Binyang Li 4db71b93b7 Move barrier into setupNvlsChannels and clean up NVLS pipeline state ...

- setupNvlsChannels now takes the Communicator and barriers internally
  after binding all switch channels, replacing the explicit
  bootstrap()->barrier() previously done only in AllreduceNvlsPacket.
- Demote nRanksPerIpcDomain_ / nBaseChannels_ to locals in
  AllreduceNvlsBlockPipeline and AllreduceNvlsWarpPipeline; they were
  never read outside initialize().
- Drive-by: pick up in-tree edits to switch_channel_device.hpp,
  executor.cc, communicator.hpp, and allreduce_rsag.cu.

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

2026-05-18 20:50:01 +00:00

..

mscclpp

Move barrier into setupNvlsChannels and clean up NVLS pipeline state

2026-05-18 20:50:01 +00:00

CMakeLists.txt

Torch integration (#692)

2026-01-21 20:32:24 -08:00