Skip to contents

Rmlx: R Interface to Apple's MLX Arrays

Rmlx-package.Rd

This package provides an R interface to Apple's MLX (Machine Learning eXchange) library for GPU-accelerated array operations on Apple Silicon.

Key Features

  • Lazy evaluation: Operations are not computed until explicitly evaluated

  • GPU acceleration: Leverage Metal on Apple Silicon

  • Familiar syntax: S3 methods for standard R operations

  • Unified memory: Efficient data sharing between CPU and GPU

Main Functions

  • as_mlx(): Convert R objects to MLX arrays

  • as.matrix.mlx(): Convert MLX arrays back to R

  • mlx_eval(): Force evaluation of lazy operations

  • Arithmetic: +, -, *, /, ^

  • Matrix ops: %*%, t, crossprod, tcrossprod

  • Reductions: sum, mean, colMeans, rowMeans

Lazy Evaluation

MLX arrays use lazy evaluation by default. Operations are recorded but not executed until:

  • You call mlx_eval(x)

  • You convert to R with as.matrix(x)

  • The result is needed for another computation

Device Management

Use mlx_default_device() to control whether arrays are created on GPU (default) or CPU. All mlx arrays are stored in float32 regardless of device. Use base R arrays if you require float64 math.

Known Limitations

Rmlx does not yet wrap every MLX primitive. Major gaps include:

  • Scatter-style updates (e.g. scatter_add, take_along_axis)

  • Bitwise integer kernels (bitwise AND/OR/XOR, shifts)

  • Advanced autograd transforms (jvp, vjp, vmap, custom VJPs)

  • N-dimensional FFT helpers (fftn, rfftn, and inverse variants)

  • Distributed collectives (NCCL/MPI backends)

See dev/mlx_coverage.txt for a living checklist of the MLX surface area.

See also

Useful links:

Author

Maintainer: David Hugh-Jones david@hughjones.com