 {D_TYPE data_d[];} d5;
+layout (binding = 6)  buffer D6 {D_TYPE data_d[];} d6;
+layout (binding = 7)  buffer D7 {D_TYPE data_d[];} d7;
+layout (binding = 8)  buffer D8 {D_TYPE data_d[];} d8;
+layout (binding = 9)  buffer D9 {D_TYPE data_d[];} d9;
+layout (binding = 10) buffer D10 {D_TYPE data_d[];} d10;
+layout (binding = 11) buffer D11 {D_TYPE data_d[];} d11;
+layout (binding = 0, std430)  buffer PartialBuf0 {float partial_sums[];} partials0;
+layout (binding = 1, std430)  buffer PartialBuf1 {float partial_sums[];} partials1;
+layout (binding = 2, std430)  buffer PartialBuf2 {float partial_sums[];} partials2;
+layout (binding = 3, std430)  buffer PartialBuf3 {float partial_sums[];} partials3;
+layout (binding = 4, std430)  buffer PartialBuf4 {float partial_sums[];} partials4;
+layout (binding = 5, std430)  buffer PartialBuf5 {float partial_sums[];} partials5;
+layout (binding = 6, std430)  buffer PartialBuf6 {float partial_sums[];} partials6;
+layout (binding = 7, std430)  buffer PartialBuf7 {float partial_sums[];} partials7;
+layout (binding = 8, std430)  buffer PartialBuf8 {float partial_sums[];} partials8;
+layout (binding = 9, std430)  buffer PartialBuf9 {float partial_sums[];} partials9;
+layout (binding = 10, std430) buffer PartialBuf10 {float partial_sums[];} partials10;
+layout (binding = 11, std430) buffer PartialBuf11 {float partial_sums[];} partials11;
 
 layout(constant_id = 0) const uint num_srcs = 2;
 
+FLOAT_TYPE load_a(uint b, uint i) {
+    switch (b) {
+    case 0:  return FLOAT_TYPE(a0.data_a[i]);
+    case 1:  return FLOAT_TYPE(a1.data_a[i]);
+    case 2:  return FLOAT_TYPE(a2.data_a[i]);
+    case 3:  return FLOAT_TYPE(a3.data_a[i]);
+    case 4:  return FLOAT_TYPE(a4.data_a[i]);
+    case 5:  return FLOAT_TYPE(a5.data_a[i]);
+    case 6:  return FLOAT_TYPE(a6.data_a[i]);
+    case 7:  return FLOAT_TYPE(a7.data_a[i]);
+    case 8:  return FLOAT_TYPE(a8.data_a[i]);
+    case 9:  return FLOAT_TYPE(a9.data_a[i]);
+    case 10: return FLOAT_TYPE(a10.data_a[i]);
+    case 11: return FLOAT_TYPE(a11.data_a[i]);
+    default: return FLOAT_TYPE(0);
+    }
+}
+
+void store_d(uint b, uint i, FLOAT_TYPE v) {
+    switch (b) {
+    case 0:  d0.data_d[i] = D_TYPE(v); break;
+    case 1:  d1.data_d[i] = D_TYPE(v); break;
+    case 2:  d2.data_d[i] = D_TYPE(v); break;
+    case 3:  d3.data_d[i] = D_TYPE(v); break;
+    case 4:  d4.data_d[i] = D_TYPE(v); break;
+    case 5:  d5.data_d[i] = D_TYPE(v); break;
+    case 6:  d6.data_d[i] = D_TYPE(v); break;
+    case 7:  d7.data_d[i] = D_TYPE(v); break;
+    case 8:  d8.data_d[i] = D_TYPE(v); break;
+    case 9:  d9.data_d[i] = D_TYPE(v); break;
+    case 10: d10.data_d[i] = D_TYPE(v); break;
+    case 11: d11.data_d[i] = D_TYPE(v); break;
+    default: break;
+    }
+}
+
+void store_partial(uint b, uint i, float v) {
+    switch (b) {
+    case 0:  partials0.partial_sums[i] = v; break;
+    case 1:  partials1.partial_sums[i] = v; break;
+    case 2:  partials2.partial_sums[i] = v; break;
+    case 3:  partials3.partial_sums[i] = v; break;
+    case 4:  partials4.partial_sums[i] = v; break;
+    case 5:  partials5.partial_sums[i] = v; break;
+    case 6:  partials6.partial_sums[i] = v; break;
+    case 7:  partials7.partial_sums[i] = v; break;
+    case 8:  partials8.partial_sums[i] = v; break;
+    case 9:  partials9.partial_sums[i] = v; break;
+    case 10: partials10.partial_sums[i] = v; break;
+    case 11: partials11.partial_sums[i] = v; break;
+    default: break;
+    }
+}
+
 uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
     return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
 }
@@ -78,10 +162,10 @@ void main() {
 
         FLOAT_TYPE sum = FLOAT_TYPE(0);
         [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
-            sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
+            sum += load_a(s, src_idx(s, i00, i01, i02, i03));
         }
         sum_sq += sum*sum;
-        d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
+        store_d(num_srcs, dst_idx(i00, i01, i02, i03), sum);
 
         idx += num_threads;
     }
@@ -104,7 +188,7 @@ void main() {
         }
 
         if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
-            partials[num_srcs + 1].partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+            store_partial(num_srcs + 1, orig_idx / (num_iter * num_threads), sum_sq);
         }
     }
 #endif
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp
new file mode 100644
index 00000000..7f9b1bce
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp
@@ -0,0 +1,20 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+    data_d[i] = D_TYPE(-float(data_a[i]));
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
index 0f3c6ca8..20e45d02 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
@@ -61,7 +61,7 @@ void quantize() {
 
     const uint a_idx = ib * 8 + iqs;
 
-    vec4 vals = a_idx < p.ne ? data_a[a_idx] : vec4(0.0f);
+    vec4 vals = a_idx < p.ne / 4 ? data_a[a_idx] : vec4(0.0f);
     const vec4 abs_vals = abs(vals);
 
     // Find absolute max for each block
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
index d5b211ff..3a47949d 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
@@ -3,6 +3,32 @@
 #include "generic_binary_head.glsl"
 #include "types.glsl"
 
+#if RMS_NORM_ROPE_FUSION
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+
+// data is passed from rms_norm -> rope through shared memory.
+// rms_norm calls this data_d, rope calls this rope_data_a.
+// Binding 2 is not used
+shared FLOAT_TYPE rope_data_a[1024];
+#define data_d rope_data_a
+
+layout (binding = 3) readonly buffer R_Y {int rope_data_pos[];};
+layout (binding = 4) readonly buffer R_Z {float rope_data_ff[];};
+layout (binding = 5) writeonly buffer R_D {ROPE_D_TYPE rope_data_d[];};
+layout (binding = 6) readonly buffer R_I {uvec2 rope_data_i[];}; // indices for set_rows
+
+#include "rope_params.glsl"
+#include "rope_funcs.glsl"
+
+#define GGML_ROPE_TYPE_NORMAL 0
+#define GGML_ROPE_TYPE_NEOX   2
+#define GGML_ROPE_TYPE_MROPE  8
+#define GGML_ROPE_TYPE_VISION 24
+
+#endif
+
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512
 
@@ -28,8 +54,12 @@ void rms_norm(uint num_iters) {
 
     uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
     uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
+#if RMS_NORM_ROPE_FUSION
+    // Per-row offset in shared memory
+    uint32_t d_offset = 0;
+#else
     uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
-
+#endif
     FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
     [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
@@ -79,6 +109,18 @@ void rms_norm(uint num_iters) {
             data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
         }
     }
+#if RMS_NORM_ROPE_FUSION
+    barrier();
+    rope_params rp = p.rope;
+    uint rope_row = (samp*nchannels + channel)*nrows + row;
+    for (uint t = 2*tid; t < ncols; t += 2*BLOCK_SIZE) {
+        if (rp.rope_mode == GGML_ROPE_TYPE_NEOX) {
+            rope_neox(t, rope_row, rp);
+        } else if (rp.rope_mode == GGML_ROPE_TYPE_NORMAL) {
+            rope_norm(t, rope_row, rp);
+        }
+    }
+#endif
 }
 
 void main() {
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl
new file mode 100644
index 00000000..1c8c6942
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl
@@ -0,0 +1,227 @@
+
+float rope_yarn_ramp(const float low, const float high, const uint i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+uint rope_a_coord(const uint i0, const uint i01, const uint i02, rope_params p) {
+#if RMS_NORM_ROPE_FUSION
+    // Per-row offset in shared memory
+    const uint ix = i0;
+#else
+    const uint ix = i02*p.nb02 + i01*p.nb01 + i0;
+#endif
+    return ix;
+}
+
+void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta, rope_params p) {
+    float mscale = p.attn_factor;
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = p.freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (p.ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
+    }
+    // Backprogagation uses inverted rotation
+    if (p.is_back != 0) {
+        theta = -theta;
+    }
+    cos_theta = cos(theta) * mscale;
+    sin_theta = sin(theta) * mscale;
+}
+
+void rope_norm(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    uint idst = i1*ne0 + i0;
+    const uint ix = rope_a_coord(i0, i01, i02, p);
+
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
+    if (p.set_rows_stride != 0) {
+        idst = i01*ne0 + i0;
+        idst += rope_data_i[i02].x * p.set_rows_stride;
+    }
+
+    if (i0 >= p.n_dims) {
+        rope_data_d[idst + 0] = ROPE_D_TYPE(rope_data_a[ix + 0]);
+        rope_data_d[idst + 1] = ROPE_D_TYPE(rope_data_a[ix + 1]);
+
+        return;
+    }
+
+    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + 1]);
+
+    rope_data_d[idst + 0] = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + 1] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+
+void rope_neox(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    uint idst = i1*ne0 + i0/2;
+    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
+    if (p.set_rows_stride != 0) {
+        idst = i01*ne0 + i0/2;
+        idst += rope_data_i[i02].x * p.set_rows_stride;
+    }
+
+    if (i0 >= p.n_dims) {
+        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
+        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);
+
+        return;
+    }
+
+    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+
+    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+
+
+void rope_multi(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+    uint ne2 = p.ne02;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    const uint idst = i1*ne0 + i0/2;
+    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+
+    if (i0 >= p.n_dims) {
+        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
+        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);
+
+        return;
+    }
+
+    const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3];
+    const int sec_w = p.sections[1] + p.sections[0];
+    const uint sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (p.is_imrope != 0) {
+        if (sector % 3 == 1 && sector < 1 + 3 * p.sections[1]) {
+            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 2 + 3 * p.sections[2]) {
+            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
+            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+        //} else {
+        //    theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < p.sections[0]) {
+            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= p.sections[0] && sector < sec_w) {
+            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
+            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + p.sections[2]) {
+            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+        }
+    }
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+
+    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+
+void rope_vision(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+    uint ne2 = p.ne02;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    const uint idst = i1*ne0 + i0/2;
+    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+
+    const int sect_dims = p.sections[0] + p.sections[1];
+    const int sec_w = p.sections[1] + p.sections[0];
+    const uint sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (sector < p.sections[0]) {
+        const uint p0 = sector;
+        theta_base = rope_data_pos[i02]*pow(p.theta_scale, p0);
+    }
+    else if (sector >= p.sections[0] && sector < sec_w) {
+        const uint p0 = sector - p.sections[0];
+        theta_base = rope_data_pos[i02 + ne2]*pow(p.theta_scale, p0);
+    }
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_dims]);
+
+    rope_data_d[idst + 0]        = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_dims] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl
index 0eda186c..d9b4d4c0 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl
@@ -3,55 +3,18 @@
 #extension GL_EXT_shader_16bit_storage : require
 
 #include "rte.glsl"
+#include "rope_params.glsl"
 
 layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer Y {int data_pos[];};
-layout (binding = 2) readonly buffer Z {float data_ff[];};
-layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
-layout (binding = 4) readonly buffer I {uvec2 data_i[];}; // indices for set_rows
+layout (binding = 0) readonly buffer X {A_TYPE rope_data_a[];};
+layout (binding = 1) readonly buffer Y {int rope_data_pos[];};
+layout (binding = 2) readonly buffer Z {float rope_data_ff[];};
+layout (binding = 3) writeonly buffer D {ROPE_D_TYPE rope_data_d[];};
+layout (binding = 4) readonly buffer I {uvec2 rope_data_i[];}; // indices for set_rows
+
 
 layout (push_constant) uniform parameter {
-    uint ncols;
-    uint n_dims;
-    float freq_scale;
-    uint p_delta_rows;
-    float freq_base;
-    float ext_factor;
-    float attn_factor;
-    float corr_dims[2];
-    float theta_scale;
-    uint has_ff;
-    uint ne02;
-    uint s1;
-    uint s2;
-    int sections[4];
-    uint is_back;
-    uint set_rows_stride;
-} p;
+    rope_params pc;
+};
 
-float rope_yarn_ramp(const float low, const float high, const uint i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta) {
-    float mscale = p.attn_factor;
-    // Get n-d rotational scaling corrected for extrapolation
-    float theta_interp = p.freq_scale * theta_extrap;
-    float theta = theta_interp;
-    if (p.ext_factor != 0.0f) {
-        float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
-        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-
-        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
-    }
-    // Backprogagation uses inverted rotation
-    if (p.is_back != 0) {
-        theta = -theta;
-    }
-    cos_theta = cos(theta) * mscale;
-    sin_theta = sin(theta) * mscale;
-}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
index 633dc20f..7c1fb1cd 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
@@ -1,52 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
-        return;
-    }
-
-    const uint row_dst = gl_GlobalInvocationID.x;
-
-    const uint row_x     = row_dst % ne1;
-    const uint channel_x = row_dst / ne1;
-
-    const uint idst = row_dst*ne0 + i0/2;
-    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
-
-    if (i0 >= p.n_dims) {
-        data_d[idst + i0/2 + 0] = data_a[ix + i0/2 + 0];
-        data_d[idst + i0/2 + 1] = data_a[ix + i0/2 + 1];
-
-        return;
-    }
-
-    const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3];
-    const int sec_w = p.sections[1] + p.sections[0];
-    const uint sector = (i0 / 2) % sect_dims;
-
-    float theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
-    if (sector % 3 == 1 && sector < 1 + 3 * p.sections[1]) {
-        theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
-    }
-    else if (sector % 3 == 2 && sector < 2 + 3 * p.sections[2]) {
-        theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
-    }
-
-    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
-
-    const float x0 = float(data_a[ix + 0]);
-    const float x1 = float(data_a[ix + p.n_dims/2]);
-
-    data_d[idst + 0]          = D_TYPE(x0*cos_theta - x1*sin_theta);
-    data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_multi(i0, i1, pc);
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
index 9f453815..68f00c18 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
@@ -1,48 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
-        return;
-    }
-
-    const uint row_dst = gl_GlobalInvocationID.x;
-
-    const uint row_x     = row_dst % ne1;
-    const uint channel_x = row_dst / ne1;
-
-    uint idst = row_dst*ne0 + i0/2;
-    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
-
-    // Fusion optimization: ROPE + VIEW + SET_ROWS..
-    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
-    if (p.set_rows_stride != 0) {
-        idst = row_x*ne0 + i0/2;
-        idst += data_i[channel_x].x * p.set_rows_stride;
-    }
-
-    if (i0 >= p.n_dims) {
-        data_d[idst + i0/2 + 0] = D_TYPE(data_a[ix + i0/2 + 0]);
-        data_d[idst + i0/2 + 1] = D_TYPE(data_a[ix + i0/2 + 1]);
-
-        return;
-    }
-
-    const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f);
-
-    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
-
-    const float x0 = float(data_a[ix + 0]);
-    const float x1 = float(data_a[ix + p.n_dims/2]);
-
-    data_d[idst + 0]          = D_TYPE(x0*cos_theta - x1*sin_theta);
-    data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_neox(i0, i1, pc);
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
index f4209ed9..28a939ec 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
@@ -1,48 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
-        return;
-    }
-
-    const uint row_dst = gl_GlobalInvocationID.x;
-
-    const uint row_x     = row_dst % ne1;
-    const uint channel_x = row_dst / ne1;
-
-    uint idst = row_dst*ne0 + i0;
-    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0;
-
-    // Fusion optimization: ROPE + VIEW + SET_ROWS..
-    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
-    if (p.set_rows_stride != 0) {
-        idst = row_x*ne0 + i0;
-        idst += data_i[channel_x].x * p.set_rows_stride;
-    }
-
-    if (i0 >= p.n_dims) {
-        data_d[idst + 0] = D_TYPE(data_a[ix + 0]);
-        data_d[idst + 1] = D_TYPE(data_a[ix + 1]);
-
-        return;
-    }
-
-    const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f);
-
-    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
-
-    const float x0 = float(data_a[ix + 0]);
-    const float x1 = float(data_a[ix + 1]);
-
-    data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
-    data_d[idst + 1] = D_TYPE(x0*sin_theta + x1*cos_theta);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_norm(i0, i1, pc);
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl
new file mode 100644
index 00000000..82f39cee
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl
@@ -0,0 +1,27 @@
+#if !defined(GGML_ROPE_PARAMS)
+#define GGML_ROPE_PARAMS
+
+#include "rte.glsl"
+
+struct rope_params {
+    uint rope_mode;
+    uint ncols;
+    uint n_dims;
+    float freq_scale;
+    uint p_delta_rows;
+    float freq_base;
+    float ext_factor;
+    float attn_factor;
+    float corr_dims[2];
+    float theta_scale;
+    uint has_ff;
+    uint ne02;
+    uint nb01;
+    uint nb02;
+    int sections[4];
+    uint is_imrope;
+    uint is_back;
+    uint set_rows_stride;
+};
+
+#endif // !defined(GGML_ROPE_PARAMS)
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp
index d37d1c10..ea1e0fdb 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp
@@ -1,47 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
-        return;
-    }
-
-    const uint row_dst = gl_GlobalInvocationID.x;
-
-    const uint row_x     = row_dst % ne1;
-    const uint channel_x = row_dst / ne1;
-
-    const uint idst = row_dst*ne0 + i0/2;
-    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
-
-    const int sect_dims = p.sections[0] + p.sections[1];
-    const int sec_w = p.sections[1] + p.sections[0];
-    const uint sector = (i0 / 2) % sect_dims;
-
-    float theta_base = 0.0;
-    if (sector < p.sections[0]) {
-        const uint p0 = sector;
-        theta_base = data_pos[channel_x]*pow(p.theta_scale, p0);
-    }
-    else if (sector >= p.sections[0] && sector < sec_w) {
-        const uint p0 = sector - p.sections[0];
-        theta_base = data_pos[channel_x + ne2]*pow(p.theta_scale, p0);
-    }
-
-    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
-
-    const float x0 = float(data_a[ix + 0]);
-    const float x1 = float(data_a[ix + p.n_dims]);
-
-    data_d[idst + 0]        = D_TYPE(x0*cos_theta - x1*sin_theta);
-    data_d[idst + p.n_dims] = D_TYPE(x0*sin_theta + x1*cos_theta);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_vision(i0, i1, pc);
 }
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/round.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/round.comp
new file mode 100644
index 00000000..e6155dcb
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/round.comp
@@ -0,0 +1,29 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    float result;
+    // Round halfway cases away from zero as roundf does.
+    if (x >= 0.0) {
+        result = floor(x + 0.5);
+    } else {
+        result = ceil(x - 0.5);
+    }
+    data_d[i] = D_TYPE(result);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp
new file mode 100644
index 00000000..323e3cde
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp
@@ -0,0 +1,23 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    const float result = (x > 20.0f) ? x : log(1.0f + exp(x));
+    data_d[i] = D_TYPE(result);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp
new file mode 100644
index 00000000..253a9e7e
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp
@@ -0,0 +1,72 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+
+layout (constant_id = 1) const uint N = 64;
+layout (constant_id = 2) const uint K = 32;
+
+layout(local_size_x = 128, local_size_y = 1, local_size_z = 1) in;
+
+uint a_base, b_base, x_base;
+
+FLOAT_TYPE get_a(uint r, uint c) {
+    return FLOAT_TYPE(data_a[a_base + r * p.nb01 + c * p.nb00]);
+}
+
+FLOAT_TYPE get_b(uint r, uint c) {
+    return FLOAT_TYPE(data_b[b_base + r * p.nb11 + c * p.nb10]);
+}
+
+void store_x(uint r, uint c, FLOAT_TYPE v) {
+    data_d[x_base + r * p.nb21 + c * p.nb20] = D_TYPE(v);
+}
+
+shared FLOAT_TYPE shA[N * N];
+shared FLOAT_TYPE shB[N * K];
+
+void main() {
+    const uint batch = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    if (batch >= p.ne02 * p.ne03) {
+        return;
+    }
+
+    const uint i3 = batch / p.ne22;
+    const uint i2 = batch % p.ne22;
+    a_base = get_aoffset() + i2 * p.nb02 + i3 * p.nb03;
+    b_base = get_boffset() + i2 * p.nb12 + i3 * p.nb13;
+    x_base = get_doffset() + i2 * p.nb22 + i3 * p.nb23;
+
+    // Load the A matrix into shA
+    [[unroll]] for (uint i = 0; i < N * N; i += gl_WorkGroupSize.x) {
+        uint idx = i + tid;
+        if (((N * N) % gl_WorkGroupSize.x == 0) || idx < N * N) {
+            shA[idx] = get_a(idx / N, idx % N);
+        }
+    }
+    // Load the B matrix into shB
+    [[unroll]] for (uint i = 0; i < N * K; i += gl_WorkGroupSize.x) {
+        uint idx = i + tid;
+        if (((N * K) % gl_WorkGroupSize.x == 0) || idx < N * K) {
+            shB[idx] = get_b(idx / K, idx % K);
+        }
+    }
+    barrier();
+
+    FLOAT_TYPE X[N];
+    // Each thread solves one column
+    if (tid < K) {
+        [[unroll]] for (int r = 0; r < N; ++r) {
+            FLOAT_TYPE b = shB[r * K + tid];
+            // Compute x[r,c] = (b[r,c] - sum(a[r,c]*x[c])) / a[r,r]
+            [[unroll]] for (int c = 0; c < r; ++c) {
+                b -= shA[r * N + c] * X[c];
+            }
+            FLOAT_TYPE x = b / shA[r * N + r];
+            X[r] = x;
+            store_x(r, tid, x);
+        }
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/step.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/step.comp
new file mode 100644
index 00000000..654a2124
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/step.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x >= 0.0f ? 1.0f : 0.0f);
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
index bc22aa7b..13ba2e99 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
@@ -1,6 +1,7 @@
 #version 450
 
 #include "types.glsl"
+#include "sum_rows.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 
@@ -11,30 +12,6 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 
-layout (push_constant) uniform parameter
-{
-    uint n_cols;
-    uint ne01, ne02;
-    uint nb01, nb02, nb03;
-    uint nb11, nb12, nb13;
-    float weight;
-    uint misalign_offsets;
-    uint ne0_12mp, ne0_12L;
-    uint ne0_1mp, ne0_1L;
-} p;
-
-uint get_aoffset() { return p.misalign_offsets >> 16; }
-uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
-
-// see init_fastdiv_values in ggml-vulkan.cpp
-uint fastdiv(uint n, uint mp, uint L) {
-    uint msbs, lsbs;
-    // msbs = mulhi(n, mp)
-    umulExtended(n, mp, msbs, lsbs);
-    return (msbs + n) >> L;
-}
-
-
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl
new file mode 100644
index 00000000..2b841baa
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl
@@ -0,0 +1,25 @@
+
+// vk_op_sum_rows_push_constants
+layout (push_constant) uniform parameter
+{
+    uint n_cols;
+    uint ne01, ne02;
+    uint nb01, nb02, nb03;
+    uint nb11, nb12, nb13;
+    float weight;
+    uint misalign_offsets;
+    uint ne0_12mp, ne0_12L;
+    uint ne0_1mp, ne0_1L;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp
new file mode 100644
index 00000000..cd858b7d
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp
@@ -0,0 +1,113 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int NCOLS_PADDED_LOG2 = 10;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// Input can either be the source (A) or intermediate values (S).
+// Similarly, output can be either destination (D) or intermediate values (S).
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 0) readonly buffer S {ivec2 data_s[];};
+layout (binding = 1) writeonly buffer D {int data_d[];};
+layout (binding = 1) writeonly buffer T {ivec2 data_t[];};
+
+layout (push_constant) uniform parameter {
+    uint orig_ncols;
+    uint ncols_input;
+    uint ncols_output;
+    uint nrows;
+    uint first_pass;
+    uint last_pass;
+} p;
+
+// pairs of (gid, value)
+shared ivec2 dst_row[BLOCK_SIZE];
+
+void topk(bool needs_bounds_check, const uint row) {
+    const int col = int(gl_LocalInvocationID.x);
+
+    // initialize indices
+    if (gl_GlobalInvocationID.x < p.ncols_input) {
+        if (p.first_pass != 0) {
+            const uint row_offset = row * p.ncols_input;
+            dst_row[col] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
+        } else {
+            const uint row_offset = row * p.orig_ncols;
+            dst_row[col] = data_s[row_offset + gl_GlobalInvocationID.x];
+        }
+    } else {
+        dst_row[col] = ivec2(p.orig_ncols, 0);
+    }
+    barrier();
+
+    if (p.ncols_output == 1) {
+        // Fast path for single output - just do a max reduction
+        [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
+            if (col < s) {
+                ivec2 a = dst_row[col];
+                ivec2 b = dst_row[col + s];
+                if (a.x >= p.orig_ncols ||
+                    b.x < p.orig_ncols && b.y > a.y) {
+                    dst_row[col] = b;
+                }
+            }
+            barrier();
+        }
+    } else {
+        // bitonic sort on this group of elements
+        uint num_outer_loop_iters = NCOLS_PADDED_LOG2;
+        for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
+            uint num_inner_loop_iters = outer_idx + 1;
+            for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
+                const int ixj = int(col ^ j);
+
+                int idx_0 = (col & k) == 0 ? col : ixj;
+                int idx_1 = (col & k) == 0 ? ixj : col;
+
+                ivec2 sh_idx_0 = dst_row[idx_0];
+                ivec2 sh_idx_1 = dst_row[idx_1];
+                bool idx_0_oob = needs_bounds_check ? sh_idx_0.x >= p.orig_ncols : false;
+                bool idx_1_oob = needs_bounds_check ? sh_idx_1.x >= p.orig_ncols : false;
+
+                if ((idx_0_oob ||
+                    (!idx_1_oob && intBitsToFloat(sh_idx_0.y) < intBitsToFloat(sh_idx_1.y))) && (ixj > col)) {
+                    dst_row[idx_0] = sh_idx_1;
+                    dst_row[idx_1] = sh_idx_0;
+                }
+
+                barrier();
+            }
+        }
+    }
+
+    if (col < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+        if (p.last_pass != 0) {
+            const uint row_offset = row * p.ncols_output;
+            data_d[row_offset + col] = dst_row[col].x;
+        } else {
+            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
+            data_t[row_offset + col] = dst_row[col];
+        }
+    }
+}
+
+void main() {
+    // Fast path for fully occupied workgroups
+    if ((p.ncols_input % BLOCK_SIZE) == 0) {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            topk(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    } else {
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            topk(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp
new file mode 100644
index 00000000..c902e602
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp
@@ -0,0 +1,199 @@
+#version 450
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_debug_printf : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_shuffle : enable
+
+#include "types.glsl"
+
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int SUBGROUP_SIZE = 32;
+layout(constant_id = 2) const int SUBGROUP_SIZE_LOG2 = 5;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// Input can either be the source (A) or intermediate values (S).
+// Similarly, output can be either destination (D) or intermediate values (S).
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 0) readonly buffer S {ivec2 data_s[];};
+layout (binding = 1) writeonly buffer D {int data_d[];};
+layout (binding = 1) writeonly buffer T {ivec2 data_t[];};
+
+layout (push_constant) uniform parameter {
+    uint orig_ncols;
+    uint ncols_input;
+    uint ncols_output;
+    uint nrows;
+    uint first_pass;
+    uint last_pass;
+} p;
+
+// pairs of (gid, value)
+shared ivec2 dst_row[BLOCK_SIZE];
+
+shared int counts[SUBGROUP_SIZE];
+shared int sh_min_idx;
+shared uint sh_total;
+shared uint offset_partials[BLOCK_SIZE / SUBGROUP_SIZE];
+
+// Map float values to uint such that comparisons still work.
+// Positive values set the high bit, negative values are inverted.
+// +0.0 -> 0x80000000, -0.0 -> 0x7FFFFFFF are in the correct places.
+uint f2ui(float x) {
+    uint y = floatBitsToUint(x);
+    if ((y & 0x80000000) != 0) {
+        y ^= ~0;
+    } else {
+        y |= 0x80000000;
+    }
+    return y;
+}
+
+void topk(const uint row) {
+    const int tid = int(gl_LocalInvocationID.x);
+
+    // initialize indices
+    if (gl_GlobalInvocationID.x < p.ncols_input) {
+        if (p.first_pass != 0) {
+            const uint row_offset = row * p.ncols_input;
+            dst_row[tid] = ivec2(gl_GlobalInvocationID.x, floatBitsToInt(data_a[row_offset + gl_GlobalInvocationID.x]));
+        } else {
+            const uint row_offset = row * p.orig_ncols;
+            dst_row[tid] = data_s[row_offset + gl_GlobalInvocationID.x];
+        }
+    } else {
+        dst_row[tid] = ivec2(p.orig_ncols, 0xFF800000); // -inf
+    }
+    barrier();
+
+    if (p.ncols_output == 1) {
+        // Fast path for single output - just do a max reduction
+        [[unroll]] for (int s = BLOCK_SIZE / 2; s >= 1; s /= 2) {
+            if (tid < s) {
+                ivec2 a = dst_row[tid];
+                ivec2 b = dst_row[tid + s];
+                if (a.x >= p.orig_ncols ||
+                    b.x < p.orig_ncols && b.y > a.y) {
+                    dst_row[tid] = b;
+                }
+            }
+            barrier();
+        }
+    } else {
+        // Do an N-ary search to find the K-th largest value.
+        // We remap the float values to be comparable as unsigned integers,
+        // and split the range into 2^N smaller ranges where N is the
+        // subgroup size. Count how many values are in each range, if the K-th
+        // largest value is in the middle of one of thee ranges then repeat
+        // and split again.
+
+        // Mask is the current set of bits we're searching. Shift is the LSB index.
+        int shift = 32 - SUBGROUP_SIZE_LOG2;
+        uint mask = ((1 << SUBGROUP_SIZE_LOG2) - 1) << shift;
+
+        // The current range.
+        uint range_min = 0;
+        uint range_max = 0xFF800000;
+        // How many are above the current range, and how many we need to find.
+        uint total = 0;
+        uint limit = min(p.ncols_output, p.ncols_input - gl_WorkGroupID.x * BLOCK_SIZE);
+
+        while (mask != 0) {
+            barrier();
+            // Initialize bucket counts to zero.
+            if (tid < SUBGROUP_SIZE) {
+                counts[tid] = 0;
+            }
+            barrier();
+            // Count how many values are in each bucket.
+            if (tid < p.ncols_input) {
+                float y = intBitsToFloat(dst_row[tid].y);
+                uint fy = f2ui(y);
+                if (fy >= range_min && fy < range_max) {
+                    uint bucket = (fy & mask) >> shift;
+                    atomicAdd(counts[bucket], 1);
+                }
+            }
+            barrier();
+
+            // On the first subgroup, do a scan to count (from the top down) how
+            // many elements are in the top N buckets. Find the index of the first
+            // that is over the limit. Copy it to the other invocations through
+            // shared memory.
+            if (tid < SUBGROUP_SIZE) {
+                uint partial_sum = counts[SUBGROUP_SIZE - 1 - tid];
+                partial_sum = subgroupInclusiveAdd(partial_sum) + total;
+                uint t = subgroupBallotFindLSB(subgroupBallot(partial_sum >= limit));
+                if (tid == t) {
+                    sh_min_idx = int(SUBGROUP_SIZE - 1 - t);
+                    sh_total = partial_sum;
+                }
+            }
+            barrier();
+            int min_idx = sh_min_idx;
+            total = sh_total;
+
+            // Update the range, and break if we've found the K-th largest.
+            range_max = range_min + ((min_idx + 1) << shift);
+            range_min = range_min + (min_idx << shift);
+
+            if (total == p.ncols_output) {
+                break;
+            }
+            total -= counts[min_idx];
+            mask >>= SUBGROUP_SIZE_LOG2;
+            shift -= SUBGROUP_SIZE_LOG2;
+            if (shift < 0) {
+                shift = 0;
+            }
+        }
+
+        ivec2 v = dst_row[tid];
+
+        // We need to compact these values to the start of the dst_row array.
+        // Have each subgroup count how many items it'll store, so other
+        // subgroups can compute their base offset.
+        bool top = f2ui(intBitsToFloat(v.y)) >= range_min;
+        uvec4 b = subgroupBallot(top);
+        uint bit_count = subgroupBallotBitCount(b);
+        if ((tid % SUBGROUP_SIZE) == 0) {
+            offset_partials[tid / SUBGROUP_SIZE] = bit_count;
+        }
+        barrier();
+
+        uint out_idx = 0;
+        [[unroll]] for (int i = 0; i < BLOCK_SIZE / SUBGROUP_SIZE; ++i) {
+            if (i < tid / SUBGROUP_SIZE) {
+                out_idx += offset_partials[i];
+            }
+        }
+
+        uint bit_count_ex = subgroupBallotExclusiveBitCount(b);
+        if (top) {
+            // TODO: Copy directly to the output?
+            dst_row[out_idx + bit_count_ex] = v;
+        }
+
+        barrier();
+    }
+
+    if (tid < p.ncols_output && gl_GlobalInvocationID.x < p.orig_ncols) {
+        if (p.last_pass != 0) {
+            const uint row_offset = row * p.ncols_output;
+            data_d[row_offset + tid] = dst_row[tid].x;
+        } else {
+            const uint row_offset = row * p.orig_ncols + gl_WorkGroupID.x * p.ncols_output;
+            data_t[row_offset + tid] = dst_row[tid];
+        }
+    }
+}
+
+void main() {
+    uint row = gl_WorkGroupID.y;
+    while (row < p.nrows) {
+        topk(row);
+        row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp
new file mode 100644
index 00000000..e18d0ffa
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp
@@ -0,0 +1,43 @@
+#version 450
+
+#include "rte.glsl"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+#define GGML_TRI_TYPE_UPPER_DIAG 0
+#define GGML_TRI_TYPE_UPPER      1
+#define GGML_TRI_TYPE_LOWER_DIAG 2
+#define GGML_TRI_TYPE_LOWER      3
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L);
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+
+    int param = floatBitsToInt(p.param1);
+    bool pass = false;
+    switch (param) {
+    case GGML_TRI_TYPE_UPPER_DIAG: pass = i00 >= i01; break;
+    case GGML_TRI_TYPE_UPPER:      pass = i00 >  i01; break;
+    case GGML_TRI_TYPE_LOWER_DIAG: pass = i00 <= i01; break;
+    case GGML_TRI_TYPE_LOWER:      pass = i00 <  i01; break;
+    }
+
+    if (pass) {
+        const float val = float(data_a[get_aoffset() + src0_idx(idx)]);
+        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val);
+    } else {
+        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(0);
+    }
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp
new file mode 100644
index 00000000..cf1b76d3
--- /dev/null
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(trunc(x));
+}
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp
index 154a2172..037ab0c7 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp
@@ -7,6 +7,7 @@ layout (push_constant) uniform parameter
     uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13;
     float sf0; float sf1; float sf2; float sf3;
+    float pixel_offset;
 } p;
 
 #include "types.glsl"
@@ -19,7 +20,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 // from ggml.h: enum ggml_scale_mode, enum ggml_scale_flag
 #define NEAREST  0
 #define BILINEAR 1
-#define ALIGN_CORNERS (1 << 8)
+#define BICUBIC  2
 
 layout (constant_id = 0) const uint scale_mode = 0;
 
@@ -52,7 +53,7 @@ float fetch_bilinear(ivec2 c0, ivec2 c1, vec2 d, uint i12, uint i13) {
 float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
     const ivec2 ne0 = ivec2(p.ne00, p.ne01);
 
-    const vec2 c = (vec2(i10, i11) + 0.5) / vec2(p.sf0, p.sf1) - 0.5;
+    const vec2 c = (vec2(i10, i11) + p.pixel_offset) / vec2(p.sf0, p.sf1) - p.pixel_offset;
     const vec2 c0f = floor(c);
     const vec2 d = c - c0f;
     const ivec2 c0 = max(ivec2(c0f), 0);
@@ -61,14 +62,37 @@ float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
     return fetch_bilinear(c0, c1, d, i12, i13);
 }
 
-float interpolate_bilinear_align_corners(uint i10, uint i11, uint i12, uint i13) {
-    const vec2 c = vec2(i10, i11) / vec2(p.sf0, p.sf1);
-    const vec2 c0f = floor(c);
-    const vec2 d = c - c0f;
-    const ivec2 c0 = ivec2(c0f);
-    const ivec2 c1 = c0 + 1;
+// Bicubic interpolation with alpha = -0.75
+// https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+const vec4 bcoeffs1 = vec4( 1.25, -2.25,  0.0, 1.0);
+const vec4 bcoeffs2 = vec4(-0.75,  3.75, -6.0, 3.0);
+vec4 powers(float x) { return vec4(x*x*x, x*x, x, 1); }
 
-    return fetch_bilinear(c0, c1, d, i12, i13);
+float bicubic(float p0, float p1, float p2, float p3, float x) {
+    return p0 * dot(bcoeffs2, powers(x + 1)) +
+           p1 * dot(bcoeffs1, powers(x    )) +
+           p2 * dot(bcoeffs1, powers(1 - x)) +
+           p3 * dot(bcoeffs2, powers(2 - x));
+}
+
+#define FETCH(a,b) data_a[base + clamp(i.x+(a), 0, res.x) * p.nb00 + clamp(i.y+(b), 0, res.y) * p.nb01]
+
+float interpolate_bicubic(uint i10, uint i11, uint i12, uint i13) {
+    const ivec2 res = ivec2(p.ne00 - 1, p.ne01 - 1);
+
+    const vec2 coord = (vec2(i10, i11) + p.pixel_offset) / vec2(p.sf0, p.sf1) - p.pixel_offset;
+    const vec2 d = fract(coord);
+    const ivec2 i = ivec2(floor(coord));
+
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+    const uint base = p.a_offset + i03 * p.nb03 + i02 * p.nb02;
+
+    return bicubic(
+        bicubic(FETCH(-1,-1), FETCH(0,-1), FETCH(1,-1), FETCH(2,-1), d.x),
+        bicubic(FETCH(-1, 0), FETCH(0, 0), FETCH(1, 0), FETCH(2, 0), d.x),
+        bicubic(FETCH(-1, 1), FETCH(0, 1), FETCH(1, 1), FETCH(2, 1), d.x),
+        bicubic(FETCH(-1, 2), FETCH(0, 2), FETCH(1, 2), FETCH(2, 2), d.x), d.y);
 }
 
 void main() {
@@ -91,8 +115,8 @@ void main() {
         case BILINEAR:
             result = interpolate_bilinear(i10, i11, i12, i13);
             break;
-        case BILINEAR | ALIGN_CORNERS:
-            result = interpolate_bilinear_align_corners(i10, i11, i12, i13);
+        case BICUBIC:
+            result = interpolate_bicubic(i10, i11, i12, i13);
             break;
     }
 
diff --git a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index e6ec589f..92bae088 100644
--- a/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -18,6 +18,7 @@
 #include <algorithm>
 #include <sys/stat.h>
 #include <sys/types.h>
+#include <filesystem>
 
 #ifdef _WIN32
     #define NOMINMAX
@@ -75,7 +76,7 @@ enum MatMulIdType {
 
 namespace {
 
-void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
+void execute_command(std::vector<std::string>& command, std::string& stdout_str, std::string& stderr_str) {
 #ifdef _WIN32
     HANDLE stdout_read, stdout_write;
     HANDLE stderr_read, stderr_write;
@@ -98,8 +99,10 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
     si.hStdOutput = stdout_write;
     si.hStdError = stderr_write;
 
-    std::vector<char> cmd(command.begin(), command.end());
-    cmd.push_back('\0');
+    std::string cmd;
+    for (const auto& part : command) {
+        cmd += part + " ";
+    }
 
     if (!CreateProcessA(NULL, cmd.data(), NULL, NULL, TRUE, 0, NULL, NULL, &si, &pi)) {
         throw std::runtime_error("Failed to create process");
@@ -137,6 +140,12 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
         throw std::runtime_error("Failed to fork process");
     }
 
+    std::vector<char*> argv;
+    for (std::string& part : command) {
+        argv.push_back(part.data());
+    }
+    argv.push_back(nullptr);
+
     if (pid == 0) {
         close(stdout_pipe[0]);
         close(stderr_pipe[0]);
@@ -144,7 +153,7 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
         dup2(stderr_pipe[1], STDERR_FILENO);
         close(stdout_pipe[1]);
         close(stderr_pipe[1]);
-        execl("/bin/sh", "sh", "-c", command.c_str(), (char*) nullptr);
+        execvp(argv[0], argv.data());
         _exit(EXIT_FAILURE);
     } else {
         close(stdout_pipe[1]);
@@ -315,20 +324,27 @@ compile_count_guard acquire_compile_slot() {
 void string_to_spv_func(std::string name, std::string in_path, std::string out_path, std::map<std::string, std::string> defines, bool coopmat, bool dep_file, compile_count_guard slot) {
     std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";
 
+    #ifdef _WIN32
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, "\"" + in_path + "\"", "-o", "\"" + out_path + "\""};
+    #else
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, in_path, "-o", out_path};
+    #endif
+
     // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
     // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
-    std::string opt_level = (coopmat || name.find("bf16") != std::string::npos) ? "" : "-O";
-
-    #ifdef _WIN32
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_path + "\""};
-    #else
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, in_path, "-o",  out_path};
-    #endif
+    // disable spirv-opt for rope shaders for https://github.com/ggml-org/llama.cpp/issues/16860
+    if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos) {
+        cmd.push_back("-O");
+    }
 
     if (dep_file) {
         cmd.push_back("-MD");
         cmd.push_back("-MF");
+#ifdef _WIN32
         cmd.push_back("\"" + target_cpp + ".d\"");
+#else
+        cmd.push_back(target_cpp + ".d");
+#endif
     }
 
     #ifdef GGML_VULKAN_SHADER_DEBUG_INFO
@@ -352,9 +368,13 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
         // }
         // std::cout << std::endl;
 
-        execute_command(command, stdout_str, stderr_str);
+        execute_command(cmd, stdout_str, stderr_str);
         if (!stderr_str.empty()) {
-            std::cerr << "cannot compile " << name << "\n\n" << command << "\n\n" << stderr_str << std::endl;
+            std::cerr << "cannot compile " << name << "\n\n";
+            for (const auto& part : cmd) {
+                std::cerr << part << " ";
+            }
+            std::cerr << "\n\n" << stderr_str << std::endl;
             return;
         }
 
@@ -428,7 +448,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
     base_dict["ACC_TYPE"     ] = f16acc ? "float16_t" : "float";
     base_dict["ACC_TYPE_VEC2"] = f16acc ? "f16vec2"   : "vec2";
     if (f16acc) {
-        base_dict["ACC_TYPE_MAX"] = "\"float16_t(65504.0)\"";
+        base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
     }
 
     if (coopmat) {
@@ -608,7 +628,7 @@ void process_shaders() {
         fa_base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
         fa_base_dict["ACC_TYPEV4"] = f16acc ? "f16vec4" : "vec4";
         if (f16acc) {
-            fa_base_dict["ACC_TYPE_MAX"] = "\"float16_t(65504.0)\"";
+            fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
         }
 
         for (const auto& tname : type_names) {
@@ -659,14 +679,20 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
         string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
 
-        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
 
         // mul mat vec with integer dot product
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        if (is_legacy_quant(tname)) {
+        if (is_legacy_quant(tname) || tname == "mxfp4" || is_k_quant(tname)) {
             string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
             string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
             string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
+
+            string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
+            string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+            string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
         }
 #endif
 
@@ -694,6 +720,8 @@ void process_shaders() {
     string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("rms_norm_mul_rope_f32_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float"}, {"RMS_NORM_ROPE_FUSION", "1"}}));
+    string_to_spv("rms_norm_mul_rope_f32_f16_rte", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RMS_NORM_ROPE_FUSION", "1"}, {"RTE16", "1"}}));
     string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
@@ -712,6 +740,9 @@ void process_shaders() {
     string_to_spv("cpy_f32_i32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
     string_to_spv("cpy_i32_f32", "copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});
 
+    string_to_spv("cpy_transpose_16", "copy_transpose.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
+    string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});
+
     for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
         string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
         string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
@@ -794,6 +825,9 @@ void process_shaders() {
         std::string suffix = rte ? "_rte" : "";
         string_to_spv("exp_f16" + suffix,        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
         string_to_spv("exp_f32" + suffix,        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}    ,   {"RTE16", rte ? "1" : "0"}});
+
+        string_to_spv("log_f16" + suffix,        "log.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("log_f32" + suffix,        "log.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
     }
     string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
@@ -805,6 +839,8 @@ void process_shaders() {
     string_to_spv("silu_f32",       "silu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("relu_f16",       "relu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("relu_f32",       "relu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("neg_f16",        "neg.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("neg_f32",        "neg.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("tanh_f16",       "tanh.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("tanh_f32",       "tanh.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("sigmoid_f16",    "sigmoid.comp",     {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
@@ -813,6 +849,30 @@ void process_shaders() {
     string_to_spv("hardsigmoid_f32","hardsigmoid.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("hardswish_f16",  "hardswish.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("hardswish_f32",  "hardswish.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("abs_f16",        "abs.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("abs_f32",        "abs.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+
+    string_to_spv("tri_f16",        "tri.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("tri_f32",        "tri.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+
+    string_to_spv("softplus_f16",   "softplus.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("softplus_f32",   "softplus.comp",    {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+
+    string_to_spv("add1_f16_f16",   "add1.comp",        {{"A_TYPE", "float16_t"},   {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
+    string_to_spv("add1_f16_f32",   "add1.comp",        {{"A_TYPE", "float16_t"},   {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
+    string_to_spv("add1_f32_f32",   "add1.comp",        {{"A_TYPE", "float"},       {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    string_to_spv("arange_f32",     "arange.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    string_to_spv("fill_f32",       "fill.comp",        {{"D_TYPE", "float"},       {"FLOAT_TYPE", "float"}});
+    string_to_spv("step_f16",       "step.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("step_f32",       "step.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("round_f16",      "round.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("round_f32",      "round.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("ceil_f16",       "ceil.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("ceil_f32",       "ceil.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("floor_f16",      "floor.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("floor_f32",      "floor.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("trunc_f16",      "trunc.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("trunc_f32",      "trunc.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
     for (auto rte : {false, true}) {
         std::string suffix = rte ? "_rte" : "";
@@ -839,31 +899,36 @@ void process_shaders() {
     string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
     string_to_spv("soft_max_back_f32", "soft_max_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
-    string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
-    string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
-    string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
-    string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
-    string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
-    string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
+    string_to_spv("argsort_large_f32", "argsort_large.comp", {{"A_TYPE", "float"}});
+
+    string_to_spv("topk_argsort_f32", "topk_argsort.comp", {{"A_TYPE", "float"}});
+    string_to_spv("topk_nary_search_f32", "topk_nary_search.comp", {{"A_TYPE", "float"}});
 
     string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
     string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
+    string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     for (std::string dim_str : {"", "_3d"}) {
         for (bool bda : {false, true}) {
@@ -888,6 +953,8 @@ void process_shaders() {
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
     string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
+    string_to_spv("solve_tri_f32", "solve_tri.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+
     for (auto transpose : {false, true}) {
         for (auto unroll : {false, true}) {
             for (auto a_f16 : {false, true}) {
@@ -1039,7 +1106,7 @@ void write_output_files() {
 
     for (const std::string& btype : btypes) {
     for (const auto& tname : type_names) {
-        if (btype == "q8_1" && !is_legacy_quant(tname)) {
+        if (btype == "q8_1" && !is_legacy_quant(tname) && tname != "mxfp4" && !is_k_quant(tname)) {
             continue;
         }
         hdr << "extern const void * arr_dmmv_"   << tname << "_" << btype << "_f32_data[3];\n";
@@ -1048,6 +1115,16 @@ void write_output_files() {
             src << "const void * arr_dmmv_"   << tname << "_" << btype << "_f32_data[3] = {mul_mat_vec_" << tname << "_" << btype << "_f32_data, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_data, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_no_shmem_data};\n";
             src << "const uint64_t arr_dmmv_" << tname << "_" << btype << "_f32_len[3] =  {mul_mat_vec_" << tname << "_" << btype << "_f32_len,  mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_len, mul_mat_vec_"  << tname << "_" << btype << "_f32_subgroup_no_shmem_len};\n";
         }
+
+        if (btype == "f16") {
+            continue;
+        }
+        hdr << "extern const void * arr_dmmv_id_"   << tname << "_" << btype << "_f32_data[3];\n";
+        hdr << "extern const uint64_t arr_dmmv_id_" << tname << "_" << btype << "_f32_len[3];\n";
+        if (basename(input_filepath) == "mul_mat_vec.comp") {
+            src << "const void * arr_dmmv_id_"   << tname << "_" << btype << "_f32_data[3] = {mul_mat_vec_id_" << tname << "_" << btype << "_f32_data, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_data, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_no_shmem_data};\n";
+            src << "const uint64_t arr_dmmv_id_" << tname << "_" << btype << "_f32_len[3] =  {mul_mat_vec_id_" << tname << "_" << btype << "_f32_len,  mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_len, mul_mat_vec_id_"  << tname << "_" << btype << "_f32_subgroup_no_shmem_len};\n";
+        }
     }
     }
 
diff --git a/ml/backend/ggml/ggml/src/ggml.c b/ml/backend/ggml/ggml/src/ggml.c
index 923c33d0..1c9e0bc0 100644
--- a/ml/backend/ggml/ggml/src/ggml.c
+++ b/ml/backend/ggml/ggml/src/ggml.c
@@ -940,6 +940,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "COS",
     "SUM",
     "SUM_ROWS",
+    "CUMSUM",
     "MEAN",
     "ARGMAX",
     "COUNT_EQUAL",
@@ -994,7 +995,10 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
+    "TOP_K",
     "LEAKY_RELU",
+    "TRI",
+    "FILL",
 
     "FLASH_ATTN_EXT",
     "FLASH_ATTN_BACK",
@@ -1007,6 +1011,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "RWKV_WKV6",
     "GATED_LINEAR_ATTN",
     "RWKV_WKV7",
+    "SOLVE_TRI",
 
     "UNARY",
 
@@ -1024,7 +1029,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 90, "GGML_OP_COUNT != 90");
+static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1044,6 +1049,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cos(x)",
     "Σx",
     "Σx_k",
+    "cumsum(x)",
     "Σx/n",
     "argmax(x)",
     "count_equal(x)",
@@ -1098,7 +1104,10 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
+    "top_k(x)",
     "leaky_relu(x)",
+    "tri(x)",
+    "fill(x, c)",
 
     "flash_attn_ext(x)",
     "flash_attn_back(x)",
@@ -1111,6 +1120,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "rwkv_wkv6(k, v, r, tf, td, s)",
     "gated_linear_attn(k, v, q, gate, s)",
     "rwkv_wkv7(r, w, k, v, a, b, s)",
+    "A X = B, A triangular, solve X",
 
     "unary(x)",
 
@@ -1128,7 +1138,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 90, "GGML_OP_COUNT != 90");
+static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1147,6 +1157,8 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "HARDSWISH",
     "HARDSIGMOID",
     "EXP",
+    "EXPM1",
+    "SOFTPLUS",
     "GELU_ERF",
     "XIELU",
     "FLOOR",
@@ -1155,7 +1167,7 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "TRUNC",
 };
 
-static_assert(GGML_UNARY_OP_COUNT == 20, "GGML_UNARY_OP_COUNT != 20");
+static_assert(GGML_UNARY_OP_COUNT == 22, "GGML_UNARY_OP_COUNT != 22");
 
 static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
     "REGLU",
@@ -2263,6 +2275,30 @@ struct ggml_tensor * ggml_log_inplace(
     return ggml_log_impl(ctx, a, true);
 }
 
+struct ggml_tensor * ggml_expm1(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_EXPM1);
+}
+
+struct ggml_tensor * ggml_expm1_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXPM1);
+}
+
+struct ggml_tensor * ggml_softplus(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_SOFTPLUS);
+}
+
+struct ggml_tensor * ggml_softplus_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SOFTPLUS);
+}
+
 // ggml_sin
 
 static struct ggml_tensor * ggml_sin_impl(
@@ -2346,6 +2382,21 @@ struct ggml_tensor * ggml_sum_rows(
     return result;
 }
 
+// ggml_cumsum
+
+struct ggml_tensor * ggml_cumsum(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    GGML_ASSERT(a->type == GGML_TYPE_F32);
+
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+    result->op     = GGML_OP_CUMSUM;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_mean
 
 struct ggml_tensor * ggml_mean(
@@ -2673,8 +2724,8 @@ struct ggml_tensor * ggml_xielu(
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params_i32(result, 0, (int32_t) GGML_UNARY_OP_XIELU);
-    ggml_set_op_params_f32(result, 1, beta + ggml_softplus(alpha_n));
-    ggml_set_op_params_f32(result, 2, ggml_softplus(alpha_p));
+    ggml_set_op_params_f32(result, 1, beta + ggml_compute_softplus_f32(alpha_n));
+    ggml_set_op_params_f32(result, 2, ggml_compute_softplus_f32(alpha_p));
     ggml_set_op_params_f32(result, 3, beta);
     ggml_set_op_params_f32(result, 4, eps);
 
@@ -4992,28 +5043,6 @@ struct ggml_tensor * ggml_roll(
     return result;
 }
 
-// ggml_arange
-
-struct ggml_tensor * ggml_arange(
-        struct ggml_context * ctx,
-        float                 start,
-        float                 stop,
-        float                 step) {
-    GGML_ASSERT(stop > start);
-
-    const int64_t steps = (int64_t) ceilf((stop - start) / step);
-
-    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
-
-    ggml_set_op_params_f32(result, 0, start);
-    ggml_set_op_params_f32(result, 1, stop);
-    ggml_set_op_params_f32(result, 2, step);
-
-    result->op = GGML_OP_ARANGE;
-
-    return result;
-}
-
 // ggml_timestep_embedding
 
 struct ggml_tensor * ggml_timestep_embedding(
@@ -5033,6 +5062,61 @@ struct ggml_tensor * ggml_timestep_embedding(
     return result;
 }
 
+// ggml_tri
+
+struct ggml_tensor * ggml_tri(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    enum ggml_tri_type    type) {
+    GGML_ASSERT(a->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(a));
+    GGML_ASSERT(a->ne[0] == a->ne[1]);
+
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+    ggml_set_op_params_i32(result, 0, type);
+
+    result->op = GGML_OP_TRI;
+    result->src[0] = a;
+
+    return result;
+}
+
+// ggml_fill
+
+static struct ggml_tensor * ggml_fill_impl(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    float                 c,
+    bool                  inplace) {
+    GGML_ASSERT(a->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(a));
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_set_op_params_f32(result, 0, c);
+
+    result->op = GGML_OP_FILL;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_fill(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    float                 c) {
+    return ggml_fill_impl(ctx, a, c, false);
+}
+
+struct ggml_tensor * ggml_fill_inplace(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    float                 c) {
+    return ggml_fill_impl(ctx, a, c, true);
+}
+
 // ggml_argsort
 
 struct ggml_tensor * ggml_argsort(
@@ -5040,6 +5124,7 @@ struct ggml_tensor * ggml_argsort(
         struct ggml_tensor   * a,
         enum ggml_sort_order   order) {
     GGML_ASSERT(a->ne[0] <= INT32_MAX);
+
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -5050,6 +5135,24 @@ struct ggml_tensor * ggml_argsort(
     return result;
 }
 
+// ggml_argsort_top_k
+
+struct ggml_tensor * ggml_argsort_top_k(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   k) {
+    GGML_ASSERT(a->ne[0] >= k);
+
+    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
+
+    result = ggml_view_4d(ctx, result,
+                k, result->ne[1], result->ne[2], result->ne[3],
+                   result->nb[1], result->nb[2], result->nb[3],
+                0);
+
+    return result;
+}
+
 // ggml_top_k
 
 struct ggml_tensor * ggml_top_k(
@@ -5058,12 +5161,32 @@ struct ggml_tensor * ggml_top_k(
         int                   k) {
     GGML_ASSERT(a->ne[0] >= k);
 
-    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_I32, k, a->ne[1], a->ne[2], a->ne[3]);
 
-    result = ggml_view_4d(ctx, result,
-                k, result->ne[1], result->ne[2], result->ne[3],
-                   result->nb[1], result->nb[2], result->nb[3],
-                0);
+    result->op     = GGML_OP_TOP_K;
+    result->src[0] = a;
+
+    return result;
+}
+
+// ggml_arange
+
+struct ggml_tensor * ggml_arange(
+        struct ggml_context * ctx,
+        float                 start,
+        float                 stop,
+        float                 step) {
+    GGML_ASSERT(stop > start);
+
+    const int64_t steps = (int64_t) ceilf((stop - start) / step);
+
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
+
+    ggml_set_op_params_f32(result, 0, start);
+    ggml_set_op_params_f32(result, 1, stop);
+    ggml_set_op_params_f32(result, 2, step);
+
+    result->op = GGML_OP_ARANGE;
 
     return result;
 }
@@ -5887,6 +6010,41 @@ struct ggml_tensor * ggml_opt_step_sgd(
     return result;
 }
 
+// solve_tri
+
+struct ggml_tensor * ggml_solve_tri(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  left,
+        bool                  lower,
+        bool                  uni) {
+    GGML_ASSERT(a->type == GGML_TYPE_F32);
+    GGML_ASSERT(b->type == GGML_TYPE_F32);
+
+    // A must be square and lower diagonal
+    GGML_ASSERT(a->ne[0] == a->ne[1]);
+    // B must have same outer dimension as A
+    GGML_ASSERT(a->ne[1] == b->ne[1]);
+
+    // batch dimensions must be equal
+    GGML_ASSERT(a->ne[2] == b->ne[2]);
+    GGML_ASSERT(a->ne[3] == b->ne[3]);
+
+    GGML_ASSERT(ggml_is_contiguous(a));
+    GGML_ASSERT(ggml_is_contiguous(b));
+
+    GGML_ASSERT(lower && left && !uni); // TODO: support other variants
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, b->ne[0], b->ne[1], b->ne[2], b->ne[3]);
+
+    result->op     = GGML_OP_SOLVE_TRI;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 struct ggml_hash_set ggml_hash_set_new(size_t size) {
@@ -6459,6 +6617,16 @@ static void ggml_compute_backward(
                         ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, tensor, grad));
                     }
                 } break;
+                case GGML_UNARY_OP_EXPM1: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_exp(ctx, src0)));
+                    }
+                } break;
+                case GGML_UNARY_OP_SOFTPLUS: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_sigmoid(ctx, src0)));
+                    }
+                } break;
                 default: {
                     fprintf(stderr, "%s: unsupported unary op for backward pass: %s\n",
                         __func__, ggml_unary_op_name(ggml_get_unary_op(tensor)));