single mention · optimization
ZeRO-2
ZeRO-2 shards optimizer-related state and gradients; the evidence does not provide further details of the base method.
The sources on this page name it without explaining how it works. The line above is the extraction's one-line gloss. See training parallelism for the methods filed beside it.
Also called ZeRO-2 optimizations.
Filed under optimization › training parallelism