# Gradient checkpointing

Stores only some activations and recomputes the rest during the backward pass.

## What it is used for

Training bigger models or longer sequences within memory limits.