Python 3.14 migration

[mattwthompson]

Upstream dependencies:

• AmberTools
  • ParmEd (Rebuild for python 3.14 conda-forge/parmed-feedstock#46)
• OpenEye Toolkits
• DGL (for NAGL training) (this probably won't happen and a refactor to PyTorch Geometric is likely necessary)
• RDKit
• PyTorch (Rebuild for python 3.14 conda-forge/pytorch-cpu-feedstock#420 ?)
• OpenMM
• Psi4
• ForceBalance

[mattwthompson]

In practical terms, this is blocked by AmberTools and OpenEye until OpenFF 2.3.0 comes out¹. Without at least one of them, there's no way for mainline force fields to assign AM1-BCC partial charges, so there isn't really a point to packaging core infrastructure.

Footnotes

1. Currently it's in a release candidate: https://github.com/openforcefield/openff-forcefields/releases/tag/2025.10.0 ↩

[mattwthompson]

I suspect the first blocker to fall would be a way to get openff-toolkit-base working on 3.14 using OpenFF NAGL (inference-only) as the only AM1-BCC ~provider

[mattwthompson]

But that's still waiting on upstreams

(openff-nagl-test) [openff-nagl] micromamba create --name temp openff-toolkit-base openff-nagl-base "python=3.14" "rdkit>=2025.09" "pytorch>=2.9"
conda-forge/noarch                                          Using cache
conda-forge/osx-arm64                                       Using cache

error    libmamba Could not solve for environment specs
    The following packages are incompatible
    ├─ openff-nagl-base =* * is installable and it requires
    │  └─ rdkit =* * with the potential options
    │     ├─ rdkit [2022.03.1|2022.03.2|...|2025.09.4] would require
    │     │  └─ python >=3.10,<3.11.0a0 *, which can be installed;
    │     ├─ rdkit [2022.09.1|2022.09.3|...|2025.09.4] would require
    │     │  └─ python >=3.11,<3.12.0a0 *, which can be installed;
    │     ├─ rdkit [2021.03.1|2021.03.2|...|2024.03.5] would require
    │     │  └─ python >=3.8,<3.9.0a0 *, which can be installed;
    │     ├─ rdkit [2021.03.1|2021.03.2|...|2025.03.5] would require
    │     │  └─ python >=3.9,<3.10.0a0 *, which can be installed;
    │     ├─ rdkit [2023.03.3|2023.09.1|...|2025.09.4] would require
    │     │  └─ python >=3.12,<3.13.0a0 *, which can be installed;
    │     ├─ rdkit [2024.09.1|2024.09.2|...|2025.09.4] would require
    │     │  └─ python [>=3.13,<3.14.0a0 *_cp313|>=3.13.0rc2,<3.14.0a0 *_cp313|>=3.13.0rc3,<3.14.0a0 *_cp313], which can be installed;
    │     └─ rdkit [2025.09.1|2025.09.2|2025.09.3|2025.09.4] would require
    │        └─ python >=3.14,<3.15.0a0 *_cp314, which conflicts with any installable versions previously reported;
    ├─ python =3.14 * is not installable because there are no viable options
    │  ├─ python [3.14.0|3.14.1|3.14.2] conflicts with any installable versions previously reported;
    │  ├─ python [3.14.0|3.14.1|3.14.2] conflicts with any installable versions previously reported;
    │  └─ python [3.14.0rc1|3.14.0rc2|3.14.0rc3] would require
    │     └─ _python_rc =* *, which does not exist (perhaps a missing channel);
    └─ pytorch >=2.9 * is installable with the potential options
       ├─ pytorch 2.9.1 would require
       │  └─ python >=3.10,<3.11.0a0 *, which can be installed;
       ├─ pytorch 2.9.1 would require
       │  └─ python >=3.11,<3.12.0a0 *, which can be installed;
       ├─ pytorch 2.9.1 would require
       │  └─ python >=3.12,<3.13.0a0 *, which can be installed;
       └─ pytorch 2.9.1 would require
          └─ python >=3.13,<3.14.0a0 *, which can be installed.
critical libmamba Could not solve for environment specs

[reneleonhardt]

Don't understand the PyTorch release process:

• Python 3.13 is supported (2.6.0 on Jan 29, 2025), free-threaded wheels since 2.7.0
• Python 3.14 is supported (2.9.0 on Oct 15, 2025), free-threaded wheels since 2.9.0 but the issue shows unsupported

pytorch/pytorch#156856 (waiting since June 25, 2025)

Let's hope the Python 3.15 release schedule is less hidden from the PyTorch release process.
https://peps.python.org/pep-0790/

[mattwthompson]

It looks like it could simply be Conda packages lagging behind the Python packages: https://pypi.org/project/torch/2.9.1/#files

[mikemhenry]

conda-forge/parmed-feedstock#49

parmed should support python 3.14 now