Environment Setup

Linux

1. Install CUDA (skip if already installed, version 12.6 or higher required): LLaMA Factory Environment Installation Documentation

2. It is recommended to use uv to install dependencies. This is a very fast Python environment manager. To install uv, refer to the official uv documentation.

3. After installing uv, use the following commands to create a new Python environment and install the dependencies:

git clone https://github.com/xming521/WeClone.git && cd WeClone
uv venv .venv --python=3.10
source .venv/bin/activate # On Windows, run .venv\Scripts\activate
uv pip install --group main -e .

4. Copy the configuration file template and rename it to settings.jsonc. Subsequent configuration changes will be made in this file:

cp settings.template.jsonc settings.jsonc

NOTE

All configurations for training, inference, etc., are unified in the settings.jsonc file.

5. Use the following command to test if the CUDA environment is correctly configured and can be recognized by PyTorch (for NVIDIA GPU users):

python -c "import torch; print('CUDA available:', torch.cuda.is_available());"

6. (Optional) Install FlashAttention to accelerate training and inference:

uv pip install flash-attn --no-build-isolation

NOTE

For version issues, you can use the pre-built wheels from FlashAttention's prebuild-wheels.

Windows

Running WeClone in a Windows environment has not been rigorously tested, and you cannot use local models to clean data. It is recommended to use WSL2 (with a GPU performance loss of about 5%). Please refer to the WSL2 installation guide.

Congratulations, you have now completed the entire environment setup!