LLM inference in C/C++
C++ ggml
llama-arch: fix DeepSeek4 APE tensor op (#25945)
helanfxz67b9b0e7f6ce
diff --git a/src/llama-arch.cpp b/src/llama-arch.cppindex 89fa097fe3..e5a9a29c7d 100644--- a/src/llama-arch.cpp+++ b/src/llama-arch.cpp@@ -666,7 +666,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {{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_APE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},{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}},@@ -833,7 +833,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {{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_APE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},{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}},