Deep Learning framework to accelerate the training of PyTorch models.

---

Boost • Docs • Community • Contribute

PyTorch-Boost is a deep learning optimization library, offering a lightweight and user-friendly wrapper for custom CUDA functions. Designed to accelerate PyTorch model training and inference, PyTorch-Boost facilitates the application of high-precision and low-precision modalities— encompassing 32-bit, 16-bit, 8-bit, 4-bit, and mixed-precision training—across multiple GPUs or machines with its distributed training capabilities.

For detailed usage, installation guidelines, and more, please refer to our comprehensive documentation.

---

Install PyTorch-Boost

Simple installation from PyPI

pip install pytorch-boost

Other installation options

Install with optional dependencies

pip install pytorch-boost['extra']

Conda

conda install pytorch-boost -c conda-forge

Install stable version

pip install https://github.com/kswain55/pytorch-boost/archive/refs/heads/release/stable.zip -U

Install bleeding-edge

pip install https://github.com/kswain55/pytorch-boost/archive/refs/heads/master.zip -U

---

Compatibility

Supported Platform:

• Linux

Supported Hardware:

• NVIDIA Hopper, Ada, Ampere, Volta, and Turing architecture
• AMD CDNA 2/3 and RDNA 2/3

Supported Software Layer:

• CUDA 11.8 - 12.2 (NVIDIA)
• ROCM 5.3 and HIP compilation tools (AMD)

---

Acknowledgements

We would like to thank Tim Dettmers for his work bitsandbytes and the team developing PyTorch, allowing us to make this possible. We would also like to thank NVIDIA and AMD for providng the neccessary hardware for testing.

---

How to cite us

If you found this library useful, please consider citing our work:

@article{boost2023pytorch,
  title={Boost},
  author={minyoung huh, kswain},
  journal={arXiv preprint arXiv:2023.2023},
  year={2023}
}