I built Advect, an automatic differentiation library for scientific Python. It provides broad NumPy API coverage, follows the Python Array API standard, and works with ordinary Python functions.

In addition to dynamic differentiation, Advect can stage a function into a reusable program that can be optimized, differentiated again, serialized, and loaded later. There is also some support for SciPy, xarray, and interoperability with other autodiff frameworks.

The linked playground runs entirely in the browser through Pyodide, so you can try tracing and differentiating an expression without installing anything.

The first public release is available with pip install advect.

Source: https://github.com/yaugenst/advect