[Bugfix] Set breakable graph env before Ray actor import (#53293)
Signed-off-by: Ace Eldeib <aeldeib@coreweave.com> Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
@@ -22,6 +22,9 @@ from vllm.utils.network_utils import (
|
||||
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
|
||||
from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
|
||||
from vllm.v1.executor.abstract import Executor
|
||||
from vllm.v1.executor.ray_env_utils import (
|
||||
update_runtime_env_for_worker_import,
|
||||
)
|
||||
from vllm.v1.executor.ray_utils import (
|
||||
WORKER_SPECIFIC_ENV_VARS,
|
||||
FutureWrapper,
|
||||
@@ -147,6 +150,7 @@ class RayDistributedExecutor(Executor):
|
||||
env_vars.update(
|
||||
{env_var: "1" for env_var in current_platform.ray_noset_device_env_vars}
|
||||
)
|
||||
update_runtime_env_for_worker_import(runtime_env)
|
||||
return ray_remote_kwargs
|
||||
|
||||
# child class could overwrite this to return actual env vars.
|
||||
|
||||
Reference in New Issue
Block a user