If you own a video adapter or computer with NVidia graphics, and get CUDA errors starting exui, you may need to install precompiled wheels for PyTorch and/or exllamav2. These instructions are written with Windows 11 in mind but they may also work for Windows 10.
- Before doing anything, make sure official NVidia drivers are installed. Versions installed automatically during Windows installation may not function as expected. Download drivers from the NVidia Download page.
- Uninstall torch and exllama2.
pip uninstall torch
pip uninstall exllamav2
NVidia documents which products support CUDA, and basically any NVidia GT, GTX and RTX device will support CUDA. However they do not document which CUDA versions are supported for each card. General guidance: Newer GTX and all RTX cards likely support CUDA 12.1 or later. Older cards are more likely to support CUDA 11.8.
- Play it safe: use CUDA 11.8
- Have newer gaming hardware? Try CUDA 12.1
Visit the PyTorch Get Started Locally page. Choose the appropriate options to be given the PIP URL to intall PyTorch.
- PyTorch Build: stable
- Your OS
- Package type: PIP
- Language: Python
- Computer Platform: select the CUDA version that matches your video adapter.
Copy and paste the command given by the PyTorch website into your Terminal.
Example: pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
Once PyTorch is installed, it is time to install exllamav2.
- Download the correct pre-compiled wheel. Choose the correct wheel. For example if
you have Windows 11 or Windows 10 running on a 64-bit CPU (most common), are using CUDA 12.1, and Python 3.10, download
exllamav2-0.0.9+cu121-cp310-cp310-win_amd64.whl
. If you are using Python 3.11, chooseexllamav2-0.0.9+cu121-cp311-cp311-win_amd64.whl
. - Use pip to install the downloaded wheel:
Example: pip install path\to\downloaded\file\exllamav2-0.0.9+cu121-cp311-cp311-win_amd64.whl.
With those packages installed, you should now be able to launch exui by entering python server.py
in the Terminal.