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, 06:00 UTC
[Bugfix] fix RAY_CGRAPH_get_timeout is not set successfully (#19725)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Jun 19, 2025 at 01:26 UTC
12575cfa7aa176e017735dd2883513b12be54c32
1 file changed
+10
-10
vllm/executor/ray_distributed_executor.py
+10
-10
@@ -557,8 +557,17 @@ class RayDistributedExecutor(DistributedExecutorBase):
557
def _compiled_ray_dag(self, enable_asyncio: bool):
558
assert self.parallel_config.use_ray
559
self._check_ray_cgraph_installation()
560
+ # Enlarge the default value of "RAY_CGRAPH_get_timeout" to 300 seconds
561
+ # (it is 10 seconds by default). This is a Ray environment variable to
562
+ # control the timeout of getting result from a compiled graph execution,
563
+ # i.e., the distributed execution that includes model forward runs and
564
+ # intermediate tensor communications, in the case of vllm.
565
+ # Note: we should set this env var before importing
566
+ # ray.dag, otherwise it will not take effect.
567
+ os.environ.setdefault("RAY_CGRAPH_get_timeout", "300") # noqa: SIM112
568
from ray.dag import InputNode, MultiOutputNode
561
-
569
+ logger.info("RAY_CGRAPH_get_timeout is set to %s",
570
+ os.environ["RAY_CGRAPH_get_timeout"]) # noqa: SIM112
571
logger.info("VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE = %s",
572
envs.VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE)
573
logger.info("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM = %s",
@@ -570,15 +579,6 @@ class RayDistributedExecutor(DistributedExecutorBase):
579
"Invalid value for VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: "
580
f"{channel_type}. Valid values are: 'auto', 'nccl', or 'shm'.")
581
573
- # Enlarge the default value of "RAY_CGRAPH_get_timeout" to 300 seconds
574
- # (it is 10 seconds by default). This is a Ray environment variable to
575
- # control the timeout of getting result from a compiled graph execution,
576
- # i.e., the distributed execution that includes model forward runs and
577
- # intermediate tensor communications, in the case of vllm.
578
- os.environ.setdefault("RAY_CGRAPH_get_timeout", "300") # noqa: SIM112
579
- logger.info("RAY_CGRAPH_get_timeout is set to %s",
580
- os.environ["RAY_CGRAPH_get_timeout"]) # noqa: SIM112
581
-
582
with InputNode() as input_data:
583
# Example DAG: PP=2, TP=4
584
#