From e3c85e75bd69065420d79f41b0899b4617c134ab Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Thu, 6 Mar 2025 18:45:09 +0100
Subject: [PATCH] CUDA: fix FA logic for PTX 7.0 and CC >= 7.5 (llama/12222)

---
 ggml/src/ggml-cuda/fattn.cu | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 24f97305..2e72fc8f 100644
--- 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;
     }