[Async Scheduling] Move the auto-selection of the backend to before the default selection.

Signed-off-by: Lehua Ding <[email protected]>

Author: lhtin

vllm/engine/arg_utils.py

@@ -1287,11 +1287,6 @@ def create_engine_config(
             is not None) else ParallelConfig.data_parallel_rpc_port
 
         if self.async_scheduling:
-            # Async scheduling does not work with the uniprocess backend.
-            if self.distributed_executor_backend is None:
-                self.distributed_executor_backend = "mp"
-                logger.info("Defaulting to mp-based distributed executor "
-                            "backend for async scheduling.")
             if self.pipeline_parallel_size > 1:
                 raise ValueError("Async scheduling is not supported with "
                                  "pipeline-parallel-size > 1.")
@@ -1344,6 +1339,14 @@ def create_engine_config(
             _api_process_rank=self._api_process_rank,
         )
 
+        if self.async_scheduling and (
+                parallel_config.distributed_executor_backend
+                not in ("mp", "ray")):
+            raise ValueError(
+                "Currently, async scheduling is only support `mp` "
+                "distributed executor backend, but you choose "
+                f"`{parallel_config.distributed_executor_backend}`.")
+
         speculative_config = self.create_speculative_config(
             target_model_config=model_config,
             target_parallel_config=parallel_config,