demucs-mlx 1.4.1

Music source separation with MLX acceleration.

demucs-mlx

Split any song into its individual stems — vocals, drums, bass, and other instruments — directly on your Mac.

demucs-mlx is a fast, native Apple Silicon port of Meta's Demucs music source separation model, built on MLX. No PyTorch required.

Features

• ~73x realtime on Apple Silicon — 2.6x faster than Demucs with PyTorch MPS
• Bit-exact parity with upstream Demucs stems (within floating-point tolerance)
• Custom fused Metal kernels (GroupNorm+GELU, GroupNorm+GLU, OLA)
• Metal-free fallbacks for non-Apple platforms (Linux)
• No PyTorch required at inference time
• Audio I/O via mlx-audio-io
• STFT/iSTFT via mlx-spectro

Requirements

• Python >= 3.10
• macOS with Apple Silicon (recommended) or Linux with MLX

Install

pip install demucs-mlx

On first run, demucs-mlx will automatically download and convert the PyTorch weights to MLX format. This requires the convert extra:

pip install 'demucs-mlx[convert]'

Once weights are cached in ~/.cache/demucs-mlx, the convert extra is no longer needed.

CLI usage

demucs-mlx /path/to/audio.wav

Options:

-n, --name          Model name (default: htdemucs)
-o, --out           Output directory (default: separated)
--shifts            Number of random shifts (default: 1)
--seed              Optional RNG seed for reproducible shifts (default: none)
--overlap           Overlap ratio (default: 0.25)
-b, --batch-size    Batch size (default: 8)
--write-workers     Concurrent writer threads (default: 1)
--list-models       List available models
-v, --verbose       Verbose logging

Python usage

from demucs_mlx import Separator

separator = Separator()
origin, stems = separator.separate_audio_file("song.wav")

# stems is a dict: {"drums": array, "bass": array, "other": array, "vocals": array}
for name, audio in stems.items():
    print(f"{name}: {audio.shape}")

To keep outputs as MLX arrays (avoids GPU-to-CPU copy):

origin, stems = separator.separate_audio_file("song.wav", return_mx=True)

For reproducible shift sampling (while keeping shifts=1 behavior), pass a seed:

separator = Separator(model="htdemucs", shifts=1, seed=0)
origin, stems = separator.separate_audio_file("song.wav")

What changed in 1.4.0

• Fixed shifted-inference TensorChunk propagation so chunk length/offset is handled correctly in all paths.
• Added optional deterministic RNG control (seed) for Python API and CLI.
• Default behavior is unchanged: shifts=1 remains stochastic unless seed is provided.

Performance

Benchmarked on a 3:15 stereo track (44.1 kHz, 16-bit) using htdemucs with default settings:

Package	Backend	Time	Speedup
demucs 4.0.1	PyTorch (CPU)	52.3s	0.1x
demucs 4.0.1	PyTorch (MPS)	6.9s	1x
demucs-mlx 1.1.0	MLX + Metal	2.7s	2.6x

Apple M4 Max, 128 GB. All runs use htdemucs with default settings and a single warm-up pass before timing.

Models

Model	Sources	Description
htdemucs	4	Hybrid Transformer Demucs (default)
htdemucs_ft	4	Fine-tuned HTDemucs
htdemucs_6s	6	6-source (adds piano, guitar)
hdemucs_mmi	4	Hybrid Demucs MMI
mdx	4	Music Demixing model
mdx_extra	4	MDX with extra training

MLX model cache

Pre-converted MLX weights are cached under ~/.cache/demucs-mlx. Delete to force re-conversion.

Documentation

• API reference: docs/api.md
• Development workflow: docs/development.md
• Platform notes: docs/platform.md

License

MIT. Based on Demucs by Meta Research. See LICENSE for details.