[Bugfix] Fix failure to launch in Tensor Parallel TP mode on macOS. (#14948)

Signed-off-by: Kebe <mail@kebe7jun.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
Kebe
2025-03-28 10:13:41 +08:00
committed by GitHub
co-authored by youkaichao
parent 726efc6a32
commit 4e0f6076be
3 changed files with 17 additions and 4 deletions
@@ -125,8 +125,13 @@ class ShmRingBuffer:
lambda *args, **kwargs: None):
try:
self.shared_memory = shared_memory.SharedMemory(name=name)
assert (
self.shared_memory.size == self.total_bytes_of_buffer)
# See https://docs.python.org/3/library/multiprocessing.shared_memory.html # noqa
# Some platforms allocate memory based on page size,
# so the shared memory block size may be larger or equal
# to the requested size. The size parameter is ignored
# when attaching to an existing block.
assert (self.shared_memory.size
>= self.total_bytes_of_buffer)
except FileNotFoundError:
# we might deserialize the object in a different node
# in this case, this object is not used,
+8
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
import os
import sys
from typing import TYPE_CHECKING, Optional
import psutil
@@ -148,6 +149,13 @@ class CpuPlatform(Platform):
# To hint IPEX uses shared memory based AllReduce
os.environ["LOCAL_WORLD_SIZE"] = str(
vllm_config.parallel_config.tensor_parallel_size)
if sys.platform == "darwin" and \
envs.VLLM_WORKER_MULTIPROC_METHOD == "fork":
if os.environ.get('VLLM_WORKER_MULTIPROC_METHOD', None) is None:
logger.warning(
"Default to spawn method on MacOS. If this is not desired,"
" set VLLM_WORKER_MULTIPROC_METHOD to fork explicitly.")
os.environ['VLLM_WORKER_MULTIPROC_METHOD'] = 'spawn'
@classmethod
def is_pin_memory_available(cls) -> bool: