pre-commit

Author: JartX

vllm/attention/layer.py

@@ -99,14 +99,19 @@ def check_upstream_fa_availability(dtype: torch.dtype):
 
 
 def maybe_get_vit_flash_attn_backend(
-    attn_backend: _Backend, use_upstream_fa: bool,
+    attn_backend: _Backend,
+    use_upstream_fa: bool,
     attn_backend_override: _Backend | None = None,
 ) -> tuple[_Backend, Callable | None]:
     if current_platform.is_rocm():
         if envs.VLLM_ROCM_USE_AITER and envs.VLLM_ROCM_USE_AITER_MHA and on_gfx9():
             attn_backend = _Backend.ROCM_AITER_FA
 
-        elif check_upstream_fa_availability(torch.get_default_dtype()) and on_gfx9() and  attn_backend_override is None:
+        elif (
+            check_upstream_fa_availability(torch.get_default_dtype())
+            and on_gfx9()
+            and attn_backend_override is None
+        ):
             attn_backend = _Backend.FLASH_ATTN
             use_upstream_fa = True
         else: