Skip to content

v0.11.0

Latest

Choose a tag to compare

@esitaridi esitaridi released this 11 Nov 01:26
201eaf7

New features:

  • Add Compute Fabric Transport (CFT) buffers, which expose CUDA fabric logical endpoints to the copy kernels
  • Add CFT fabric kernels driven from TMA: try_put/get, counted put, and multimem put/pullred for unicast and multicast traffic on Blackwell (sm_100) and later
  • Localize copies to the locality domain under test on Rubin, in the intra device copy tests and in the device-to-device CE tests on both single-node and multinode runs
  • Add device_to_device_latency_counted and multinode_multicast_reduce tests
  • Print the fabric cluster and clique membership of every rank before the tests run, and warn when a node's fabric is unhealthy

Improvements:

  • Append the PCI bus, device and function ID to the device listing

Bug fixes:

  • Wait for every stream to finish warming up before timing the concurrent copy tests, to avoid under-reporting their aggregate bandwidth
  • Restore multi-GPU device listing on non-MPI runs