inference/research
verifiedhigh · 0.85updated 2026-10-03

NCCL issue 582: OOM during init rank causes deadlock

repo

Fields

id
sources/mgpu-nccl-582-init-oom-deadlock
URL
https://github.com/NVIDIA/nccl/issues/582
Publisher
NVIDIA/nccl
Kind
repo
Published
2021-10-15
Accessed
2026-10-03
notes
Points at group.cc: when one init thread errors (e.g. OOM), the remaining threads block in bootstrapAllGather on a socket with no timeout, converting a local error into a hang.

Notes

Points at group.cc: when one init thread errors (e.g. OOM), the remaining threads block in bootstrapAllGather on a socket with no timeout, converting a local error into a hang.

Cross-references

This record references no other records.