[V1] Fix non-cudagraph op name (vllm-project#10166)

Signed-off-by: Woosuk Kwon <[email protected]>
Signed-off-by: OmerD <[email protected]>

vllm/v1/worker/gpu_model_runner.py

@@ -411,7 +411,7 @@ def load_model(self) -> None:
             set_compilation_config(
                 CompilationConfig(
                     use_cudagraph=True,
-                    non_cudagraph_ops=["vllm.unified_flash_attention"],
+                    non_cudagraph_ops=["vllm.unified_v1_flash_attention"],
                     use_inductor=True,
                 ))