ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix bug in MMVQ kernel

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-05-23 08:02:47 +03:00
parent 7f2edd1a85
commit 193a15b465
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
ggml/src/ggml-cuda/mmvq.cu
View File

@@ -72,10 +72,13 @@ static __device__ void mul_mat_vec_q(
constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
//int64_t rows_per_cuda_block = ggml_cuda_info().devices[id].cc < CC_RDNA2 ?
// ncols_y < 4 ? 1 : 2 : 1;
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
constexpr int rows_per_cuda_block = 1;
#else
constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
constexpr int rows_per_cuda_block = ncols_y < 4 ? 1 : 2;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;