public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-17 20:03:45 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2026-07-10 13:06:42 +03:00
parent afa6bffd2b
commit 289ecb05c9
32 changed files with 4809 additions and 313 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/talk-llama/llama-arch.cpp
+57
View File
@@ -77,6 +77,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
{ LLM_ARCH_DEEPSEEK2OCR, "deepseek2-ocr" },
{ LLM_ARCH_DEEPSEEK32, "deepseek32" },
{ LLM_ARCH_DEEPSEEK4, "deepseek4" },
{ LLM_ARCH_CHATGLM, "chatglm" },
{ LLM_ARCH_GLM4, "glm4" },
{ LLM_ARCH_GLM4_MOE, "glm4moe" },
@@ -129,6 +130,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_PANGU_EMBED, "pangu-embedded" },
{ LLM_ARCH_MISTRAL3, "mistral3" },
{ LLM_ARCH_EAGLE3, "eagle3" },
{ LLM_ARCH_DFLASH, "dflash" },
{ LLM_ARCH_MISTRAL4, "mistral4" },
{ LLM_ARCH_PADDLEOCR, "paddleocr" },
{ LLM_ARCH_MIMO2, "mimo2" },
@@ -249,9 +251,19 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, "%s.attention.indexer.head_count" },
{ LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, "%s.attention.indexer.key_length" },
{ LLM_KV_ATTENTION_INDEXER_TOP_K, "%s.attention.indexer.top_k" },
{ LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT, "%s.attention.output_group_count" },
{ LLM_KV_ATTENTION_OUTPUT_LORA_RANK, "%s.attention.output_lora_rank" },
{ LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE, "%s.attention.compress_rope_freq_base" },
{ LLM_KV_ATTENTION_COMPRESS_RATIOS, "%s.attention.compress_ratios" },
{ LLM_KV_ATTENTION_SHARED_KV_LAYERS, "%s.attention.shared_kv_layers" },
{ LLM_KV_ATTENTION_RECURRENT_LAYERS, "%s.attention.recurrent_layers" },
{ LLM_KV_HYPER_CONNECTION_COUNT, "%s.hyper_connection.count" },
{ LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, "%s.hyper_connection.sinkhorn_iterations" },
{ LLM_KV_HYPER_CONNECTION_EPSILON, "%s.hyper_connection.epsilon" },
{ LLM_KV_HASH_LAYER_COUNT, "%s.hash_layer_count" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_COUNT_SWA, "%s.rope.dimension_count_swa" },
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
@@ -439,6 +451,23 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" },
{ LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" },
{ LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" },
{ LLM_TENSOR_ATTN_KV, "blk.%d.attn_kv" },
{ LLM_TENSOR_ATTN_KV_NORM, "blk.%d.attn_kv_a_norm" },
{ LLM_TENSOR_ATTN_OUT_A, "blk.%d.attn_output_a" },
{ LLM_TENSOR_ATTN_OUT_B, "blk.%d.attn_output_b" },
{ LLM_TENSOR_HC_HEAD_FN, "output_hc_fn" },
{ LLM_TENSOR_HC_HEAD_BASE, "output_hc_base" },
{ LLM_TENSOR_HC_HEAD_SCALE, "output_hc_scale" },
{ LLM_TENSOR_HC_ATTN_FN, "blk.%d.hc_attn_fn" },
{ LLM_TENSOR_HC_ATTN_BASE, "blk.%d.hc_attn_base" },
{ LLM_TENSOR_HC_ATTN_SCALE, "blk.%d.hc_attn_scale" },
{ LLM_TENSOR_HC_FFN_FN, "blk.%d.hc_ffn_fn" },
{ LLM_TENSOR_HC_FFN_BASE, "blk.%d.hc_ffn_base" },
{ LLM_TENSOR_HC_FFN_SCALE, "blk.%d.hc_ffn_scale" },
{ LLM_TENSOR_ATTN_COMPRESSOR_WKV, "blk.%d.attn_compressor_kv" },
{ LLM_TENSOR_ATTN_COMPRESSOR_WGATE, "blk.%d.attn_compressor_gate" },
{ LLM_TENSOR_ATTN_COMPRESSOR_APE, "blk.%d.attn_compressor_ape" },
{ LLM_TENSOR_ATTN_COMPRESSOR_NORM, "blk.%d.attn_compressor_norm" },
{ LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "per_layer_token_embd" },
{ LLM_TENSOR_PER_LAYER_MODEL_PROJ, "per_layer_model_proj" },
{ LLM_TENSOR_PER_LAYER_PROJ_NORM, "per_layer_proj_norm" },
@@ -565,6 +594,11 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_INDEXER_PROJ, "blk.%d.indexer.proj" },
{ LLM_TENSOR_INDEXER_ATTN_K, "blk.%d.indexer.attn_k" },
{ LLM_TENSOR_INDEXER_ATTN_Q_B, "blk.%d.indexer.attn_q_b" },
{ LLM_TENSOR_INDEXER_COMPRESSOR_WKV, "blk.%d.indexer_compressor_kv" },
{ LLM_TENSOR_INDEXER_COMPRESSOR_WGATE, "blk.%d.indexer_compressor_gate" },
{ LLM_TENSOR_INDEXER_COMPRESSOR_APE, "blk.%d.indexer_compressor_ape" },
{ LLM_TENSOR_INDEXER_COMPRESSOR_NORM, "blk.%d.indexer_compressor_norm" },
{ LLM_TENSOR_FFN_GATE_TID2EID, "blk.%d.ffn_gate_tid2eid" },
{ LLM_TENSOR_MASKED_EMBD_CENTROIDS, "masked_embd_centroids" },
{ LLM_TENSOR_MASKED_EMBD_ORDERING, "masked_embd_ordering" },
{ LLM_TENSOR_FC, "fc" },
@@ -615,6 +649,23 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV_A_MQA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_KV_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_OUT_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_OUT_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_HC_HEAD_FN, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_HC_HEAD_BASE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_ADD}},
{LLM_TENSOR_HC_HEAD_SCALE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
{LLM_TENSOR_HC_ATTN_FN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_HC_ATTN_BASE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_HC_ATTN_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_HC_FFN_FN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_HC_FFN_BASE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_HC_FFN_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_COMPRESSOR_WKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_COMPRESSOR_WGATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_COMPRESSOR_APE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_ATTN_COMPRESSOR_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_K_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_V_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_SINKS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SCALE}},
@@ -778,6 +829,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_INDEXER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_COMPRESSOR_WKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_COMPRESSOR_WGATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_INDEXER_COMPRESSOR_APE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_INDEXER_COMPRESSOR_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_FFN_GATE_TID2EID, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
{LLM_TENSOR_NEXTN_PROJ_PRE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_NEXTN_PROJ_POST, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
// NextN/MTP tensors are stored per-block (blk.%d.nextn.*) even though only the
@@ -932,6 +988,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
case LLM_ARCH_OLMOE:
case LLM_ARCH_DEEPSEEK2:
case LLM_ARCH_DEEPSEEK32:
case LLM_ARCH_DEEPSEEK4:
case LLM_ARCH_GLM_DSA:
case LLM_ARCH_BITNET:
case LLM_ARCH_T5: