EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

musa: refine compute capability (#12493)

Browse Source 
* musa: refine compute capability

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Address review comments

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

---------

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
This commit is contained in:
R0CKSTAR
2025-03-22 17:11:37 +08:00
committed by GitHub
parent eddfb43850
commit fac63a3d78
5 changed files with 44 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-cuda/mmq.cu
View File

@@ -28,7 +28,7 @@ void ggml_cuda_op_mul_mat_q(
// Also its fixup needs to allocate a temporary buffer in the memory pool.
// There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
const bool use_stream_k = ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA &&
cc < GGML_CUDA_CC_OFFSET_AMD && src1_ncols == ne11;
GGML_CUDA_CC_IS_NVIDIA(cc) && src1_ncols == ne11;
const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
switch (src0->type) {
@@ -145,7 +145,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
return true;
#endif //GGML_CUDA_FORCE_MMQ
if (cc < GGML_CUDA_CC_OFFSET_AMD) {
if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}