v0.1.0 — first release

First release of demucs-onnx: working ONNX export + inference for HT-Demucs FT.

Install

pip install demucs-onnx
demucs-onnx separate song.mp3 stems/

Models auto-download from Hugging Face on first run.

The 4 export blockers this package solves

Vanilla torch.onnx.export fails on htdemucs_ft for four reasons. This package fixes all of them and parity-checks the result vs PyTorch fp32 (< 1e-3 max abs diff) before writing the .onnx file.

1. torch.stft returns complex tensors → replaced with Conv1d of precomputed sin/cos DFT bases (export/stft.py).
2. model.segment = Fraction(39, 5) breaks dynamo / lowering → coerced to float (export/segment.py).
3. random.randrange in the transformer pos-embedding → patched to deterministic shift=0 (export/pos_embed.py).
4. aten::_native_multi_head_attention has no ONNX symbolic → manual scaled-dot-product attention (export/mha.py).

What you get

• Inference: from demucs_onnx import separate, separate_stem — pure numpy + onnxruntime, no PyTorch needed. CPU / CoreML / CUDA / DirectML EPs.
• CLI: demucs-onnx separate ..., demucs-onnx export ..., demucs-onnx list-models.
• Export: from demucs_onnx.export import export_to_onnx (requires pip install "demucs-onnx[export]").

Links

• PyPI: https://pypi.org/project/demucs-onnx/
• Pre-exported ONNX models on Hugging Face:
  • https://huggingface.co/StemSplitio/htdemucs-ft-onnx (full 4-stem bag)
  • https://huggingface.co/StemSplitio/htdemucs-ft-drums-onnx
  • https://huggingface.co/StemSplitio/htdemucs-ft-bass-onnx
  • https://huggingface.co/StemSplitio/htdemucs-ft-other-onnx
  • https://huggingface.co/StemSplitio/htdemucs-ft-vocals-onnx

MIT licensed, matching the original facebookresearch/demucs.