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

diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 24f973056aa9a78a3cf1347e64adba2bb678ec00..2e72fc8fd380bb68d7cbdb2500b056925c5504a7 100644 (file)
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -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;
     }