mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml update to b7108 (#12992)

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

This reverts commit 3a9e8e9fd4.

* 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
This commit is contained in:
Daniel Hiltgen
2025-12-03 19:43:29 -08:00
committed by GitHub
parent 854d40edc5
commit 0cf7794b16
303 changed files with 32711 additions and 23435 deletions
Download Patch File Download Diff File Expand all files Collapse all files

172
llama/patches/0004-solar-pro.patch
View File

@@ -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;
+}