[Core] Eliminate parallel worker per-step task scheduling overhead (#4894)

vllm/engine/async_llm_engine.py

@@ -234,6 +234,14 @@ class _AsyncLLMEngine(LLMEngine):
         # Log stats.
         self.do_log_stats(scheduler_outputs, output)
 
+        if not request_outputs:
+            # Stop the execute model loop in parallel workers until there are
+            # more requests to process. This avoids waiting indefinitely in
+            # torch.distributed ops which may otherwise timeout, and unblocks
+            # the RPC thread in the workers so that they can process any other
+            # queued control plane messages, such as add/remove lora adapters.
+            await self.model_executor.stop_remote_worker_execution_loop_async()
+
         return request_outputs
 
     async def encode_request_async(
@@ -687,7 +695,7 @@ class AsyncLLMEngine:
             multi_modal_data: Multi modal data per request.
 
         Yields:
-            The output `EmbeddingRequestOutput` objects from the LLMEngine 
+            The output `EmbeddingRequestOutput` objects from the LLMEngine
             for the request.
 
         Details: