try old wmma

LostRuins · Feb 16, 2025 · 505aede · 505aede
1 parent 15ae98c
commit 505aede
Showing 1 changed file with 6 additions and 5 deletions.
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
@@ -263,10 +263,11 @@ 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) {
-        ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
-        return;
-    }
+    ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+    // if (cc == GGML_CUDA_CC_VOLTA) {
+    //     ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+    //     return;
+    // }
 
-    ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
+    // ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
 }