pytorch_lightning version issue

[kaitrepte]

Hi,

upon training, I see the following error:

module 'pytorch_lightning.utilities.seed' has no attribute 'seed_everything'.

I am using the most up-to-date version (2.6.5).
Please advise whether a different version is needed, or whether the source code can be updated wrt the newest version.

[QuantumLab-ZY]

Hi, thanks for reporting this.

The error comes from an API change in newer pytorch_lightning releases. In the current codebase, HamGNN calls pl.utilities.seed.seed_everything(...), which works with older Lightning versions but is no longer valid in pytorch_lightning 2.6.5.

At the moment, the provided environment file is still pinned to pytorch-lightning=1.5.10, so the current release is effectively aligned with the 1.x API. That is why you see this issue when using the latest Lightning version.

For now, you have two options:

1. Use the environment/version set recommended by the project, especially pytorch-lightning 1.5.10.
2. Update the source to the newer API (pl.seed_everything(...)) so that it works with newer Lightning releases.

We will look into updating the code to improve compatibility with recent pytorch_lightning versions. Thanks again for pointing this out.

[kaitrepte]

I saw a commit saying that compatibility was established for Pytorch 2.x
Is this the current version available on github?

[QuantumLab-ZY]

Thanks for the follow-up.

You are right that the repository already includes some PyTorch / PyTorch Lightning 2.x compatibility fixes, but the specific seeding call reported in this issue was still using the older Lightning API:

pl.utilities.seed.seed_everything(...)

This has now been updated to the public Lightning API:

pl.seed_everything(...)

pl.seed_everything is supported by both older Lightning 1.x releases and newer Lightning 2.x releases, so this should resolve the error with recent pytorch_lightning versions such as 2.6.5 without requiring a version-specific workaround.
I also verified the change with:

python -m compileall hamgnn/main.py
python -c "import pytorch_lightning as pl; print(pl.seed_everything)"

Both checks passed.

[kaitrepte]

Hi,

thanks for clarifying.
The major issue I saw after fixing the source code wrt seed_everything was that
the Trainer class inside pytorch_lightning has changed almost all its inputs.
Have you adjusted those as well?

Best,
Kai

[QuantumLab-ZY]

Thanks for the follow-up.

You are right that full compatibility with the Lightning 2.x API would require updating a few more code paths beyond seed_everything, especially the Trainer arguments and some epoch-end hooks. The required source changes are not very large, and we originally planned to make them immediately.

However, because HamGNN still needs to preserve compatibility with the historical training setup and checkpoints, we are being cautious about changing the training stack too aggressively. For the current release, the recommended solution is to create the Python environment from the provided HamGNN.yaml file in the repository root:

conda env create -f ./HamGNN.yaml
conda activate HamGNN

This environment is intended to be fully compatible with the current HamGNN codebase and pins the dependency versions accordingly, including pytorch-lightning=1.5.10. Using this environment should avoid the Lightning 2.x API incompatibilities you encountered.
We will continue evaluating full Lightning 2.x support, but for now please use the provided HamGNN.yaml environment for the most reliable setup.

[kaitrepte]

Ok.

Please let me know once the full compatibility is achieved.

Thanks,
Kai

[kaitrepte]

Hi,

any updates on this?

Thanks,
Kai

[kaitrepte]

Is there a requirements.txt file that one can use instead of the yaml file which requires conda?

[QuantumLab-ZY]

Hi Kai,

A PyTorch Lightning 2.x compatible version is available on the feat/pl2x-delta-learning branch:

https://github.com/QuantumLab-ZY/HamGNN/tree/feat/pl2x-delta-learning

We understand that a pip-based installation is useful for virtual environments, containers, and systems where Conda is unavailable. However,
simply converting the YAML file to requirements.txt would not be sufficient: PyTorch and PyTorch Geometric packages may require different wheels depending on the CUDA version.

For now, please refer to this branch for Lightning 2.x support.

Best,
Yang