biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[distributed][misc] use fork by default for mp (#5669)

Browse Source
This commit is contained in:
youkaichao
2024-06-20 17:06:34 -07:00
committed by GitHub
parent 8065a7e220
commit 6c5b7af152
3 changed files with 38 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/envs.py
View File

@@ -29,7 +29,7 @@ if TYPE_CHECKING:
VLLM_CPU_KVCACHE_SPACE: int = 0
VLLM_XLA_CACHE_PATH: str = "~/.vllm/xla_cache/"
VLLM_USE_RAY_COMPILED_DAG: bool = False
VLLM_WORKER_MULTIPROC_METHOD: str = "spawn"
VLLM_WORKER_MULTIPROC_METHOD: str = "fork"
VLLM_IMAGE_FETCH_TIMEOUT: int = 5
VLLM_TARGET_DEVICE: str = "cuda"
MAX_JOBS: Optional[str] = None
@@ -212,7 +212,7 @@ environment_variables: Dict[str, Callable[[], Any]] = {
# Use dedicated multiprocess context for workers.
# Both spawn and fork work
"VLLM_WORKER_MULTIPROC_METHOD":
lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn"),
lambda: os.getenv("VLLM_WORKER_MULTIPROC_METHOD", "fork"),
# Timeout for fetching images when serving multimodal models
# Default is 5 seconds