[V1] Multiprocessing Tensor Parallel Support for v1 (#9856)

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>

vllm/v1/engine/async_llm.py

@@ -20,7 +20,7 @@ from vllm.v1.engine.async_stream import AsyncStream
 from vllm.v1.engine.core_client import EngineCoreClient
 from vllm.v1.engine.detokenizer import Detokenizer
 from vllm.v1.engine.processor import Processor
-from vllm.v1.executor.gpu_executor import GPUExecutor
+from vllm.v1.executor.abstract import Executor
 
 logger = init_logger(__name__)
 
@@ -30,7 +30,7 @@ class AsyncLLM(EngineClient):
     def __init__(
         self,
         vllm_config: VllmConfig,
-        executor_class: Type[GPUExecutor],
+        executor_class: Type[Executor],
         log_stats: bool,
         usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
         stat_loggers: Optional[Dict[str, StatLoggerBase]] = None,
@@ -119,14 +119,24 @@ class AsyncLLM(EngineClient):
     def shutdown(self):
         """Shutdown, cleaning up the background proc and IPC."""
 
-        self.engine_core.shutdown()
+        if engine_core := getattr(self, "engine_core", None):
+            engine_core.shutdown()
 
         if handler := getattr(self, "output_handler", None):
             handler.cancel()
 
     @classmethod
     def _get_executor_cls(cls, vllm_config: VllmConfig):
-        return GPUExecutor
+        distributed_executor_backend = (
+            vllm_config.parallel_config.distributed_executor_backend)
+        if distributed_executor_backend == "mp":
+            from vllm.v1.executor.multiproc_executor import MultiprocExecutor
+            executor_class = MultiprocExecutor
+        else:
+            assert (distributed_executor_backend is None)
+            from vllm.v1.executor.uniproc_executor import UniprocExecutor
+            executor_class = UniprocExecutor
+        return executor_class
 
     async def add_request(
         self,