v4.1.3

Fix: catch all exceptions from torchaudio.load() for torchcodec fallback

v4.1.2

What's New

Added

• Google Colab notebook (notebooks/quickstart_colab.ipynb)
  • 7-section workflow: install, download model, upload audio, run separation, check outputs, preview results, download results
  • Support for both CLI and Python API methods
  • Colab badge and "Try it in Colab" section in README
• Soundfile fallback for audio I/O operations
  • save_audio() falls back to soundfile when torchaudio fails
  • _load_audio() now has 3 fallback levels: ffmpeg → torchaudio → soundfile

Fixed

• Python 3.12+ compatibility: Fixed torchaudio/torchcodec dependency issue in newer Python environments (e.g., Google Colab)

Full Changelog: v4.1.1...v4.1.2

v4.1.1

Highlights

• Added a full pytest suite (unit + slow integration) and CI configuration so we automatically verify imports, CLI usage, and model loading before every release.
• Documented the new testing workflow in docs/dev/PRINCIPLES.md and docs/dev/MAINTENANCE.md so future updates stay consistent.
• Added module aliasing in demucs_infer.compat to keep pretrained checkpoints working after the package rename, and fixed the CLI parser name to show demucs-infer everywhere.
• Simplified packaging by making sure YAML/TXT configs ship with the wheel and adding validation to the publish workflow to catch missing artifacts.

Testing

• uv run pytest tests/ -v
• uv run python examples/basic_separation.py
• uv run demucs-infer --help
• uv run demucs-infer --list-models

v4.1.0

Overview

demucs-infer is an inference-only fork of the original Demucs by Alexandre Défossez and Meta AI Research. This package was created to provide ongoing maintenance and PyTorch 2.x compatibility for Demucs inference capabilities, as the original repository is no longer actively maintained.

Added

• PyTorch 2.x Support: Full compatibility with PyTorch 2.0+ and modern torchaudio versions

  • Removed torchaudio<2.1 version restriction
  • Updated all dependencies for PyTorch 2.x compatibility

• Inference-Only Packaging: Streamlined package focusing solely on inference

  • Removed training code (~50% package size reduction)
  • Removed evaluation scripts
  • Kept all inference models and algorithms (100% unchanged)

• Modern Dependency Management:

  • UV package manager support with fast installation
  • Minimal core dependencies (7 packages vs 15+ in original)
  • Optional dependency groups: [mp3], [quantized], [dev]

• Enhanced Documentation:

  • Comprehensive README.md with installation guides
  • Migration guide from original Demucs
  • Implementation notes and technical details
  • Prominent attribution to original research

• CLI Tool: demucs-infer command (avoids conflicts with original demucs)

Changed

• Package Name: demucs → demucs-infer (no naming conflicts)
• Import Name: demucs → demucs_infer (explicit, no conflicts)
• License: Updated with dual copyright attribution
  • Original: Copyright (c) Meta Platforms, Inc. and affiliates
  • Modifications: Copyright (c) 2025 Bo-Yu Chen

Removed

• Training Infrastructure:
  • Training scripts (train.py, solver.py, etc.)
  • Training dependencies (hydra, dora-search, omegaconf, submitit)
  • Dataset utilities (musdb, museval)
  • Evaluation scripts

Unchanged

All core functionality from original Demucs is preserved with zero modifications:

• ✅ All separation models (HTDemucs, HTDemucs-FT, HTDemucs-6s, MDX, MDX-Extra, quantized variants)
• ✅ Model architectures (identical neural networks)
• ✅ Separation algorithms (identical audio processing)
• ✅ Model weights (same pretrained checkpoints from official repositories)
• ✅ Audio quality (100% identical output to original Demucs)

Credits

All credit for the models, algorithms, and research belongs to:

• Alexandre Défossez (Original author)
• Meta AI Research (Original research team)

Research papers:

• Hybrid Demucs (2021): arXiv:2111.03600
• Hybrid Transformer Demucs (2022): arXiv:2211.08553

Maintenance and PyTorch 2.x compatibility:

• Bo-Yu Chen and worzpro Development Team