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 14, 05:30 UTC
[Logging] Remove TORCH_NCCL_AVOID_RECORD_STREAMS to squash a warning (#25532)
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Tyler Michael Smith committed
Sep 25, 2025 at 11:16 UTC
2e5df88c92fd066d05b90a37a6ff5f802f7be814
1 file changed
-8
vllm/v1/worker/gpu_worker.py
-8
@@ -155,14 +155,6 @@ class Worker(WorkerBase):
155
156
def init_device(self):
157
if self.device_config.device.type == "cuda":
158
- # torch.distributed.all_reduce does not free the input tensor until
159
- # the synchronization point. This causes the memory usage to grow
160
- # as the number of all_reduce calls increases. This env var disables
161
- # this behavior.
162
- # Related issue:
163
- # https://discuss.pytorch.org/t/cuda-allocation-lifetime-for-inputs-to-distributed-all-reduce/191573
164
- os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
165
-
158
# This env var set by Ray causes exceptions with graph building.
159
os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
160
self.device = torch.device(f"cuda:{self.local_rank}")