Installation¶

Prerequisites¶

Python 3.10+
PyTorch >= 2.0 (with CUDA support for GPU training)
PyTorch Geometric >= 2.7.0

Install PyTorch first

PyTorch and PyTorch Geometric must be installed before installing LUMINA, as they are not auto-installed by pip.

Step 1: Install PyTorch¶

Follow the official PyTorch installation guide for your platform and CUDA version.

# Example: PyTorch with CUDA 12.8
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128

Step 2: Install PyTorch Geometric¶

pip install torch-geometric

Step 3: Install LUMINA¶

Development install (recommended)¶

git clone https://github.com/argonne-gridfm/lumina-sdk.git
cd lumina-sdk
make dev

This installs LUMINA in editable mode (pip install -e .).

Minimum supported install includes [acopf]

The base install above is not sufficient to load any OPF dataset — lumina/dataset/opf/utils.py unconditionally imports pandapower, which only ships with the [acopf] extra. Importing OPFDataset will raise ModuleNotFoundError: pandapower otherwise.

For most workflows the minimum supported install is:

pip install -e ".[acopf]"
# or equivalently:
make install-acopf

With optional dependencies¶

# Testing
make install-test       # or: pip install -e ".[test]"

# Required for OPFDataset loading + ACOPF evaluation (pandapower, pypower)
make install-acopf      # or: pip install -e ".[acopf]"

# Hyperparameter search (wandb, optuna)
make install-hps        # or: pip install -e ".[hps]"

# Documentation
pip install -e ".[doc]"

# Everything
make install-all        # or: pip install -e ".[test,acopf,hps]"

Verify Installation¶

import lumina
from lumina.dataset.opf.opf_dataset import OPFDataset
from lumina.model.opf.losses import OPFLossManager

print(f"LUMINA installed successfully! Version {lumina.__version__}")

Check detected versions:

make info

HPC Systems¶

For Polaris or Perlmutter, see the HPC Training guide for system-specific setup instructions.