jax-graft: run JAX on your Mac's GPU
An open-source JAX plugin that runs JAX programs on Apple Silicon GPUs through Metal. It's for people who write JAX and want their Mac's GPU to do the work: training small models, running LLM inference, or scientific code that's happy in float32.
Under the hood it treats Metal as one more XLA:GPU platform, next to CUDA and ROCm. XLA's own GPU compiler fuses your program and generates the kernels, which the plugin translates to Metal Shading Language.
The JAX platform name is "mtl" (Metal's own prefix, as in MTLDevice),
so it doesn't collide with Apple's closed-source jax-metal plugin, which
uses "metal". This project isn't affiliated with or endorsed by Apple or
Google.
Status
Early, but it works. Float32, float16 and bfloat16 programs run end to
end on one GPU: training loops, sorting, FFTs, host callbacks and float32
linear algebra. Performance is broadly comparable to MLX and PyTorch's MPS
backend on the workloads we've tried, with some slow spots (see
docs/performance.md). Everything so far has been tested on
a single machine (an M3 with 8 GB, macOS 26.2), so expect rough edges
elsewhere, and please report them.
Requirements
- An Apple Silicon Mac. Intel Macs aren't supported.
- macOS 26 or later.
- Python 3.12+ with
jaxandjaxlib0.10.0 or later (tested: 0.10.0 through 0.11.2 and a 0.12 nightly). Older versions load with a warning.
Install
pip install jax-graft
Quick start
import jax
import jax.numpy as jnp
x = jnp.arange(4.0)
y = jax.jit(lambda v: v * 2)(x)
print(y, y.devices()) # [0. 2. 4. 6.] {MtlDevice(id=0)}
JAX prints a warning that the "mtl" platform is experimental; that's
expected. Like JAX's other GPU plugins, once installed the plugin is JAX's
default backend on a Mac with a Metal GPU, and programs it can't run
(float64, for example) raise an error rather than falling back to the CPU.
To keep CPU as the default, run with JAX_PLATFORMS=cpu; to keep it
available too, use JAX_PLATFORMS=cpu,mtl and place work on the GPU
explicitly with jax.device_put(x, jax.devices("mtl")[0]).
Before running anything heavy, read Is it safe for my
GPU?: a kernel that runs too long
makes macOS reset the GPU for the whole machine. The
FAQ also covers what works, accuracy, memory and the
compilation cache, and docs/troubleshooting.md
maps the plugin's error messages to what to do.
Examples
Each example has a README with results and what it took to make it fast. Run their commands from the root of this repository.
examples/llm: Qwen3 inference in pure JAX, in bf16, int8 and int4.examples/lora: LoRA fine-tuning of Qwen3.examples/cifar: training a CIFAR-10 classifier to 94% accuracy (airbench94).
Learn more
docs/faq.md: what works, GPU safety, accuracy, memory and the compilation cache.docs/design.md: how the plugin works, and how it differs from MLX.docs/op-coverage.md,docs/accuracy.md,docs/callbacks.md: what runs, how accurately, and host callbacks.docs/performance.md: measurements and how they were taken.docs/integration-notes.md: the details of how the plugin plugs into XLA.docs/development.md: building, testing and environment variables.docs/roadmap.md: what's next, andCHANGELOG.md: what changed.
Contributing
Issues and pull requests are welcome. CONTRIBUTING.md
says what makes a useful report, and
docs/development.md covers building and testing.
Please run GPU tests under scripts/device_lock.py. For security issues,
see SECURITY.md.
License
Apache-2.0 (LICENSE). The plugin includes third-party code:
matmul, convolution and FFT kernels ported from
MLX (MIT), Apple's metal-cpp headers
(Apache-2.0), and a statically linked XLA. Their notices are in
THIRD_PARTY_NOTICES, which ships in both wheels.
Metadata
Release files for jax-graft 0.0.4
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| jax_graft-0.0.4-py3-none-any.whl | Python 3 | none | any | Details for jax_graft-0.0.4-py3-none-any.whl |
Release files / jax_graft-0.0.4-py3-none-any.whl
| Download URL | jax_graft-0.0.4-py3-none-any.whl |
|---|---|
| Size | 59.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
dbe5cf768675296fb5a47ea0b1827a4294d6c10902d2e1352130f999c9190df2
|
|
BLAKE2b-256 checksum How to use checksums |
f6bdbdc270b254fa765775849b406eb6af036c57427c3850db3720ba5ede8999
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 10, 2026.
Transparency logSecurity for jax-graft 0.0.4
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
1 file in this release has been uploaded with provenance metadata.
Attestation for jax_graft-0.0.4-py3-none-any.whl
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 10, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.