vLLM PR 52555: Add opt-in custom all-reduce max-size override

Fields

id

sources/mgpu-vllm-52555-custom-ar-nvlink

URL

https://github.com/vllm-project/vllm/pull/52555

Publisher

vllm-project/vllm

Kind

repo

Published

2026-08-16

Accessed

2026-10-03

notes

Authoritative statement of the custom-all-reduce eligibility rule: at TP=2 custom all-reduce can operate over PCIe when functional GPU P2P exists and NVLink is not required; at TP>2 the constructor requires full one-hop NVLink and otherwise disables custom all-reduce entirely.

Notes

Authoritative statement of the custom-all-reduce eligibility rule: at TP=2 custom all-reduce can operate over PCIe when functional GPU P2P exists and NVLink is not required; at TP>2 the constructor requires full one-hop NVLink and otherwise disables custom all-reduce entirely.

Cross-references

This record references no other records.