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

vLLM issue 58290: Multi-node TP=8 (2 nodes x 4 H200) hangs in ncclCommInitRank during startup

repo

Fields

id
sources/mgpu-vllm-58290-multinode-init-hang
URL
https://github.com/vllm-project/vllm/issues/58290
Publisher
vllm-project/vllm
Kind
repo
Published
2026-09-23
Accessed
2026-10-03
notes
Even local_rank ranks report Init COMPLETE while odd ranks block in the graph-connect phase; startup never returns.

Notes

Even local_rank ranks report Init COMPLETE while odd ranks block in the graph-connect phase; startup never returns.

Cross-references

This record references no other records.