xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-27 02:41:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Logging] Remove TORCH_NCCL_AVOID_RECORD_STREAMS to squash a warning (#25532)

Browse Source 
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Tyler Michael Smith 2025-09-25 11:16:06 -04:00 committed by GitHub
parent 0754ac4c49
commit 2e5df88c92
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 0 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/v1/worker/gpu_worker.py
View File

@ -155,14 +155,6 @@ class Worker(WorkerBase):
def init_device(self): def init_device(self):
if self.device_config.device.type == "cuda": if self.device_config.device.type == "cuda":
# torch.distributed.all_reduce does not free the input tensor until
# the synchronization point. This causes the memory usage to grow
# as the number of all_reduce calls increases. This env var disables
# this behavior.
# Related issue:
# https://discuss.pytorch.org/t/cuda-allocation-lifetime-for-inputs-to-distributed-all-reduce/191573
os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
# This env var set by Ray causes exceptions with graph building. # This env var set by Ray causes exceptions with graph building.
os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None) os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
self.device = torch.device(f"cuda:{self.local_rank}") self.device = torch.device(f"cuda:{self.local_rank}")