Installation¶
This page covers installing remedi and its dependencies.
Prerequisites¶
- Python ≥ 3.12
- A CUDA 12.6 capable GPU for training and for computing descriptors at scale (CPU works for small runs).
uvis recommended, since it resolves the pinned torch/CUDA and git dependencies for you.
Install with uv (recommended)¶
git clone https://github.com/molsuit/Rem3Di.git
cd Rem3Di
uv sync # core install
uv sync --extra eval # + downstream learners (LightGBM, scikit-learn)
uv sync uses the pins in pyproject.toml: the CUDA-12.6 torch wheels
(pytorch-cu126 index) and the torch-sim MACE backend from git
(torch-sim-atomistic).
Install with pip¶
pip install -e . # core
pip install -e ".[eval]" # + downstream learners
With pip you must install the git-sourced dependencies yourself:
pip install \
"torch-sim-atomistic @ git+https://github.com/TorchSim/torch-sim.git@main"
Verify¶
python - <<'PY'
from remedi.evaluation.evaluation_utils import (
evaluate_molecular_descriptor_on_dataset,
)
print("ok")
PY
Gotchas¶
- zarr v3 only. Datasets use
zarr>=3.2.polaris-libpinszarr<3and cannot share this environment; ingest Polaris data out-of-venv viascripts/dataset_download/dump_polaris.py. - numba. Pinned to
numba>=0.61(older versions fail to build on Python 3.12). - MACE weights. The model uses a MACE foundation model as its frontend. Point
the checkpoint's
mace_config.model_pathat a MACE model file on your machine (see Concepts). torch.loadweights-only error on import (e3nnconstants.pt, PyTorch ≥ 2.6): exportTORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1.
Next steps¶
- Quickstart: get descriptors in 5 minutes.