onnxruntime-mlx 0.1.0

MLX-native ONNX Runtime execution provider (plugin EP) for Apple Silicon

onnxruntime-mlx (Python package)

Pip-installable MLX-native execution provider for ONNX Runtime on Apple Silicon. The wheel bundles the cargo-built libonnxruntime_mlx_ep.dylib plugin EP (plus its mlx-c / mlx dynamic dependencies and mlx.metallib) and a thin pure-Python locator that finds it, so Python users can register the EP with a stock onnxruntime wheel — no ONNX Runtime fork or manual dylib path required.

• Import name: onnxruntime_mlx
• Distribution name: onnxruntime-mlx
• Platform: macOS 14+ on Apple Silicon (arm64) only
• Requires: onnxruntime >= 1.22 (plugin-EP C API, ORT_API_VERSION >= 27)

Install

The wheel is built from a full repository checkout (the build hook runs cargo build --release in the sibling rust/ crate). Install a Rust toolchain (rustup) and brew install mlx-c first, then:

# from the repo root; point at the ORT C-API headers (or set ORT_HOME):
ORT_INCLUDE_DIR=/path/to/onnxruntime/include python -m build --wheel ./python
# or install directly:
ORT_INCLUDE_DIR=/path/to/onnxruntime/include pip install ./python

Build backend: hatchling with a custom build hook (hatch_build.py). There is no compiled Python extension — the EP is a Rust cdylib, and the Python layer is a pure-Python locator. The hook:

1. runs cargo build --release in ../rust (honouring ORT_INCLUDE_DIR, else $ORT_HOME/include), then
2. bundles the resulting dylib + the mlx-c/mlx runtime into the package (relinked to @loader_path), and
3. forces a single py3-none-macosx_*_arm64 platform wheel.

Because the wheel ships zero CPython-ABI code, one wheel installs on CPython 3.12, 3.13 and the free-threaded (3.13t/3.14t) builds — and abi3audit --strict is clean by construction (nothing auditable).

Usage

import onnxruntime as ort
import onnxruntime_mlx

onnxruntime_mlx.register_execution_provider_library()   # once per process
sess = ort.InferenceSession(
    "model.onnx",
    providers=["MLXExecutionProvider", "CPUExecutionProvider"],
)

Public API (onnxruntime_mlx)

Function	Returns
ep_name()	"MLXExecutionProvider"
version()	the installed package version string
vendor()	"onnxruntime-mlx"
library_path()	absolute path to the bundled libonnxruntime_mlx_ep.dylib

Plus register_execution_provider_library() and append_to_session_options().

How the dylib + mlx deps are bundled

The build hook (hatch_build.py):

1. copies libmlxc.dylib, libmlx.dylib, and mlx.metallib next to the plugin inside the package;
2. rewrites the plugin's mlx dependencies to @loader_path/lib{mlxc,mlx}.dylib and the bundled install ids to match, then re-signs (ad-hoc) each mutated binary.

The Rust EP does not link libonnxruntime (it reaches ORT purely through the OrtApi function-pointer table), so there is no onnxruntime dependency to rewrite; onnxruntime is never vendored and must match the host at runtime.