Fix DeepSeek FA (#382)

ikawrakow · Iwan Kawrakow · web-flow · commit e3fec17347d5 · 2025-05-05T08:39:10.000+03:00
Co-authored-by: Iwan Kawrakow &lt;iwan.kawrakow@gmail.com&gt;
diff --git 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);

Original file line number	Diff line number	Diff line change
`@@ -521,6 +521,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst`
`521`	`521`	`// K and V head sizes.`
`522`	`522`	`if (K->ne[0] != V->ne[0]) {`
`523`	`523`	`ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);`
	`524`	`+ return;`
`524`	`525`	`}`
`525`	`526`
`526`	`527`	`ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);`