Skip to content

RCCL 2.26.6 for ROCm 7.0.2

Choose a tag to compare

@rocm-ci rocm-ci released this 10 Oct 12:09
01dfdac

Added

  • Enabled double-buffering in reduceCopyPacks to trigger pipelining, especially to overlap bf16 arithmetic.
  • Added --force-reduce-pipeline as an option that can be passed to the install.sh script. Passing this option will enable software-triggered pipelining bfloat16 reductions (i.e. all_reduce, reduce_scatter and reduce).