single mention · optimization
Sequence parallelism for activations
The TP all-reduce is decomposed into reduce-scatter and all-gather, with the intra-block kernel placed between the collectives to materialize activations.
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 sequence parallelism (SP) for activations.
Filed under optimization › training parallelism