[Frontend] Disable uvicorn signal handlers instead of racing them (#50916)

Signed-off-by: Nick Hill <nickhill123@gmail.com>

tests/entrypoints/openai/test_dp_supervisor.py

@@ -312,7 +312,7 @@ async def test_shutdown_if_supervisor_server_error_on_startup(
         await self._shutdown_event.wait()
 
     monkeypatch.setattr(dp_sup.asyncio, "get_running_loop", lambda: FakeLoop())
-    monkeypatch.setattr(dp_sup.uvicorn, "Server", FakeServer)
+    monkeypatch.setattr(dp_sup, "NoSignalServer", FakeServer)
     monkeypatch.setattr(DPSupervisor, "_shutdown_children", fake_shutdown_children)
     monkeypatch.setattr(DPSupervisor, "_start_children", fake_start_children)
     monkeypatch.setattr(DPSupervisor, "_monitor_children", fake_monitor_children)

vllm/entrypoints/launcher.py

@@ -2,8 +2,10 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import asyncio
+import contextlib
 import signal
 import socket
+from collections.abc import Generator
 from functools import partial
 from typing import Any
 
@@ -23,6 +25,18 @@ from vllm.utils.network_utils import find_process_using_port
 logger = init_logger(__name__)
 
 
+class NoSignalServer(uvicorn.Server):
+    """Uvicorn server that never installs its own SIGINT/SIGTERM handlers.
+
+    Callers register their own handlers on the event loop for graceful
+    shutdown; uvicorn's would race with and override them (see #49668).
+    """
+
+    @contextlib.contextmanager
+    def capture_signals(self) -> Generator[None, None, None]:
+        yield
+
+
 async def serve_http(
     app: FastAPI,
     sock: socket.socket | None,
@@ -73,7 +87,7 @@ async def serve_http(
     config.h11_max_incomplete_event_size = h11_max_incomplete_event_size
     config.h11_max_header_count = h11_max_header_count
     config.load()
-    server = uvicorn.Server(config)
+    server = NoSignalServer(config)
     app.state.server = server
 
     loop = asyncio.get_running_loop()
@@ -92,18 +106,6 @@ async def serve_http(
         )
     )
 
-    # make sure uvicorn server signal handler has been registered.
-    # Otherwise the app signal handler below will be overwritten in server.serve.
-    # server.started is set to True after server complete server.startup.
-    # We need to constraint the time sequence.
-    logger.info("API server: waiting for HTTP server to start")
-    while not server.started and not server_task.done():
-        await asyncio.sleep(0.1)
-    if server_task.done():
-        # Propagate startup failure (e.g. port bind error) instead of hanging.
-        await server_task
-    logger.info("API server: HTTP server started")
-
     shutdown_event = asyncio.Event()
 
     def signal_handler() -> None:

vllm/entrypoints/openai/dp_supervisor.py

@@ -23,6 +23,7 @@ import uvloop
 from fastapi import FastAPI, Response
 
 import vllm.envs as envs
+from vllm.entrypoints.launcher import NoSignalServer
 from vllm.logger import init_logger
 from vllm.utils.system_utils import (
     decorate_logs,
@@ -335,7 +336,7 @@ class DPSupervisor:
             ssl_cert_reqs=self.args.ssl_cert_reqs,
             ssl_ciphers=self.args.ssl_ciphers,
         )
-        supervisor_server = uvicorn.Server(config)
+        supervisor_server = NoSignalServer(config)
         supervisor_server_task = asyncio.create_task(
             supervisor_server.serve(),
             name="dp-supervisor",