Open Models Almanac
Methodsoptimizationtraining parallelism

single mention · optimization

Dedicated reduce-scatter process group

A separate NCCL process group lets reduce-scatter communication overlap with all-gathers instead of serializing behind them.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See training parallelism for the methods filed beside it.

Also called set_separate_reduce_scatter_group().

Filed under optimization › training parallelism