[Feat][Executor] Introduce RayExecutorV2 (#36836)
Signed-off-by: Jeffrey Wang <jeffreywang@anyscale.com>
This commit is contained in:
@@ -59,6 +59,7 @@ if TYPE_CHECKING:
|
||||
VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
|
||||
VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM: bool = False
|
||||
VLLM_USE_RAY_WRAPPED_PP_COMM: bool = True
|
||||
VLLM_USE_RAY_V2_EXECUTOR_BACKEND: bool = False
|
||||
VLLM_XLA_USE_SPMD: bool = False
|
||||
VLLM_WORKER_MULTIPROC_METHOD: Literal["fork", "spawn"] = "fork"
|
||||
VLLM_ASSETS_CACHE: str = os.path.join(VLLM_CACHE_ROOT, "assets")
|
||||
@@ -753,6 +754,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
|
||||
"VLLM_USE_RAY_WRAPPED_PP_COMM": lambda: bool(
|
||||
int(os.getenv("VLLM_USE_RAY_WRAPPED_PP_COMM", "1"))
|
||||
),
|
||||
# When True and distributed_executor_backend="ray", use RayExecutorV2
|
||||
# (MQ-based) instead of RayDistributedExecutor (compiled-graph backend).
|
||||
# TODO (jeffreywang): Enabled by default in vLLM 0.20.0.
|
||||
"VLLM_USE_RAY_V2_EXECUTOR_BACKEND": lambda: bool(
|
||||
int(os.getenv("VLLM_USE_RAY_V2_EXECUTOR_BACKEND", "0"))
|
||||
),
|
||||
# Use dedicated multiprocess context for workers.
|
||||
# Both spawn and fork work
|
||||
"VLLM_WORKER_MULTIPROC_METHOD": env_with_choices(
|
||||
|
||||
Reference in New Issue
Block a user