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 7, 14:00 UTC
[EPLB] Enforce sync eplb for NCCL-based all2all backend (#35212)
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Ilya Markov committed
Feb 28, 2026 at 06:47 UTC
b2d8b422b2014b23c44fea703c70331eef35e7a1
1 file changed
+11
vllm/config/parallel.py
+11
@@ -774,6 +774,17 @@ class ParallelConfig:
774
"backend is mp, uni or external_launcher."
775
)
776
777
+ if (
778
+ self.all2all_backend in ("allgather_reducescatter", "naive")
779
+ and self.eplb_config.use_async
780
+ ):
781
+ logger.warning(
782
+ "Async EPLB causes hangs with the '%s' all2all backend. "
783
+ "Forcing synchronous EPLB.",
784
+ self.all2all_backend,
785
+ )
786
+ self.eplb_config.use_async = False
787
+
788
@property
789
def use_ray(self) -> bool:
790
return self.distributed_executor_backend == "ray" or (