Torch compatibility

PyTorch’s C++ ABI is not guaranteed to be stable across minor releases, so a NEML2 wheel from PyPI is only guaranteed to load and run correctly with the torch versions listed below. Every combination shown here is exercised by NEML2’s CI on each change.

Supported range

torch >= 2.10.0, <= 2.12.1

The bounds are asymmetric on purpose:

  • The lower bound is a hard pyproject.toml constraint; older torch versions have a known-incompatible ABI.

  • The upper bound is advisory. A newly-released torch may or may not work; rather than blocking users, pip install neml2 accepts any newer torch. The upper bound documents which versions have actually been verified.

Combinations outside this list may still work — for example, a freshly-released torch that hasn’t been verified yet — but they are not regression-tested. If you need an additional combination, please open an issue.

Supported combinations

linux

  • Python 3.10: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.11: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.12: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.13: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.14: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

macOS

  • Python 3.10: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.11: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.12: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.13: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1

  • Python 3.14: torch 2.10.0, 2.11.0, 2.12.0, 2.12.1