torchannex

torchannex is a curated distribution ("annex") of small, pure-Python libraries that extend PyTorch. Installing torchannex gives you a set of sub-libraries pinned to versions that are known to be compatible with a specific PyTorch release, although these libraries are intended to be loosely coupled with PyTorch and therefore compatible with a variety of releases.

The annex is designed to be cheap: every package in it is a lightweight pure Python package that installs instantly and pulls in no heavy native dependencies of its own beyond PyTorch.

When should I use torchannex?

You do not have to depend on torchannex directly. Two patterns are supported:

1. Batteries-included. Depend on torchannex. You get the full curated set, versioned in lockstep with the PyTorch release you have installed.
2. Hand-picked. Depend on the individual annex packages you care about and pin them yourself. Useful when you have an opinion about a specific sub-library's version and do not want the whole set.

The annex packages are ordinary PyPI packages. The metapackage is a convenience, not a gate.

Installation

pip install torchannex
# or
uv pip install torchannex

If torch is already installed (or pinned elsewhere in your environment), the resolver will select the torchannex version whose PyTorch pin matches:

pip install "torch==2.5.3" torchannex   # resolves to torchannex 2.5.x

Versioning

torchannex versions track PyTorch on the X.Y components only:

Component	Meaning
X	PyTorch major version (lockstep)
Y	PyTorch minor version (lockstep)
Z	Independent annex bugfix counter for the X.Y series

A torchannex X.Y.Z release depends on torch ~= X.Y.0 (i.e. >= X.Y.0, < X.(Y+1).0). This means:

• Any torch X.Y.* patch release is compatible with any torchannex X.Y.*.
• If we ship a fix to an annex sub-library without a new PyTorch release, we bump only the annex Z. For example, for the PyTorch 2.5 series:
  • torchannex 2.5.0 — first annex release for torch 2.5
  • torchannex 2.5.1 — bugfix to an annex sub-library (torch unchanged)
  • torchannex 2.5.2 — another annex bugfix
• When PyTorch 2.6 ships we start a new series at torchannex 2.6.0.

Relationship between torch Z and torchannex Z

They are independent. torchannex 2.5.7 does not imply torch 2.5.7; it simply means the 8th annex release cut for the 2.5 series. Whatever torch 2.5.* you have installed is fine.

What's in the annex?

In the short term, we plan to put spmd_types in the annex to allow for some closer integration with DTensor local map.

Building and releasing

torchannex is a pure metapackage — no Python source — built with setuptools. The empty py-modules / packages entries in pyproject.toml tell setuptools there is nothing to find.

uv build       # produces sdist and wheel
uv publish     # uploads to PyPI (requires credentials)

See TESTING.md for the manual acceptance tests to run before a release.

License

BSD-3-Clause.