From f455ead8aaed6159077ab629b3a9e6c34d388fbf Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Mon, 5 May 2025 08:31:55 +0300
Subject: [PATCH] Fix DeepSeek FA

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

diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index fce66c20..ea52fa02 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -521,6 +521,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     // K and V head sizes.
     if (K->ne[0] != V->ne[0]) {
         ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+        return;
     }
 
     ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);