diff --git a/tests/distributed/test_shm_broadcast.py b/tests/distributed/test_shm_broadcast.py index 7debed4cbb..d28732261a 100644 --- a/tests/distributed/test_shm_broadcast.py +++ b/tests/distributed/test_shm_broadcast.py @@ -1,11 +1,13 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project +import gc import io import pickle import random import threading import time +import weakref from types import SimpleNamespace from unittest import mock @@ -634,6 +636,26 @@ def test_acquire_read_releases_slot_when_reader_raises(): reader.shutdown() +def test_reader_in_reference_cycle_is_collected_without_blocking(): + writer = MessageQueue( + n_reader=1, + n_local_reader=1, + max_chunk_bytes=1024 * 1024, + max_chunks=1, + ) + reader = MessageQueue.create_from_handle(writer.export_handle(), rank=0) + reader.cycle = reader + reader_ref = weakref.ref(reader) + del reader + + collector = threading.Thread(target=gc.collect, daemon=True) + collector.start() + collector.join(timeout=10) + + assert not collector.is_alive() + assert reader_ref() is None + + def test_warning_logs(caplog_vllm): """Test that warning logs are emitted at VLLM_RINGBUFFER_WARNING_INTERVAL intervals when indefinite=False, and are not emitted when indefinite=True. diff --git a/vllm/distributed/device_communicators/shm_broadcast.py b/vllm/distributed/device_communicators/shm_broadcast.py index bafc5728b8..da2288f77b 100644 --- a/vllm/distributed/device_communicators/shm_broadcast.py +++ b/vllm/distributed/device_communicators/shm_broadcast.py @@ -9,6 +9,7 @@ import shutil import sys import threading import time +import weakref from contextlib import contextmanager from dataclasses import dataclass, field from multiprocessing import shared_memory @@ -472,6 +473,12 @@ class Handle: remote_addr_ipv6: bool = False +def _close_zmq(context: zmq.Context, sockets: list[zmq.Socket]) -> None: + for socket in sockets: + socket.close(linger=0) + context.term() + + class MessageQueue: def __init__( self, @@ -561,9 +568,26 @@ class MessageQueue: remote_subscribe_addr=remote_subscribe_addr, remote_addr_ipv6=remote_addr_ipv6, ) + self._close_zmq_on_collect(context) logger.debug("vLLM message queue communication handle: %s", self.handle) + def _close_zmq_on_collect(self, context: zmq.Context) -> None: + # pyzmq's Context.__del__ can block forever in term() when the context + # and its sockets are collected in the same GC cycle, so the finalizer + # keeps them alive and closes the sockets before terminating. + sockets = [self.local_socket, self.remote_socket] + if self._spin_condition is not None: + sockets += [ + self._spin_condition.local_notify_socket, + self._spin_condition.read_cancel_socket, + self._spin_condition.write_cancel_socket, + ] + finalizer = weakref.finalize( + self, _close_zmq, context, [s for s in sockets if s is not None] + ) + finalizer.atexit = False # type: ignore[misc] + def export_handle(self) -> Handle: return self.handle @@ -613,6 +637,7 @@ class MessageQueue: self._spin_condition = None # type: ignore self.shutting_down = False + self._close_zmq_on_collect(context) return self def wait_until_ready(self):