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

Update GGML to b6646 (#12245)

Browse Source 
Notable EOLs with this change:
- MacOS v12 and v13 are no longer supported (v14+ required)
- AMD gfx900 and gfx906 are no longer supported
This commit is contained in:
Daniel Hiltgen
2025-10-02 14:47:10 -07:00
committed by GitHub
parent fdb109469f
commit c68f367ef6
326 changed files with 30615 additions and 20624 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
llama/patches/0004-solar-pro.patch
View File

@@ -15,10 +15,10 @@ adds support for the Solar Pro architecture
7 files changed, 248 insertions(+)
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 18dcc6dd..4b285646 100644
index 4e8d54c4..f98a3574 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -78,6 +78,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
@@ -81,6 +81,7 @@ 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" },
@@ -26,15 +26,15 @@ index 18dcc6dd..4b285646 100644
{ LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
{ LLM_ARCH_PLM, "plm" },
{ LLM_ARCH_BAILINGMOE, "bailingmoe" },
@@ -164,6 +165,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
@@ -177,6 +178,7 @@ 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" },
+ { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION, "%s.attention.block_skip_connection" },
{ LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
{ LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
@@ -1794,6 +1796,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
@@ -1879,6 +1881,24 @@ 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 +59,7 @@ index 18dcc6dd..4b285646 100644
{
LLM_ARCH_WAVTOKENIZER_DEC,
{
@@ -2219,6 +2239,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
@@ -2368,6 +2388,7 @@ 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 +68,10 @@ index 18dcc6dd..4b285646 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 7af587e7..3ea994c7 100644
index b5c6f3d7..aa8e0e7b 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -82,6 +82,7 @@ enum llm_arch {
@@ -85,6 +85,7 @@ enum llm_arch {
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_GRANITE_HYBRID,
LLM_ARCH_CHAMELEON,
@@ -79,15 +79,15 @@ index 7af587e7..3ea994c7 100644
LLM_ARCH_WAVTOKENIZER_DEC,
LLM_ARCH_PLM,
LLM_ARCH_BAILINGMOE,
@@ -168,6 +169,7 @@ enum llm_kv {
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
LLM_KV_ATTENTION_SLIDING_WINDOW,
@@ -181,6 +182,7 @@ enum llm_kv {
LLM_KV_ATTENTION_SCALE,
LLM_KV_ATTENTION_OUTPUT_SCALE,
LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
+ LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
LLM_KV_ATTENTION_KEY_LENGTH_MLA,
LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
@@ -394,6 +396,7 @@ enum llm_tensor {
@@ -417,6 +419,7 @@ enum llm_tensor {
LLM_TENSOR_ENC_OUTPUT_NORM,
LLM_TENSOR_CLS,
LLM_TENSOR_CLS_OUT,
@@ -96,10 +96,10 @@ index 7af587e7..3ea994c7 100644
LLM_TENSOR_CONVNEXT_DW,
LLM_TENSOR_CONVNEXT_NORM,
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index 7a06368d..35fc054f 100644
index c04ac58f..24a515a0 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -146,6 +146,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
@@ -147,6 +147,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
}
@@ -115,10 +115,10 @@ index 7a06368d..35fc054f 100644
if (il < n_layer) {
return swa_layers[il];
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index bd231224..29bd9056 100644
index 0fe4b569..eb13709f 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -62,6 +62,8 @@ struct llama_hparams {
@@ -64,6 +64,8 @@ struct llama_hparams {
std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
@@ -127,7 +127,7 @@ index bd231224..29bd9056 100644
uint32_t n_layer_dense_lead = 0;
uint32_t n_lora_q = 0;
uint32_t n_lora_kv = 0;
@@ -220,6 +222,9 @@ struct llama_hparams {
@@ -236,6 +238,9 @@ struct llama_hparams {
uint32_t n_pos_per_embd() const;
@@ -135,10 +135,10 @@ index bd231224..29bd9056 100644
+ bool n_bskcn(uint32_t n, uint32_t il) const;
+
bool is_swa(uint32_t il) const;
};
bool has_kv(uint32_t il) const;
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index f71c40f8..7eab9b68 100644
index 8182a9ad..daef900c 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -465,6 +465,7 @@ namespace GGUFMeta {
@@ -150,10 +150,10 @@ index f71c40f8..7eab9b68 100644
llama_model_loader::llama_model_loader(
const std::string & fname,
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 58ca7df7..280129e1 100644
index 2470f878..0398b553 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -1706,6 +1706,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
@@ -1845,6 +1845,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
@@ -175,7 +175,7 @@ index 58ca7df7..280129e1 100644
case LLM_ARCH_WAVTOKENIZER_DEC:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -4793,6 +4808,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
@@ -5113,6 +5128,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@@ -210,7 +210,7 @@ index 58ca7df7..280129e1 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);
@@ -15495,6 +15538,165 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
@@ -16273,6 +16316,165 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
}
};
@@ -229,7 +229,7 @@ index 58ca7df7..280129e1 100644
+ 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_unified();
+ 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;
+
@@ -316,7 +316,7 @@ index 58ca7df7..280129e1 100644
+
+ cur = build_attn(inp_attn,
+ model.layers[il].wo, model.layers[il].bo,
+ Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+ Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+ cb(cur, "attn_out", il);
+ }
+
@@ -376,7 +376,7 @@ index 58ca7df7..280129e1 100644
// ref: https://github.com/facebookresearch/chameleon
// based on the original build_llama() function, changes:
// * qk-norm
@@ -18439,6 +18641,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
@@ -19552,6 +19754,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_chameleon>(*this, params);
} break;
@@ -387,7 +387,7 @@ index 58ca7df7..280129e1 100644
case LLM_ARCH_WAVTOKENIZER_DEC:
{
llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params);
@@ -18652,6 +18858,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
@@ -19770,6 +19976,7 @@ 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:
@@ -396,10 +396,10 @@ index 58ca7df7..280129e1 100644
case LLM_ARCH_NEO_BERT:
case LLM_ARCH_SMOLLM3:
diff --git a/src/llama-model.h b/src/llama-model.h
index 6fcd74d5..09964533 100644
index d73ce969..c086f94e 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -70,6 +70,7 @@ enum llm_type {
@@ -76,6 +76,7 @@ enum llm_type {
LLM_TYPE_15B,
LLM_TYPE_16B,
LLM_TYPE_20B,
@@ -407,7 +407,7 @@ index 6fcd74d5..09964533 100644
LLM_TYPE_27B,
LLM_TYPE_30B,
LLM_TYPE_32B,
@@ -367,6 +368,8 @@ struct llama_layer {
@@ -380,6 +381,8 @@ struct llama_layer {
// openai-moe
struct ggml_tensor * attn_sinks = nullptr;