Gradient Checkpointing

AI and Machine Learning · Efficiency and Deployment · 2016 · gradient-checkpointing.yaml

Keeps only a subset of activations and recomputes the rest during the backward pass, trading arithmetic for memory at roughly the square root of the naive cost.

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 backpropagation Backpropagation gradient-checkpointing Gradient Checkpointing gradient-checkpointing->backpropagation keeping every activation exhausts memory zero-sharding ZeRO Sharding zero-sharding->gradient-checkpointing shards state across devices instead of recomputing it

This node

Referenced by

References