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

38
ggml/src/ggml-cuda/mmq.cu
View File

@ -333,23 +333,29 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
}
if (amd_wmma_available(cc)) {
// High expert counts almost always better on MMQ
// due to a large amount of graph splits
// https://github.com/ggml-org/llama.cpp/pull/18202
if (n_experts >= 64) {
return true;
}
switch (type) {
// These quants are really bad on MMQ
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q6_K:
// These quants are usually worse but not always
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
return ne11 <= 128;
default:
// 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
// due to a large amount of graph splits
// https://github.com/ggml-org/llama.cpp/pull/18202
if (n_experts >= 64) {
return true;
}
switch (type) {
// These quants are really bad on MMQ
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q6_K:
// These quants are usually worse but not always
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
return ne11 <= 128;
default:
return true;
}
} else {
return true;
}
}