Fixing model loading logic (names and cache dir) for fairchem models

[niklashoelter]

Previously: model (str | Path | None) was passed as to fairchem's get_predict_unit - although this is expecting a model name and thereby throws an error even if a correct model path is set.

According to Fairchem, when loading a cached model, the cache_dir parameter needs to be set in addition to a valid model_name. This is fixed here.

Summary

According to the fairchem's pretrained_mlip.get_predict_unit logic, I made the model_name a required parameter when initalizing a FairChemModel object and added a second, optional model_cache_dir string parameter that is passed to the get_predict_unit function if set.

As this is my first contribution, hope that I did not forget anything!

Checklist

Before a pull request can be merged, the following items must be checked:

• Doc strings have been added in the Google docstring format.
• Run ruff on your code.
• Tests have been added for any new functionality or bug fixes.

[CompRhys]

Thanks for contributing!

I would need a bit more time to dig into what's happened here but this may be because of #211 and #270 where FairChemModel now corresponds to the v2 fairchem library. Older fairchem models can be accessed in the farichem_legacy.py / FairChemV1Model. Does just swapping to using FairChemV1Model get the behavior you want?

[niklashoelter]

@CompRhys thanks for your quick answer!
I did not check with the v1 fairchem models as I am relying on the UMA ones. The simple fix from my branch however worked for me.

[CompRhys]

I think the mistake is that we need something along the lines of https://github.com/facebookresearch/fairchem/blob/dcd772e7921969643dc8ff71f2430326e2afad57/src/fairchem/core/calculate/ase_calculator.py#L132-L149. We would still want to allow users to load checkpoints from paths even if we expect majority to reuse models released by meta.

[niklashoelter]

Tried to implement it as you suggested while keeping the cache dir option which is None by default.

[CompRhys]

Made a few tweaks for better back compatibility and to get the tests to work.

[CompRhys]

Fairchem tests fail because I believe your branch won't have access to the HF_TOKEN we have in the repo secrets. GraphPES errors are known to be flaky.