RCCL 2.26.6 for ROCm 7.0.2
Added
- Enabled double-buffering in
reduceCopyPacksto trigger pipelining, especially to overlap bf16 arithmetic. - Added
--force-reduce-pipelineas an option that can be passed to theinstall.shscript. Passing this option will enable software-triggered pipeliningbfloat16reductions (i.e.all_reduce,reduce_scatterandreduce).