RCCL PR 2187: Force Simple protocol for gfx12 single-node collectives to fix deadlock

Fields

id

sources/mgpu-rccl-2187-gfx12-ll-deadlock

URL

https://github.com/ROCm/rccl/pull/2187

Publisher

ROCm/rccl

Kind

repo

Published

2026-08-20

Accessed

2026-10-03

notes

Names the regression commit and the unfixed primitive: prims_ll.h never received the gfx12 memory-ordering fixes that prims_simple.h got, so selecting LL on gfx1201 deadlocks. Fixes a vLLM TP=2 inference deadlock on dual R9700.

Notes

Names the regression commit and the unfixed primitive: prims_ll.h never received the gfx12 memory-ordering fixes that prims_simple.h got, so selecting LL on gfx1201 deadlocks. Fixes a vLLM TP=2 inference deadlock on dual R9700.

Cross-references

This record references no other records.