safemlx

safemlx provides safe, low-level Rust bindings to MLX: arrays and native operations, devices and streams, lazy graph evaluation, function transforms, native serialization, and accelerator/runtime facilities. It deliberately contains no neural-network layers, models, optimizers, checkpoint policy, or other framework abstractions. Eredu's MLX-specific abstractions live in eredu-backend-mlx.

Most applications should use eredu. Depend on this crate directly only when building MLX-specific operations or backend components.

Features

• accelerate: Accelerate-backed operations on Apple platforms.
• metal: Metal execution on Apple platforms.
• cuda: CUDA execution on x86-64 Linux or Windows.
• nccl: CUDA plus MLX's optional Linux NCCL backend.
• safetensors: conversion between Array and safetensors::TensorView.

The default feature set enables Accelerate and Metal where those backends are available.

Operations are lazy and run on an explicit stream:

use safemlx::{array, Device, DeviceType, Stream};

let stream = Stream::new_with_device(&Device::new(DeviceType::Cpu, 0));
let values = array!([1.0, 2.0, 3.0]);
let squares = values.square(&stream)?.into_evaluated()?;

assert_eq!(squares.as_slice::<f32>(), &[1.0, 4.0, 9.0]);
# Ok::<(), safemlx::error::Exception>(())

Rustdoc contains the API guide and examples for arrays, indexing, lazy evaluation, graph transforms, and I/O.

Notable APIs

• Completion events submit selected lazy graphs and support host observation or same-device stream ordering without a whole-stream drain.
• Typed host-transfer buffers provide explicit CPU, Metal shared, CUDA pinned, or CUDA managed storage policies.
• The distributed module wraps MLX groups, collectives, point-to-point operations, and their native execution semantics.

See the safemlx implementation guides for completion events, asynchronous device timing, and host-transfer buffers.

Platforms

Eredu's MLX implementation supports Apple silicon on macOS and selected Apple device targets, x86-64 Linux with CPU or CUDA, and native x86-64 Windows with CPU or CUDA. The compressed Metal library is embedded automatically. Backend prerequisites and platform packaging details are documented in the MLX backend's platform setup guide.

The minimum supported Rust version is 1.89.

License

Licensed under either MIT or Apache-2.0.