Open Models Almanac
Methodsoptimizationtraining parallelism

single mention · optimization

Pipeline ZeRO-2 gradient sharding with CPU offloading

Pipeline ZeRO-2 shards gradients across data-parallel ranks and stores the sharded gradients in CPU memory.

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.

Filed under optimization › training parallelism