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 4, 07:30 UTC
[Ray] Enable RayExecutorV2 by default (#41421)
Signed-off-by: Jeffrey Wang <jeffreywang@anyscale.com>
Jeffrey Wang committed
May 4, 2026 at 21:27 UTC
f04fd1677b6cbaa03145e4b1e5c6696321aefdbe
1 file changed
+1
-2
vllm/envs.py
+1
-2
@@ -782,9 +782,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
782
),
783
# When True and distributed_executor_backend="ray", use RayExecutorV2
784
# (MQ-based) instead of RayDistributedExecutor (compiled-graph backend).
785
- # TODO (jeffreywang): Enabled by default in vLLM 0.20.0.
785
"VLLM_USE_RAY_V2_EXECUTOR_BACKEND": lambda: bool(
787
- int(os.getenv("VLLM_USE_RAY_V2_EXECUTOR_BACKEND", "0"))
786
+ int(os.getenv("VLLM_USE_RAY_V2_EXECUTOR_BACKEND", "1"))
787
),
788
# Use dedicated multiprocess context for workers.
789
# Both spawn and fork work