mlx-sparse

Warning

This is an early beta. macOS and Linux are supported, but Linux is CPU-only in this release. APIs may change, bugs are expected, and some features are still incomplete, especially GPU kernels for various operations. Feedback and issue reports are very welcome!

A lot of the functionality is new and is currently being tested. We welcome any and all feedback! Not all solvers are GPU supported (see the solver support matrix).

Note

GPU support in this version is Apple Silicon Metal only. CUDA is not currently supported. Linux wheels use native CPU kernels only, there is no CUDA, ROCm, Accelerate, BLAS, or Sparse BLAS backend on Linux yet.

mlx-sparse is an attempt at an MLX-native sparse array package. The public API is Python, while performance-critical operations are implemented as MLX primitives in C++ with CPU backends and Metal kernels for fixed-shape sparse operations.

The supported format surface is COO, CSR, and CSC for 2D sparse arrays. Current functionality includes construction, validation, format conversions, canonicalization, sparse-dense products, batched dense RHS products, sparse-sparse products, transpose, and Hermitian transpose. mlx-sparse also supports sparse linalg solvers (cg, bicgstab, gmres, minres), reusable direct solves (factorized), sparse spectral routines (eigsh, eigs, svds), sparse Cholesky/LU factors, sparse triangular solves, sparse dot/vdot, and autodiff through sparse values and dense RHS operands, including complex64.

Supported value dtypes are float32, float16, bfloat16, and complex64. Supported index dtypes are int32 and int64 on CPU and GPU.

Quick Start

Install from PyPI:

python -m pip install mlx-sparse

Published macOS wheels are built with optional Apple Accelerate sparse direct solver support enabled. Published Linux wheels are CPU-only. For editable/source macOS builds, enable the same Accelerate path with:

CMAKE_ARGS="-DMLX_SPARSE_ENABLE_ACCELERATE=ON" python -m pip install -e .

See the full installation guide for platform requirements, feature gates, and verification commands.

import mlx.core as mx
import numpy as np

import mlx_sparse as ms

ms.use_cpu()

data = [2.0, -1.0, 4.0]
row = [0, 0, 1]
col = [0, 2, 1]

a = ms.coo_array((data, (row, col)), shape=(2, 3)).tocsr(canonical=True)
x = mx.array(np.array([3.0, 10.0, 7.0], dtype=np.float32))

y = a @ x
dense = a.todense()
at = a.T

b = mx.array(np.array([1.0, 2.0], dtype=np.float32))
spd = ms.csr_array(
    (
        [4.0, 1.0, 1.0, 3.0],
        [0, 1, 0, 1],
        [0, 2, 4],
    ),
    shape=(2, 2),
    canonical=True,
)
solution, info = ms.linalg.cg(spd, b)
factor = ms.linalg.sparse_cholesky(spd)

The package build compiles src/sparse/*.metal into mlx_sparse/mlx_sparse.metallib when the macOS Metal toolchain is available, and macOS wheels ship that metallib beside the Python package. Linux wheels do not build or ship Metal assets.

Development

For contributors, use an editable install from the repository root. This builds the native extension and installs the development tooling.

python -m pip install -e ".[dev]"

License

This package is licensed under the Apache License 2.0.