fix log spew on drain exit; fix Ctrl-C killing workers immediately when TP>1 by isolating process with setpgrp()

Signed-off-by: Will Eaton <[email protected]>

Author: wseaton

vllm/entrypoints/serve/middleware.py

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-from collections.abc import Awaitable
+import asyncio
 
 from fastapi.responses import JSONResponse
 from starlette.types import ASGIApp, Receive, Scope, Send
@@ -23,26 +23,34 @@ def set_rejecting_requests(value: bool) -> None:
 
 
 class ServiceUnavailableMiddleware:
-    """
-    Middleware that checks if the server is currently unavailable
-    (e.g., scaling or draining) and returns a 503 Service Unavailable.
-
-    """
+    """Returns 503 during drain; suppresses CancelledError on force-exit."""
 
     def __init__(self, app: ASGIApp) -> None:
         self.app = app
 
-    def __call__(self, scope: Scope, receive: Receive, send: Send) -> Awaitable[None]:
+    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
         if scope["type"] != "http":
-            return self.app(scope, receive, send)
+            await self.app(scope, receive, send)
+            return
 
-        path = scope.get("path", "")
-        rejecting = is_rejecting_requests()
-        if rejecting and path not in _EXEMPT_PATHS:
+        if is_rejecting_requests() and scope.get("path", "") not in _EXEMPT_PATHS:
             response = JSONResponse(
                 content={"error": "Server is unavailable. Please try again later."},
                 status_code=503,
             )
-            return response(scope, receive, send)
-
-        return self.app(scope, receive, send)
+            await response(scope, receive, send)
+            return
+
+        try:
+            await self.app(scope, receive, send)
+        except asyncio.CancelledError:
+            if not is_rejecting_requests():
+                raise
+            try:
+                response = JSONResponse(
+                    content={"error": "Server is shutting down."},
+                    status_code=503,
+                )
+                await response(scope, receive, send)
+            except (Exception, asyncio.CancelledError):
+                pass

vllm/v1/engine/core.py

@@ -991,9 +991,12 @@ def signal_handler(*_):
                         (EngineCoreRequestType.SHUTDOWN, None)
                     )
 
-        # Either SIGTERM or SIGINT will terminate the engine_core
         signal.signal(signal.SIGTERM, signal_handler)
-        signal.signal(signal.SIGINT, signal_handler)
+        if shutdown_pipe is not None:
+            # Isolate from terminal Ctrl-C; parent uses shutdown pipe.
+            os.setpgrp()
+        else:
+            signal.signal(signal.SIGINT, signal_handler)
         try:
             vllm_config: VllmConfig = kwargs["vllm_config"]
             parallel_config: ParallelConfig = vllm_config.parallel_config