A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 01:00 UTC
[Misc] Copy HF_TOKEN env var to Ray workers (#21406)
Signed-off-by: Rui Qiao <ruisearch42@gmail.com>
Rui Qiao committed
Jul 22, 2025 at 16:18 UTC
35bc8bd5fb302dee4c2a7c7bc7e2e1f303478b09
2 files changed
+8
-3
vllm/executor/ray_distributed_executor.py
+5
-1
@@ -58,6 +58,9 @@ class RayDistributedExecutor(DistributedExecutorBase):
58
"VLLM_HOST_IP", "VLLM_HOST_PORT", "LOCAL_RANK", "CUDA_VISIBLE_DEVICES"
59
}
60
61
+ # These non-vLLM env vars are copied from the driver to workers
62
+ ADDITIONAL_ENV_VARS = {"HF_TOKEN", "HUGGING_FACE_HUB_TOKEN"}
63
+
64
uses_ray: bool = True
65
66
def _init_executor(self) -> None:
@@ -326,7 +329,8 @@ class RayDistributedExecutor(DistributedExecutorBase):
329
# Environment variables to copy from driver to workers
330
env_vars_to_copy = get_env_vars_to_copy(
331
exclude_vars=self.WORKER_SPECIFIC_ENV_VARS,
329
- additional_vars=set(current_platform.additional_env_vars),
332
+ additional_vars=set(current_platform.additional_env_vars).union(
333
+ self.ADDITIONAL_ENV_VARS),
334
destination="workers")
335
336
# Copy existing env vars to each worker's args
vllm/ray/ray_env.py
+3
-2
@@ -43,6 +43,8 @@ def get_env_vars_to_copy(exclude_vars: Optional[set[str]] = None,
43
exclude_vars: A set of vllm defined environment variables to exclude
44
from copying.
45
additional_vars: A set of additional environment variables to copy.
46
+ If a variable is in both exclude_vars and additional_vars, it will
47
+ be excluded.
48
destination: The destination of the environment variables.
49
Returns:
50
A set of environment variables to copy.
@@ -52,10 +54,9 @@ def get_env_vars_to_copy(exclude_vars: Optional[set[str]] = None,
54
55
env_vars_to_copy = {
56
v
55
- for v in envs.environment_variables
57
+ for v in set(envs.environment_variables).union(additional_vars)
58
if v not in exclude_vars and v not in RAY_NON_CARRY_OVER_ENV_VARS
59
}
58
- env_vars_to_copy.update(additional_vars)
60
61
to_destination = " to " + destination if destination is not None else ""
62