biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Perf] use cpu all reduce to avoid sync when async_scheduling & dp > 1 (#29311)

Browse Source 
Signed-off-by: zhuhaoran <zhuhaoran.zhr@alibaba-inc.com>
This commit is contained in:
zhrrr
2025-11-25 15:31:07 +08:00
committed by GitHub
parent 888152bf87
commit f242cfcdd5
1 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/engine/arg_utils.py
View File

@@ -1570,6 +1570,12 @@ class EngineArgs:
model_config.skip_tokenizer_init = True model_config.skip_tokenizer_init = True
logger.info("Skipping tokenizer initialization for tokens-only mode.") logger.info("Skipping tokenizer initialization for tokens-only mode.")
if self.async_scheduling and not self.disable_nccl_for_dp_synchronization:
logger.info(
"Disabling NCCL for DP synchronization when using async scheduling."
)
self.disable_nccl_for_dp_synchronization = True
# Forward the deprecated CLI args to the EPLB config. # Forward the deprecated CLI args to the EPLB config.
if self.num_redundant_experts is not None: if self.num_redundant_experts is not None:
self.eplb_config.num_redundant_experts = self.num_redundant_experts self.eplb_config.num_redundant_experts = self.num_redundant_experts