---
id: orbax-checkpoint
version: "0.12.4"
license: unclear
license_treatment: permissive
maintenance: active
---

# orbax-checkpoint — Orbax Checkpoint

License: permissive · Maintenance: active · Downloads: 5.7M/mo

## What it is and what it does

Orbax Checkpoint is a checkpointing library built for JAX machine learning workflows, designed to handle the state-saving needs of training pipelines. It wraps JAX's pytree structures and provides asynchronous I/O capabilities to avoid blocking training loops while persisting model weights and optimizer state to disk or cloud storage. The library sits on top of tensorstore for flexible storage backends and uses aiofiles and uvloop to manage non-blocking I/O operations.

The package targets researchers and practitioners who need reliable, efficient checkpointing during long-running JAX training jobs. It abstracts away the complexity of coordinating I/O with training steps, supporting various serialization formats and storage systems. With multiple runtime dependencies including protobuf, msgpack, and pyyaml, it provides a complete ecosystem for checkpoint management rather than a minimal utility.

Use it for:

- Save JAX model state asynchronously during training without pausing gradient computation.
- Restore model weights and optimizer state from checkpoints to resume interrupted training runs.
- Manage multiple checkpoint versions and storage backends through a unified API.
- Coordinate checkpointing across distributed JAX training setups with tensorstore integration.
- Serialize complex nested pytree structures with custom types using pluggable handlers.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Orbax Checkpoint provides asynchronous checkpointing for JAX machine learning workflows, supporting multiple storage formats and customizable serialization to save and restore model state during training.

Yes. Orbax Checkpoint is actively maintained (release 2 days old), has no known vulnerabilities, and solves a real problem for JAX practitioners: efficient, non-blocking checkpoint management. The permissive license and low install friction make it a straightforward choice. Install it if you're running JAX training jobs that need reliable state persistence.

## Install

pip install orbax-checkpoint
uv add orbax-checkpoint
poetry add orbax-checkpoint

## Installing orbax-checkpoint

Before you install: Low friction install with a pure-Python wheel. Active maintenance with a release 2 days old and recent commits. Depends on JAX and many other runtime packages including tensorstore, aiofiles, and uvloop for async I/O support.

License in practice: Permissive Apache license allows free use, modification, and distribution in both open and closed projects with minimal restrictions.

Quickstart:

pip install orbax-checkpoint

import orbax.checkpoint

# Basic checkpoint save/restore with JAX arrays
checkpointer = orbax.checkpoint.PyTreeCheckpointer()
checkpointer.save(path, pytree_state)

Requires Python 3.11 or later and JAX installed; asynchronous features depend on uvloop and aiofiles being available.

Verify before relying:

- Whether the package's 'highly customizable and composable API' covers specific storage backends beyond what tensorstore provides.
- Performance characteristics of asynchronous checkpointing relative to synchronous alternatives in typical training loops.
- Compatibility guarantees with different JAX versions and distributed training frameworks.

## Package facts

- License: not declared (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 5.7M/month (top 5,000 on PyPI)
- Known vulnerabilities: none known

## Tags

jax checkpointing library, async model checkpoints, jax training state save, machine learning checkpoint management, jax model persistence, distributed training checkpoints, asynchronous model saving, jax, async-io, model-persistence

[View on SkillFed](https://skillfed.io/packages/orbax-checkpoint) · [View on PyPI](https://pypi.org/project/orbax-checkpoint/)