tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-09 18:40:20 -04:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: faster Deepseek FA, add Turing support (#13435)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-05-14 16:08:20 +02:00
committed by GitHub
parent 5e7d95e22e
commit 6da34fa276
4 changed files with 277 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -3222,7 +3222,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
#endif // FLASH_ATTN_AVAILABLE
if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
if (!new_mma_available(cc) || cc < GGML_CUDA_CC_AMPERE) {
if (!new_mma_available(cc)) {
return false;
}
const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];