CUDA: fix FA logic for PTX 7.0 and CC >= 7.5 (llama/12222)

This commit is contained in:
Johannes Gäßler 2025-03-06 18:45:09 +01:00 committed by Georgi Gerganov
parent b9eab73fa2
commit e3c85e75bd

View File

@ -310,7 +310,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
}
// The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
if (cc == GGML_CUDA_CC_VOLTA) {
if (fp16_mma_available(cc) && !new_mma_available(cc)) {
ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
return;
}