PyPI · #2206 most downloaded on PyPI

autograd

1.9.13.3M downloads/moHIPS/autograd ↗New releases in your daily digest, changelog included.

Efficiently computes derivatives of NumPy code.

Last release 3 months ago
02 Jul 2026
Release timing varies
gaps range from 5 weeks to 2.7 years
Rarely documented
notes for 5 of 33 stable releases
Nothing withdrawn
no release was ever pulled
11 years old
33 releases · first in 2015

Release timeline

33 releases since 2015

2015201720192021

Releases

1.9.12 Jul 2026
This is a small patch release that fixes a gradient regression in reduction operations.
Autograd v1.9.1

29 June 2026

This is a small patch release that fixes a gradient regression in reduction operations.

Bug fixes

Reduction gradient functions now accept **kwargs, so no longer fail with a TypeError when NumPy passes an out=None argument through to them. This restores differentiation of reductions over ArrayBox values in a few edge cases that broke after #717 (#789, by @fjosw).

Full CHANGELOG: v1.9.0...v1.9.1
Open source →
1.9.026 Jun 2026
23 June 2026, released on PyPI on 26 June 2026
Autograd v1.9.0

23 June 2026, released on PyPI on 26 June 2026

We are happy to announce the Autograd v1.9.0 release. It brings broader NumPy 2.x and Python compatibility, several gradient correctness fixes, and better integration with array-like containers such as xarray's DataArray.

This release supports both NumPy 1.x and 2.x. Python versions 3.10 through 3.14 are supported. We have dropped support for Python 3.9, which is now end of life (#748).

Features and general enhancements

Improved integration with NumPy ufuncs and array-like containers, making it possible to differentiate through xarray.DataArrays that hold ArrayBoxes (#717, #716).

Added a 'same' mode for scipy.signal.convolve, including its forward pass and VJP (#721).

Bug fixes

np.where now preserves native Python types instead of promoting them to float64 (#779, fixes #671).

Fixed np.mean, np.std, np.var, and linalg.norm failing when passed a list or tuple containing ArrayBox objects (#757).

Added the missing VJP for np.astype, which was introduced in NumPy 2.0 (#714).

Added the missing JVP for np.conjugate (#713).

Fixed np.absolute gradients so they handle the division-by-zero case the same way np.abs already did (#712).

Registered np.empty_like and np.full_like as non-differentiable functions so they pass through tracing cleanly (#710).

Fixed linalg.norm VJP and JVP for complex numbers (#705, fixes #666).

Maintenance

Adopted PEP 639 for license metadata and dropped the now-redundant license trove classifier (#783, #787).

Followed PEP 765 (#740).

Miscellaneous

Security hardening: added a Dependabot config for GitHub Actions, added Zizmor as a CI job, and pinned and updated pre-commit hooks (#785).

Added exclude-newer = "3 days ago" to guard against supply-chain attacks during dependency resolution (#775).

Replaced the retired macos-13 CI image with macos-15-intel (#747).

Removed the outdated in-tree conda package recipe (#784).

Documentation

Added a tutorial on working with xarray.DataArrays (#782).

Added an examples section with a matplotlib dependency, and tidied up the existing examples (#756, #720).

Updated the README and fixed the instructions for running a specific test (#704, #702, #711).

New contributors

@jeremy-feng made their first contribution in #704

@danra made their first contribution in #705

@jvansanten made their first contribution in #720

@Schefflera-Arboricola made their first contribution in #740

@joewallwork made their first contribution in #756

@aaishwarymishra made their first contribution in #777

@mrityunjai01 made their first contribution in #721

@tylerflex made their first contribution in #757

Full CHANGELOG: v1.8.0...v1.9.0

One column per quarter.

1.8.05 May 2025Deprecated

Fix numpy.core.einsumfunc deprecation by @fjosw in #628

Autograd v1.8.0

We are happy to announce the release of Autograd version 1.8.0.

This version supports Python versions 3.9–3.13 and includes minor features, some bug fixes, and many maintenance updates.

What's changed

Fix numpy.core.einsumfunc deprecation by @fjosw in #628
Run test workflow at the beginning of every month and enable manual trigger of the pipeline by @fjosw in #629
Update tanh(x) derivative README example by @agriyakhetarpal in #635, #636
Add __array_namespace__ to ArrayBox by @yaugenst-flex in #647
Drop Python 3.8 and start testing against Python 3.13 by @agriyakhetarpal in #648
Fix scipy.special.logsumexp type promotion, test against NumPy and SciPy nightlies by @agriyakhetarpal in #643
Pin to NumPy <3 by @agriyakhetarpal in #658
Update license.txt, fix copyright license year by @JasonnnW3000 in #668
Remove scipy.misc by @fjosw in #669
Replace finally statement by except BaseException to suppress all exceptions by @fjosw in #653
Add Linux arm64 CI, pin GitHub Actions to hashes by @agriyakhetarpal in #678
Fix deep learning summer school link by @Saransh-cpp in #682
Skip test_dtypes for now by @agriyakhetarpal in #690
Enable PyPy tests on Linux arm64, and remove pypy-3.9 by @agriyakhetarpal in #697
Export NumPy version from Autograd wrapper by @agriyakhetarpal in #689

New contributors

@yaugenst-flex made their first contribution in #647
@JasonnnW3000 made their first contribution in #668
@Saransh-cpp made their first contribution in #682

Full Changelog: v1.7.0...v1.8.0

Open source →

1.7.022 Aug 2024

This release includes support for NumPy version 2.0.0