Installation#
TERRA is published on PyPI as terra-st (the import name is terra) and
requires an NVIDIA GPU. Install in two steps — PyTorch first, so it matches
your GPU, then TERRA. We recommend uv.
1. Install PyTorch for your hardware#
Install the PyTorch build that matches your GPU driver
before installing TERRA, so the correct CUDA wheel is used (otherwise a plain
install pulls the default wheel, which may not match your driver). Run
nvidia-smi and read the “CUDA Version” in the top-right, then install the
matching CUDA build — see the
official PyTorch install guide — e.g.:
uv pip install torch --index-url https://download.pytorch.org/whl/cu124
2. Install TERRA#
uv pip install terra-st
Plain pip install terra-st works too. Verify the install (PyTorch sees your
GPU, and TERRA imports):
python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
python -c "import terra; print(terra.__version__)"
The last value printed by the first command should be True — TERRA requires a
GPU, so torch.cuda.is_available() must report one. If it prints False,
PyTorch can’t see your GPU (usually a CUDA build that doesn’t match your driver
— revisit step 1).
Development install#
For a development install from a clone of the repository (after step 1 above):
git clone https://github.com/Lotfollahi-lab/terra.git
cd terra
uv pip install -e ".[dev,test,doc]"
Optional extras#
TERRA ships several optional dependency groups:
Extra |
Install |
Purpose |
|---|---|---|
|
|
Publish/download model bundles on the Hugging Face Hub ( |
|
|
JupyterLab + ipykernel to run the tutorial notebooks. |
|
|
Evaluation utilities (CellPhoneDB, Omnipath). |
|
|
Build the documentation. |
|
|
Run the test suite. |
Reproducible environment#
For the exact, fully-pinned environment TERRA is developed and tested against, use the committed lockfile with uv:
uv sync