ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

HIP: add GGML_CUDA_CC_IS_* for amd familys as increasing cc archtectures for amd gpus are not supersets of eatch other (#11601)

Browse source 
This fixes a bug where RDNA1 gpus other than gfx1010 where not handled correctly
This commit is contained in:
uvos 2025-02-02 22:08:05 +01:00 • committed by GitHub
parent 90f9b88afb
commit 4d0598e144
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 11 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml/src/ggml-cuda/mmq.cu
View file

@ -148,5 +148,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
return cc < GGML_CUDA_CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}
return (cc < GGML_CUDA_CC_RDNA3 && cc != GGML_CUDA_CC_CDNA && cc != GGML_CUDA_CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc) && !GGML_CUDA_CC_IS_GCN(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}