[Frontend] Improve Startup Failure UX

tests/entrypoints/openai/test_mp_api_server.py

@@ -1,3 +1,5 @@
+import time
+
 import pytest
 
 from vllm.entrypoints.openai.api_server import build_async_engine_client
@@ -8,19 +10,20 @@
 @pytest.mark.asyncio
 async def test_mp_crash_detection():
 
-    with pytest.raises(RuntimeError) as excinfo:
-        parser = FlexibleArgumentParser(
-            description="vLLM's remote OpenAI server.")
-        parser = make_arg_parser(parser)
-        args = parser.parse_args([])
-        # use an invalid tensor_parallel_size to trigger the
-        # error in the server
-        args.tensor_parallel_size = 65536
-
-        async with build_async_engine_client(args):
-            pass
-    assert "The server process died before responding to the readiness probe"\
-          in str(excinfo.value)
+    parser = FlexibleArgumentParser(description="vLLM's remote OpenAI server.")
+    parser = make_arg_parser(parser)
+    args = parser.parse_args([])
+    # use an invalid tensor_parallel_size to trigger the
+    # error in the server
+    args.tensor_parallel_size = 65536
+
+    start = time.perf_counter()
+    async with build_async_engine_client(args):
+        pass
+    end = time.perf_counter()
+
+    assert end - start < 60, ("Expected vLLM to gracefully shutdown in <60s "
+                              "if there is an error in the startup.")
 
 
 @pytest.mark.asyncio

vllm/entrypoints/openai/api_server.py

@@ -8,7 +8,7 @@
 from argparse import Namespace
 from contextlib import asynccontextmanager
 from http import HTTPStatus
-from typing import AsyncIterator, Set
+from typing import AsyncIterator, Optional, Set
 
 from fastapi import APIRouter, FastAPI, Request
 from fastapi.exceptions import RequestValidationError
@@ -60,6 +60,7 @@
 openai_serving_tokenization: OpenAIServingTokenization
 prometheus_multiproc_dir: tempfile.TemporaryDirectory
 
+# Cannot use __name__ (https://github.com/vllm-project/vllm/pull/4765)
 logger = init_logger('vllm.entrypoints.openai.api_server')
 
 _running_tasks: Set[asyncio.Task] = set()
@@ -94,7 +95,15 @@ async def _force_log():
 
 @asynccontextmanager
 async def build_async_engine_client(
-        args: Namespace) -> AsyncIterator[AsyncEngineClient]:
+        args: Namespace) -> AsyncIterator[Optional[AsyncEngineClient]]:
+    """
+    Create AsyncEngineClient, either:
+        - in-process using the AsyncLLMEngine Directly
+        - multiprocess using AsyncLLMEngine RPC
+
+    Returns the Client or None if the creation failed.
+    """
+
     # Context manager to handle async_engine_client lifecycle
     # Ensures everything is shutdown and cleaned up on error/exit
     global engine_args
@@ -157,11 +166,13 @@ async def build_async_engine_client(
                 try:
                     await rpc_client.setup()
                     break
-                except TimeoutError as e:
+                except TimeoutError:
                     if not rpc_server_process.is_alive():
-                        raise RuntimeError(
-                            "The server process died before "
-                            "responding to the readiness probe") from e
+                        logger.error(
+                            "RPCServer process died before responding "
+                            "to readiness probe")
+                        yield None
+                        return
 
             yield async_engine_client
         finally:
@@ -410,6 +421,10 @@ async def run_server(args, **uvicorn_kwargs) -> None:
     logger.info("args: %s", args)
 
     async with build_async_engine_client(args) as async_engine_client:
+        # If None, creation of the client failed and we exit.
+        if async_engine_client is None:
+            return
+
         app = await init_app(async_engine_client, args)
 
         shutdown_task = await serve_http(