From e3fec17347d5e85c262f857aab82fa1af3c76304 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Mon, 5 May 2025 08:39:10 +0300 Subject: Fix DeepSeek FA (#382) Co-authored-by: Iwan Kawrakow --- ggml/src/ggml-cuda/fattn.cu | 1 + 1 file changed, 1 insertion(+) (limited to 'ggml/src') 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); -- cgit v1.2.3