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 11, 22:30 UTC
[Misc] Move `disable_nccl_for_dp_synchronization` init logic into `VllmConfig` (#30161)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Dec 5, 2025 at 20:59 UTC
4026ae31e910d50da2b80c1c386f1d1db7f1b7d8
2 files changed
+9
-6
vllm/config/vllm.py
+9
@@ -586,6 +586,15 @@ class VllmConfig:
586
else:
587
self.scheduler_config.async_scheduling = True
588
589
+ if (
590
+ self.scheduler_config.async_scheduling
591
+ and not self.parallel_config.disable_nccl_for_dp_synchronization
592
+ ):
593
+ logger.info(
594
+ "Disabling NCCL for DP synchronization when using async scheduling."
595
+ )
596
+ self.parallel_config.disable_nccl_for_dp_synchronization = True
597
+
598
from vllm.platforms import current_platform
599
600
if (
vllm/engine/arg_utils.py
-6
@@ -1602,12 +1602,6 @@ class EngineArgs:
1602
model_config.skip_tokenizer_init = True
1603
logger.info("Skipping tokenizer initialization for tokens-only mode.")
1604
1605
- if self.async_scheduling and not self.disable_nccl_for_dp_synchronization:
1606
- logger.info(
1607
- "Disabling NCCL for DP synchronization when using async scheduling."
1608
- )
1609
- self.disable_nccl_for_dp_synchronization = True
1610
-
1605
parallel_config = ParallelConfig(
1606
pipeline_parallel_size=self.pipeline_parallel_size,
1607
tensor_parallel_size=self.tensor_parallel_size,