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 12, 14:45 UTC
[Perf] use cpu all reduce to avoid sync when async_scheduling & dp > 1 (#29311)
Signed-off-by: zhuhaoran <zhuhaoran.zhr@alibaba-inc.com>
zhrrr committed
Nov 25, 2025 at 15:31 UTC
f242cfcdd5f1db4e005503a02a1317369d2a8e3d
1 file changed
+6
vllm/engine/arg_utils.py
+6
@@ -1570,6 +1570,12 @@ class EngineArgs:
1570
model_config.skip_tokenizer_init = True
1571
logger.info("Skipping tokenizer initialization for tokens-only mode.")
1572
1573
+ if self.async_scheduling and not self.disable_nccl_for_dp_synchronization:
1574
+ logger.info(
1575
+ "Disabling NCCL for DP synchronization when using async scheduling."
1576
+ )
1577
+ self.disable_nccl_for_dp_synchronization = True
1578
+
1579
# Forward the deprecated CLI args to the EPLB config.
1580
if self.num_redundant_experts is not None:
1581
self.eplb_config.num_redundant_experts = self.num_redundant_experts