Fixes needed to run with current (2026) package versions

[AndreasPlt]

Thanks for publishing this codebase, it has been a useful reference for my own work!

Running the training scripts with current package versions required a few changes. Documenting them here in case it helps others who fork this repo (not expecting any merges into the main branch):

1. Hardcoded user paths in PathUtils.scratch_dir (utils.py:38-42)

scratch_dir is hardcoded to /Users/Matteo/scratch/ (macOS) and /local/scratch/mb2389/ (Linux). Replacing this with an environment variable + fallback default works:

scratch_dir = Path(os.environ.get("HETGPPO_SCRATCH_DIR", Path.home() / ".hetgppo" / "scratch"))

Additionally, the _temp_dir passed to ray.init() should be removed, as on macOS, long scratch paths exceed the 103-byte AF_UNIX socket limit. Ray's default (/tmp/ray) works fine.

2. Pydantic v1/v2 incompatibility

Ray 2.1.0 uses pydantic.fields.ModelField, which was removed in pydantic v2. Pinning pydantic<2 resolves this.

3. RelVel transform (models/gppo.py:100)

torch_geometric.transforms.BaseTransform now requires a forward method instead of __call__. Renaming __call__ to forward fixes:

TypeError: Can't instantiate abstract class RelVel with abstract method forward

Tested with: Python 3.10, Ray 2.1.0, pydantic 1.10.26, torch-geometric 2.7+, wandb 0.25.1.