DeepSpeed
88b4d506 - fix(comm): bounded handshake gates the custom AR — PCIe P2P hangs fall back

Commit
5 days ago
fix(comm): bounded handshake gates the custom AR — PCIe P2P hangs fall back The flag protocol's spin hangs on PCIe P2P pairs (A40: watchdog kill at 480s) where NVLink honors it; publish-and-wait probe with a bounded spin disables the fast path gracefully instead. Signed-off-by: Guokai Ma <guokai.ma@intel.com>
Author
Parents
Loading