Skip to content

perf: eliminate end_sync in custom allreduce by delaying input tensor release until next sync#4084

Open
jpy794 wants to merge 2 commits into
ROCm:mainfrom
RadeonFlow:custom-ar-hold-input
Open

perf: eliminate end_sync in custom allreduce by delaying input tensor release until next sync#4084
jpy794 wants to merge 2 commits into
ROCm:mainfrom
RadeonFlow:custom-ar-hold-input

Commits

Commits on Jul 5, 2026