ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-29 10:51:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support GigaChat3 (#995)

Browse Source 
* Fixing Gigachat support

* Gigachat: CUDA FA (needs 192 x 192 for MLA = 3)

* Gigachat: CPU FA (needs 192 x 192 for MLA = 3)

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-11-24 06:55:14 +01:00
committed by GitHub
parent f6163dd58f
commit 920f424929
11 changed files with 103 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama-arch.cpp
View File

@@ -142,6 +142,8 @@ 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_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
{ LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },

2
src/llama-arch.h
View File

@@ -135,6 +135,8 @@ enum llm_kv {
LLM_KV_ATTENTION_SCALE,
LLM_KV_ATTENTION_OUTPUT_SCALE,
LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
LLM_KV_ATTENTION_KEY_LENGTH_MLA,
LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_DIMENSION_SECTIONS,

2
src/llama-build-context.cpp
View File

@@ -5931,7 +5931,7 @@ ggml_cgraph * llm_build_context::build_deepseek2() {
// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
bool is_lite = (hparams.n_layer == 27);
bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
// 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.

4
src/llama-hparams.cpp
View File

@@ -762,8 +762,10 @@ void llm_load_hparams(
for (auto& item : hparams.n_head_kv_arr) item = n_nead_kv;
hparams.n_embd_head_k = 192;
hparams.n_embd_head_v = 128;
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k);
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v);
}
bool is_lite = (hparams.n_layer == 27);
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) {

2
src/llama-load-tensors.cpp
View File

@@ -1617,7 +1617,7 @@ bool create_tensors_helper::create_arctix_tensors(const LLM_TN & tn) {
bool create_tensors_helper::create_deepseek2_tensors(const LLM_TN & tn) {
LOADING_PRELUDE
const bool is_lite = (hparams.n_layer == 27);
const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
const int64_t n_embd_head_qk_rope = hparams.n_rot;
const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;