single mention · optimization
Keep weights unsharded after backward across gradient-accumulation steps
Weights remain unsharded between backward and the next accumulation-step forward to avoid another all-gather, then are forcibly resharded before optimization.
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_reshard_after_backward().
Filed under optimization › training parallelism