mirror of
https://github.com/microsoft/mscclpp.git
synced 2026-05-11 17:00:22 +00:00
Disable NVLS zero-copy when symmetric memory is not enabled, and allow the RSAG zero-copy path to participate in MNNVL tuning for non-symmetric memory. Cache RSAG zero-copy contexts by the concrete buffer pointers so CUDA graph capture does not create a new registration for every execute call, and cap requested blocks at the channel count. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>