CUDA: fix FA logic for PTX 7.0 and CC >= 7.5 (ggml-org#12222)

JohannesGaessler · web-flow · commit 5220a16d1856 · 2025-03-06T18:45:09.000+01:00
diff --git 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;
     }

Original file line number	Diff line number	Diff line change
`@@ -310,7 +310,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst`
`310`	`310`	`}`
`311`	`311`
`312`	`312`	`// The MMA implementation needs Turing or newer, use the old WMMA code for Volta:`
`313`		`- if (cc == GGML_CUDA_CC_VOLTA) {`
	`313`	`+ if (fp16_mma_available(cc) && !new_mma_available(cc)) {`
`314`	`314`	`ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);`
`315`	`315`	`return;`
`316`	`316`	`}`