Binyang Li 3bc00cb7f0 Enable NVLS zero-copy without symmetric memory flag ...

Allow default_allreduce_nvls_zero_copy to run when the public symmetric_memory flag is false; the algorithm already binds the concrete input and output allocations in its context. Include that fast path in MNNVL tuning and bound allpair/NVLS packet candidates to small sizes so large-message no-symmetric tuning avoids slow or unsafe packet variants.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

2026-04-28 08:24:49 +00:00

..

customized-collective-algorithm

Support E4M3B15 datatype (#765)

2026-04-07 13:37:02 -07:00

torch-integration

Enable NVLS zero-copy without symmetric memory flag

2026-04-28 08:24:49 +00:00

tutorials

Mahdieh/switchchannel test clean (#751)

2026-02-20 22:46:32 -05:00