This repository uses the uv package manager To install uv pleaes refer to the official uv website and make sure to add the uv binary to $PATH.
Installing dependency is as simple as,
uv sync
The Flash Attention build script is not compatible with uv and require you CUDA 11.4+
To properly install Flash Attentoin:
- Load CUDA
module load cuda/12.4.1-fasrc01
- Install FlashAttention build script dependencies and build with --no-build-isolation
uv pip install torch setuptools
uv add flash-attn --no-build-isolation
uv sync
Notebooks are provided to play with the sampling algorithms with trained models. To use VSCode Jupyter with GPUs, the easiest way is to login to a compute node and establish a connection tunnel. This is very easy to do with the VSCode server module on the FASRC cluster,
salloc [GPU Things]
module load vscode
code tunnel
The codebase employ a ruff pre-commit hook for style fomartting.
After you've installed the necessary dependencies, install the pre-commit hooks by,
pre-commit install