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

MUSA: support ARM64 and enable dp4a .etc (#11843)

Browse source 
* MUSA:  support ARM64 and enable __dp4a .etc

* fix cross entropy loss op for musa

* update

* add cc info log for musa

* add comment for the MUSA .cc calculation block

---------

Co-authored-by: Bodhi Hu <huaishun.hu@mthreads.com>
This commit is contained in:
Bodhi 2025-02-21 15:46:23 +08:00 • committed by GitHub
parent ee02ad02c5
commit 0b3863ff95
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 25 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml/src/ggml-cuda/common.cuh
View file

@ -411,13 +411,13 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
return __dp4a(a, b, c);
#else // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
#else // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
const int8_t * a8 = (const int8_t *) &a;
const int8_t * b8 = (const int8_t *) &b;
return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
}