[Bugfix] fix OpenAI API server startup with --disable-frontend-multiprocessing (#8537)

tests/entrypoints/openai/test_basic.py

 from http import HTTPStatus
+from typing import List
 
 import openai
 import pytest
@@ -12,8 +13,44 @@ from ...utils import RemoteOpenAIServer
 MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
 
 
+@pytest.fixture(scope='module')
+def server_args(request: pytest.FixtureRequest) -> List[str]:
+    """ Provide extra arguments to the server via indirect parametrization
+
+    Usage:
+
+    >>> @pytest.mark.parametrize(
+    >>>     "server_args",
+    >>>     [
+    >>>         ["--disable-frontend-multiprocessing"],
+    >>>         [
+    >>>             "--model=NousResearch/Hermes-3-Llama-3.1-70B",
+    >>>             "--enable-auto-tool-choice",
+    >>>         ],
+    >>>     ],
+    >>>     indirect=True,
+    >>> )
+    >>> def test_foo(server, client):
+    >>>     ...
+
+    This will run `test_foo` twice with servers with:
+    - `--disable-frontend-multiprocessing`
+    - `--model=NousResearch/Hermes-3-Llama-3.1-70B --enable-auto-tool-choice`.
+
+    """
+    if not hasattr(request, "param"):
+        return []
+
+    val = request.param
+
+    if isinstance(val, str):
+        return [val]
+
+    return request.param
+
+
 @pytest.fixture(scope="module")
-def server():
+def server(server_args):
     args = [
         # use half precision for speed and memory savings in CI environment
         "--dtype",
@@ -23,6 +60,7 @@ def server():
         "--enforce-eager",
         "--max-num-seqs",
         "128",
+        *server_args,
     ]
 
     with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
@@ -35,6 +73,15 @@ async def client(server):
         yield async_client
 
 
+@pytest.mark.parametrize(
+    "server_args",
+    [
+        pytest.param([], id="default-frontend-multiprocessing"),
+        pytest.param(["--disable-frontend-multiprocessing"],
+                     id="disable-frontend-multiprocessing")
+    ],
+    indirect=True,
+)
 @pytest.mark.asyncio
 async def test_show_version(client: openai.AsyncOpenAI):
     base_url = str(client.base_url)[:-3].strip("/")
@@ -45,6 +92,15 @@ async def test_show_version(client: openai.AsyncOpenAI):
     assert response.json() == {"version": VLLM_VERSION}
 
 
+@pytest.mark.parametrize(
+    "server_args",
+    [
+        pytest.param([], id="default-frontend-multiprocessing"),
+        pytest.param(["--disable-frontend-multiprocessing"],
+                     id="disable-frontend-multiprocessing")
+    ],
+    indirect=True,
+)
 @pytest.mark.asyncio
 async def test_check_health(client: openai.AsyncOpenAI):
     base_url = str(client.base_url)[:-3].strip("/")

vllm/entrypoints/openai/api_server.py

@@ -537,8 +537,11 @@ async def run_server(args, **uvicorn_kwargs) -> None:
         raise KeyError(f"invalid tool call parser: {args.tool_call_parser} "
                        f"(chose from {{ {','.join(valide_tool_parses)} }})")
 
-    temp_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
-    temp_socket.bind(("", args.port))
+    # workaround to make sure that we bind the port before the engine is set up.
+    # This avoids race conditions with ray.
+    # see https://github.com/vllm-project/vllm/issues/8204
+    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    sock.bind(("", args.port))
 
     def signal_handler(*_) -> None:
         # Interrupt server on sigterm while initializing
@@ -552,8 +555,6 @@ async def run_server(args, **uvicorn_kwargs) -> None:
         model_config = await engine_client.get_model_config()
         init_app_state(engine_client, model_config, app.state, args)
 
-        temp_socket.close()
-
         shutdown_task = await serve_http(
             app,
             host=args.host,
@@ -564,6 +565,7 @@ async def run_server(args, **uvicorn_kwargs) -> None:
             ssl_certfile=args.ssl_certfile,
             ssl_ca_certs=args.ssl_ca_certs,
             ssl_cert_reqs=args.ssl_cert_reqs,
+            fd=sock.fileno(),
             **uvicorn_kwargs,
         )