Contributing
Contributions are welcome, especially independently reproduced RDNA 4 results.
Before opening a pull request:
ruff check .
ruff format --check .
pytest
python -m build
twine check dist/*
GPU-dependent changes must include:
- GPU model and architecture;
- OS, kernel, host ROCm, PyTorch, torchvision, Triton, and Transformers versions;
- the exact model revision;
- the command used;
- peak allocated VRAM and wall time;
- a statement about output comparison or quality evidence;
- confirmation that unrelated GPUs and system Python were not modified.
Do not commit model weights, generated OCR from private documents, Hugging Face tokens, or raw home-directory paths.