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

CUDA: app option to compile without FlashAttention (#12025)

Browse source
This commit is contained in:
Johannes Gäßler 2025-02-22 20:44:34 +01:00 • committed by GitHub
parent 36c258ee92
commit a28e0d5eb1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
13 changed files with 46 additions and 31 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml/src/ggml-cuda/common.cuh
View file

@ -204,9 +204,9 @@ typedef float2 dfloat2;
#define CP_ASYNC_AVAILABLE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
#define FLASH_ATTN_AVAILABLE
#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
static bool fp16_available(const int cc) {
return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;