Skip to main content

Installation

Apple Silicon required

VeloxQuant-MLX uses Metal GPU kernels compiled at runtime. It requires macOS on an Apple M-series chip (M1 or later). Intel Macs and Linux are not supported.

Requirements​

RequirementVersion
macOS13 Ventura or later
Apple SiliconM1, M2, M3, M4 (any tier)
Python3.10 – 3.14
MLX≥ 0.18
NumPy≥ 1.26

Install from PyPI​

pip install veloxquant-mlx

This installs the library and its runtime dependencies (MLX, NumPy, psutil).

Install from source​

Use this if you want the latest development version or plan to contribute.

git clone https://github.com/rajveer43/VeloxQuant-MLX.git
cd VeloxQuant-MLX
pip install -e ".[dev]"

The [dev] extra installs the test suite and tooling dependencies (pytest, psutil, build, twine, ruff). SciPy is a core runtime dependency (used for codebook training), so it is already installed with the base package — no extra needed.

Optional extras​

pip install "veloxquant-mlx[plots]"

The [plots] extra installs Matplotlib, needed only for the benchmark-plotting utilities. It is not required for compression itself.

Verify the installation​

import veloxquant_mlx

print(veloxquant_mlx.__version__)  # should print the version you just installed

import mlx.core as mx

print(mx.default_device())  # Device(gpu, 0)

If you see Device(cpu, 0) instead of gpu, MLX is not using Metal. See the troubleshooting section below.

Most users will want to run VeloxQuant-MLX with mlx_lm for model loading and generation:

pip install mlx-lm

Check compatibility:

python -c "import mlx_lm; print(mlx_lm.__version__)"

Troubleshooting​

Device(cpu, 0) — Metal not active​

MLX falls back to CPU when Metal is unavailable. Common causes:

  1. Running in a VM or Docker — Metal is not forwarded; run natively on the Mac host.
  2. Rosetta 2 Python — Install an arm64 Python via Homebrew or the official installer: 
    brew install python@3.12
  3. Conda environment — Use miniforge (arm64) instead of standard Anaconda: 
    brew install miniforge
    conda create -n velox python=3.12
    conda activate velox
    pip install veloxquant-mlx

ImportError: No module named 'mlx'​

MLX is only published for Apple Silicon. If pip install mlx fails, you are on an incompatible platform.

Metal kernel compilation errors​

VeloxQuant-MLX compiles Metal kernels on first use with mx.fast.metal_kernel. If you see compilation errors:

# Ensure Xcode Command Line Tools are installed
xcode-select --install

# Confirm Metal availability
python -c "import mlx.core as mx; mx.fast.metal_kernel"

ArtifactNotFoundError on first run​

Some algorithms (VecInfer, SpectralQuant) require calibration artifacts before first use. There is no single precompute --method ... --model ... command — each method calibrates differently:

  • VecInfer — calibrate smooth factors and train a codebook from your model's own key/value activations (calibrate_smooth_factors, train_codebook).
  • SpectralQuant — calibrate per-layer rotation matrices (calibrate_spectral_rotation, cached to disk by model name).
  • Zero-calibration methods (TurboQuant, QJL, PolarQuant) can precompute their analytical codebooks/rotation matrices ahead of time with: 
    python -m veloxquant_mlx precompute --head_dim 128 --bits 1 2 3 4 --jl_dim 128 --output_dir ./artifacts/

See the Calibration Guide for the full, verified walkthrough of each method's real calibration flow.

Next steps​