ZeRO Sharding

AI and Machine Learning · Distributed Training · 2019 · also: ZeRO · zero-sharding.yaml

Partitions optimizer state, gradients and parameters across data- parallel ranks instead of replicating them, gathering each shard only while it is needed.

Colour is the family; a dashed line is the second member of it.

Drag to pan · scroll to zoom · click a node to open it
G data-parallelism Data Parallelism fully-sharded-data-parallel Fully Sharded Data Parallel zero-sharding ZeRO Sharding fully-sharded-data-parallel->zero-sharding shards parameters as well as optimizer state, per layer gradient-checkpointing Gradient Checkpointing zero-sharding->data-parallelism every replica redundantly stores the same optimizer state zero-sharding->gradient-checkpointing shards state across devices instead of recomputing it

This node

Referenced by

References