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

rocBLAS: Avoid fp32->fp16->fp32 conversion on cdna (#11356)

Browse source
This commit is contained in:
uvos 2025-01-24 17:50:49 +01:00 • committed by GitHub
parent 9755129c27
commit 9fbadaef4f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 40 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ggml/src/ggml-cuda/mmvq.cu
View file

@ -142,7 +142,7 @@ static void mul_mat_vec_q_cuda(
int64_t nwarps = 1;
int64_t rows_per_cuda_block = 1;
if (ggml_cuda_info().devices[id].cc < GGML_CUDA_CC_CDNA || ggml_cuda_info().devices[id].cc == GGML_CUDA_CC_RDNA1) { // NVIDIA and AMD older than RDNA2 but not CDNA
if (ggml_cuda_info().devices[id].cc < GGML_CUDA_CC_RDNA2) { // NVIDIA and AMD older than RDNA2
switch(ncols_y) {
case 1:
nwarps = 4;
@ -166,6 +166,7 @@ static void mul_mat_vec_q_cuda(
break;
}
}
const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
const dim3 block_nums(nblocks, 1, 1);
const dim3 block_dims(WARP_SIZE, nwarps, 1);