Skip to content

How to install for Nvidia 50 series GPUs? #208

@quasiblob

Description

@quasiblob

You repository readme recommends installing from pypi package.

This didn't work for me at all for some reason, Windows 11 and RTX 50 series GPU.

Can't remember the errors, but I got several and I tried different things until I managed to get this thing installed.


If someone else runs into issues with install process, here are the steps, these worked for me at least.
Note - I didn't bother with Flash Attention!

Before install

  • Install huggingface CLI, if it is not installed
  • Required for gated model download - NOTE - install it for your system Python:
    pip install -U "huggingface_hub[cli]

Install steps

  1. Clone the repo
    git clone https://github.com/Stability-AI/stable-audio-tools.git

  2. Go to folder
    cd stable-audio-tools

  3. Create virtual env (my Python version was 3.10.11, you'll have to have a version installed in your system)
    py -3.10 -m venv venv

  4. Activate venv
    Gitbash: source ./venv/scripts/activate or for cmd: call .\venv\scripts\activate

  5. Update pip

  • I did several install experiments, IIRC I got errors about typing-extensions conflict when installing PyTorch CUDA version, but that didn't happen with newer pip 25.x version:
    python.exe -m pip install --upgrade pip
  1. Modify setup.py (NOTE! - this step wasn't required actually, skip it - previous install experiment needed this)
    Open file: notepad setup.py
    Modify line of sentencepiece: sentencepiece==0.2.0

  2. Install torch CUDA version (before installing other dependecies)

  • Installer does not force exact versions of PyTorch, so I installed RTX 50 suitable version beforehand, as otherwise you get CPU versions:
    pip install torch==2.7.0 torchaudio --index-url https://download.pytorch.org/whl/cu128
  1. Then run the setup script:
    pip install .

  2. Do a pip check

  • It should show 'no broken requirements':
    pip check

Run the app:
python ./run_gradio.py --pretrained-name stabilityai/stable-audio-open-1.0

Note;
I get this warning ConnectionResetError: [WinError 10054] An existing connection was forcibly closed... after every audio clip generation, seems like this happens on Windows and is harmless but it looks ugly in terminal.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions