happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

mmq.cu: move amd wmma mmq/wmma switching behind IS_RDNA3

This commit is contained in:
Beinsezii 2026-01-02 15:05:30 -08:00
parent a435c7725b
commit 3326fa2387
1 changed files with 22 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-cuda/mmq.cu
View File

@ -333,6 +333,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
} }
if (amd_wmma_available(cc)) { if (amd_wmma_available(cc)) {
// RDNA 4 is consistently worse on rocblas
// https://github.com/ggml-org/llama.cpp/pull/18537#issuecomment-3706422301
if (GGML_CUDA_CC_IS_RDNA3(cc)) {
// High expert counts almost always better on MMQ // High expert counts almost always better on MMQ
// due to a large amount of graph splits // due to a large amount of graph splits
// https://github.com/ggml-org/llama.cpp/pull/18202 // https://github.com/ggml-org/llama.cpp/pull/18202
@ -351,6 +354,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
default: default:
return true; return true;
} }
} else {
return true;
}
} }
return (!GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; return (!GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;