Pangu Weather DCU AI4S

Pangu-Weather inference and training optimizations for the AI4S DCU/ROCm competition environment.

Installation

Install the competition runtime first. torch, onescience and apex must match the provided DCU/ROCm environment, so they are deliberately not declared as automatic PyPI dependencies.

pip install EASONEASON

The lightweight Python dependencies (numpy, h5py, tqdm) install automatically. The model checkpoint and ERA5 data are not bundled.

Running

Run in a directory that contains conf/config.yaml and whose relative data and checkpoint paths match that file:

pangu-weather-infer
# or
python -m inference

The package also installs pangu-weather-train.

The default configuration is included in the distribution under pangu_weather_dcu/conf/config.yaml; copy it into the competition workspace before editing paths. Model weights (*.pth) and generated data are excluded from the release on purpose.

Publishing a release

1. Replace the package name and homepage in pyproject.toml if necessary. PyPI names are global: a name already taken cannot be published.

2. Create a new version for every upload (for example, change 0.1.0 to 0.1.1). PyPI does not allow replacing an uploaded file.

3. Build and check the artifacts:

   python -m pip install --upgrade build twine
   python -m build
   python -m twine check dist/*
   

4. Test on TestPyPI first, then publish to PyPI using a scoped API token:

   python -m twine upload --repository testpypi dist/*
   python -m twine upload dist/*
   

Do not put tokens, checkpoints, or competition datasets in this repository or the distribution.