Loosen torchaudio version in requirements_minimal.txt

[samuelbradshaw]

Hi! This pull request loosens the dependency version for torchaudio in requirements_minimal.txt.

If I understand the README correctly, there are two sets of dependencies:

• requirements_minimal.txt for simple stem separation. This is what's used when installing via pip install demucs.
• requirements.txt or environment-[cpu|cuda].yml. This is for training.

Training requires torchaudio<2.2, but stem separation appears to work correctly with later versions of torchaudio (see #7). Therefore, I think the dependencies in requirements_minimal.txt can be safely loosened to allow later torchaudio versions. I have tested stem separation with torchaudio 2.8.0 and everything appears to work correctly.

[akx]

Yep, this works great for inference.

[DaveParr]

I wonder if >2.9 might cause problems? uv tool install demucs --with torchcodec failed to run uvx demucs /home/dave/Development/stems/tmp.wav with:

uvx demucs /home/dave/Development/stems/tmp.wav
Important: the default model was recently changed to `htdemucs` the latest Hybrid Transformer Demucs model. In some cases, this model can actually perform worse than previous models. To get back the old default model use `-n mdx_extra_q`.
Selected model is a bag of 1 models. You will see that many progress bars per track.
Separated tracks will be stored in /home/dave/separated/htdemucs
Namespace(tracks=[PosixPath('/home/dave/Development/stems/tmp.wav')], sig=None, name='htdemucs', repo=None, verbose=False, out=PosixPath('separated'), filename='{track}/{stem}.{ext}', device='cuda', shifts=1, overlap=0.25, split=True, segment=None, stem=None, int24=False, float32=False, clip_mode='rescale', flac=False, mp3=False, mp3_bitrate=320, mp3_preset=2, jobs=0)
Separating track /home/dave/Development/stems/tmp.wav
100%|██████████████████████████████████████████████████████████████████████| 204.75/204.75 [00:08<00:00, 24.41seconds/s]
/home/dave/.local/share/uv/tools/demucs/lib/python3.12/site-packages/torchaudio/__init__.py:178: UserWarning: The 'encoding' parameter is not fully supported by TorchCodec AudioEncoder.
  return save_with_torchcodec(
/home/dave/.local/share/uv/tools/demucs/lib/python3.12/site-packages/torchaudio/__init__.py:178: UserWarning: The 'bits_per_sample' parameter is not directly supported by TorchCodec AudioEncoder.
  return save_with_torchcodec(

in environment:

Resolved 43 packages in 192ms
Installed 43 packages in 162ms
 + antlr4-python3-runtime==4.9.3
 + cloudpickle==3.1.2
 + demucs==4.0.1
 + dora-search==0.1.12
 + einops==0.8.1
 + filelock==3.20.1
 + fsspec==2025.12.0
 + jinja2==3.1.6
 + julius==0.2.7
 + lameenc==1.8.1
 + markupsafe==3.0.3
 + mpmath==1.3.0
 + networkx==3.6.1
 + numpy==2.4.0
 + nvidia-cublas-cu12==12.8.4.1
 + nvidia-cuda-cupti-cu12==12.8.90
 + nvidia-cuda-nvrtc-cu12==12.8.93
 + nvidia-cuda-runtime-cu12==12.8.90
 + nvidia-cudnn-cu12==9.10.2.21
 + nvidia-cufft-cu12==11.3.3.83
 + nvidia-cufile-cu12==1.13.1.3
 + nvidia-curand-cu12==10.3.9.90
 + nvidia-cusolver-cu12==11.7.3.90
 + nvidia-cusparse-cu12==12.5.8.93
 + nvidia-cusparselt-cu12==0.7.1
 + nvidia-nccl-cu12==2.27.5
 + nvidia-nvjitlink-cu12==12.8.93
 + nvidia-nvshmem-cu12==3.3.20
 + nvidia-nvtx-cu12==12.8.90
 + omegaconf==2.3.0
 + openunmix==1.3.0
 + pyyaml==6.0.3
 + retrying==1.4.2
 + setuptools==80.9.0
 + submitit==1.5.4
 + sympy==1.14.0
 + torch==2.9.1
 + torchaudio==2.9.1
 + torchcodec==0.9.1
 + tqdm==4.67.1
 + treetable==0.2.6
 + triton==3.5.1
 + typing-extensions==4.15.0
Installed 1 executable: demucs

however uv tool install demucs --with torch==2.8.0 ran uvx demucs /home/dave/Development/stems/tmp.wav with:

Important: the default model was recently changed to `htdemucs` the latest Hybrid Transformer Demucs model. In some cases, this model can actually perform worse than previous models. To get back the old default model use `-n mdx_extra_q`.
Selected model is a bag of 1 models. You will see that many progress bars per track.
Separated tracks will be stored in /home/dave/separated/htdemucs
Namespace(tracks=[PosixPath('/home/dave/Development/stems/tmp.wav')], sig=None, name='htdemucs', repo=None, verbose=False, out=PosixPath('separated'), filename='{track}/{stem}.{ext}', device='cuda', shifts=1, overlap=0.25, split=True, segment=None, stem=None, int24=False, float32=False, clip_mode='rescale', flac=False, mp3=False, mp3_bitrate=320, mp3_preset=2, jobs=0)
Separating track /home/dave/Development/stems/tmp.wav
100%|██████████████████████████████████████████████████████████████████████| 204.75/204.75 [00:08<00:00, 25.11seconds/s]
/home/dave/.local/share/uv/tools/demucs/lib/python3.12/site-packages/torchaudio/_backend/utils.py:337: UserWarning: In 2.9, this function's implementation will be changed to use torchaudio.save_with_torchcodec` under the hood. Some parameters like format, encoding, bits_per_sample, buffer_size, and ``backend`` will be ignored. We recommend that you port your code to rely directly on TorchCodec's encoder instead: https://docs.pytorch.org/torchcodec/stable/generated/torchcodec.encoders.AudioEncoder
  warnings.warn(
/home/dave/.local/share/uv/tools/demucs/lib/python3.12/site-packages/torchaudio/_backend/ffmpeg.py:247: UserWarning: torio.io._streaming_media_encoder.StreamingMediaEncoder has been deprecated. This deprecation is part of a large refactoring effort to transition TorchAudio into a maintenance phase. The decoding and encoding capabilities of PyTorch for both audio and video are being consolidated into TorchCodec. Please see https://github.com/pytorch/audio/issues/3902 for more information. It will be removed from the 2.9 release. 
  s = torchaudio.io.StreamWriter(uri, format=muxer, buffer_size=buffer_size)

in environment:

Resolved 41 packages in 192ms
Prepared 4 packages in 1m 23s
Uninstalled 6 packages in 121ms
Installed 4 packages in 100ms
 - nvidia-nccl-cu12==2.27.5
 + nvidia-nccl-cu12==2.27.3
 - nvidia-nvshmem-cu12==3.3.20
 - torch==2.9.1
 + torch==2.8.0
 - torchaudio==2.9.1
 + torchaudio==2.8.0
 - torchcodec==0.9.1
 - triton==3.5.1
 + triton==3.4.0
Installed 1 executable: demucs

Output completed though and audio separation seemed successful with 4 decent outputs
🥇

[samuelbradshaw]

Thanks! I updated this pull request to use torchaudio>=0.8,<2.9.