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

fix AMD

Browse source
This commit is contained in:
JohannesGaessler 2024-02-10 09:54:36 +01:00
parent 97f8a7a2fa
commit 2bb97fca5e
1 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml-cuda.cu
View file

@ -5319,13 +5319,13 @@ static __global__ void mul_mat_vec_q(
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
#if __CUDA_ARCH__ < CC_RDNA2
constexpr int nwarps = ncols_y <= 4 ? 4 : 2;
constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
#else
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
constexpr int nwarps = 1;
constexpr int rows_per_cuda_block = 1;
#endif // __CUDA_ARCH__ < CC_RDNA2
#else
constexpr int nwarps = ncols_y <= 4 ? 4 : 2;
constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;