[Bugfix] Fix HPU multiprocessing executor (#12167)
Signed-off-by: Konrad Zawora <kzawora@habana.ai>
This commit is contained in:
@@ -1,7 +1,9 @@
|
||||
import os
|
||||
from typing import TYPE_CHECKING, Optional
|
||||
|
||||
import torch
|
||||
|
||||
from vllm import envs
|
||||
from vllm.logger import init_logger
|
||||
|
||||
from .interface import Platform, PlatformEnum, _Backend
|
||||
@@ -58,6 +60,22 @@ class HpuPlatform(Platform):
|
||||
cache_config = vllm_config.cache_config
|
||||
if cache_config and cache_config.block_size is None:
|
||||
cache_config.block_size = 128
|
||||
if (parallel_config.distributed_executor_backend == 'mp'
|
||||
and envs.VLLM_WORKER_MULTIPROC_METHOD == 'fork'):
|
||||
if os.environ.get("VLLM_WORKER_MULTIPROC_METHOD",
|
||||
None) is not None:
|
||||
logger.warning("On HPU, VLLM_WORKER_MULTIPROC_METHOD=fork "
|
||||
"might cause application hangs on exit. Using "
|
||||
"VLLM_WORKER_MULTIPROC_METHOD=fork anyway, "
|
||||
"as it was explicitly requested.")
|
||||
else:
|
||||
logger.warning(
|
||||
"On HPU, VLLM_WORKER_MULTIPROC_METHOD=fork "
|
||||
"might cause application hangs on exit. Setting "
|
||||
"VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. "
|
||||
"To override that behavior, please set "
|
||||
"VLLM_WORKER_MULTIPROC_METHOD=fork explicitly.")
|
||||
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
|
||||
|
||||
@classmethod
|
||||
def is_pin_memory_available(cls):
|
||||
|
||||
Reference in New Issue
Block a user