Model Parallelism

Sharded Data Parallelism

A combination of data parallelism and the ZeRO strategy, where model weights are partitioned (sharded) among workers while maintaining data parallelism.

← Terug