Taylor Mode Autodiff in PyTorch

This library provides a PyTorch implementation of Taylor mode automatic differentiation, a generalization of forward mode to higher-order derivatives. It is similar to JAX's Taylor mode (jax.experimental.jet).

The repository also hosts the Python functionality+experiments and LaTeX source for our NeurIPS 2025 paper "Collapsing Taylor Mode Automatic Differentiation", which allows to further accelerate Taylor mode for many practical differential operators.

🔪 Warning: expect rough edges! 🔪

This is a research prototype with various limitations (e.g. operator coverage). We highly recommend double-checking your results with PyTorch's autodiff. Please help us improve the package by providing feedback, filing issues, and opening pull requests.

Getting Started

Installation

pip install jet-for-pytorch

Examples

See the documentation.

Citing

If you find the jet package useful for your research, consider citing

@inproceedings{dangel2025collapsing,
  title =        {Collapsing Taylor Mode Automatic Differentiation},
  author =       {Felix Dangel and Tim Siebert and Marius Zeinhofer and Andrea
                  Walther},
  year =         2025,
  booktitle =    {Advances in Neural Information Processing Systems (NeurIPS)},
}