metal-linalg for Python

QR, symmetric eigendecomposition and SVD for batches of matrices on Apple GPUs, on MLX arrays. For PyTorch tensors, install metal-linalg-torch instead.

import mlx.core as mx
import metal_linalg as ml

a = mx.random.normal((1000, 64, 32))     # a batch of 1000 matrices

Q, R = ml.qr(a)                          # Q (1000, 64, 32), R (1000, 32, 32)
R = ml.qr(a, mode="r")                   # R alone, as numpy's mode="r": Q never formed
Q, R = ml.qr(a, mode="complete")         # Q (1000, 64, 64), R (1000, 64, 32)
U, S, Vt = ml.svd(a)                     # thin: U (1000, 64, 32), S (1000, 32), Vt (1000, 32, 32)
w, V = ml.eigh(a.swapaxes(-1, -2) @ a)   # w ascending (1000, 32), V (1000, 32, 32)

Each call is routed to the fastest Metal kernel for its shape and batch, or to LAPACK on the CPU, by a policy measured on the Mac it runs on:

ml.device_name()                      # 'Apple M5 Pro'
ml.eigh_policy_source()               # 'tuned:Apple M5 Pro'
ml.eigh_backend(32, 4096)             # 'ql': 4096 matrices of 32x32 go to the GPU
ml.eigh_backend(512, 64)              # 'cpu': the CPU's cores win 64 of 512x512
ml.set_eigh_policy(gpu_min_batch=1)   # override the measured policy

Inputs may be mx.array, NumPy arrays or nested lists; outputs are float32 mx.array. ml.eigvalsh and ml.svdvals return the values alone, for less work; ml.cholesky(p, upper=False) factors symmetric positive definite matrices (since 2.18.0; a matrix that is not comes back all NaN), and ml.cholesky_ex(p) also returns LAPACK's info for each; ml.lu_factor, ml.solve and ml.inv work as mx.linalg's, on the GPU for large matrices, and their _ex forms return info too; ml.solve_triangular(a, b, upper=False) as mx.linalg.solve_triangular. The kernels and the routing are described in the main README.

Installing

pip install metal-linalg

On an Apple Silicon Mac with macOS 14 or later and Python 3.10 to 3.14, this installs a prebuilt wheel and the mlx it was built against; nothing is compiled.

MLX is pinned to one release. The package shares MLX's own library and passes mx.array objects through MLX's internals, so it works only with the MLX it was built against, and asks pip for exactly that one (mlx==0.32.3 for this release). Installing it may therefore move your mlx to that version, and upgrading mlx past it on its own breaks the import with an ImportError that says so; upgrade both together, pip install -U metal-linalg mlx, once a release of this package for the newer MLX is out.

From source

On a Mac with no matching wheel, pip builds the package from source, which needs Xcode's command line tools (xcode-select --install) and fetches CMake and the rest itself. To build a clone, or the main branch:

pip install .                                                  # in a clone
pip install git+https://github.com/c0rmac/metal-linalg.git     # main branch

Against another MLX

To build against an MLX other than the pinned one, a source build of MLX or Homebrew's, turn off pip's build isolation so the build sees your MLX, and --no-deps so pip leaves it alone. The build must also use the nanobind your MLX was built with: the pip MLX 0.32.3 uses nanobind 3.0.1, Homebrew's MLX 0.32.1 uses 2.15.0. The build checks this before compiling and stops with a message naming the mismatch.

With Homebrew's Python and MLX (brew install mlx includes the mlx Python package), in a virtual environment that can see Homebrew's packages:

python3 -m venv --system-site-packages ~/.venvs/metal-linalg
source ~/.venvs/metal-linalg/bin/activate
pip install scikit-build-core "nanobind==2.15.0"
pip install --no-build-isolation --no-deps git+https://github.com/c0rmac/metal-linalg.git

After upgrading that MLX, reinstall the package the same way.

Calibration

Routing comes from measurements made on each kind of Mac. Where there are none for yours, or they are stale or incomplete, importing the package raises a metal_linalg.CalibrationWarning once per decomposition, saying so and how to measure your Mac (about an hour and a half) and submit the results:

ml.calibration_status()     # {'qr': 'current', 'eigh': 'current', 'svd': 'stale'}

warnings.filterwarnings("ignore", category=ml.CalibrationWarning), or METAL_LINALG_NO_CALIBRATION_NOTICE=1, silences it. Which chips are measured: the measurements page.

Tests

python -m unittest discover -s python/tests -v