Remat

Applies gradient checkpointing (rematerialization) to a function to save memory at the cost of compute.

Abstract Signature:

Remat(f: Callable)

PyTorchJAX (Core)Flax NNX

PyTorch

API: torch.utils.checkpoint.checkpoint

Strategy: Plugin (checkpoint_wrapper)

JAX (Core)

API: jax.checkpoint

Strategy: Direct Mapping

Official Docs ↗

Flax NNX

API: flax.nnx.remat

Strategy: Direct Mapping

Official Docs ↗