Installation

cBottle has been tested on these software configurations - Linux OS (Ubuntu 22.04) - CUDA 12 - pytorch 2.6 and 2.7 - earth2grid v2025.6.1, v2025.7.1

Inference has been tested on - Data center GPUS: H100, A100 - Desktop GPU: RTX 6000 Ada Generation

Training has been tested on - H100, A100

It typically takes about 10 minutes to install earth2grid from scratch on a desktop computer, assuming that CUDA has been installed. 5 of that is to build earth2grid from source. This could be avoided if there are binaries available for your platform (see instructions).

PIP

If necessary install pytorch. Then, we need to install earth2grid

pip install torch # if necessary
pip install setuptools hatchling
pip install --no-build-isolation https://github.com/NVlabs/earth2grid/archive/main.tar.gz

Warning: make sure to include --no-build-isolation above to avoid building earth2grid against the wrong version of pytorch. This will cause very confusing error messages during runtime.

Now, install cbottle

pip install -e .

NERSC (instructions)

Install the pre-requisites

module load pytorch/2.6.0
python3 -m venv --system-site-packages cbottle-env
source cbottle-env/bin/activate
CC=gcc CXX=g++ pip install --no-build-isolation https://github.com/NVlabs/earth2grid/archive/main.tar.gz

Then clone cBottle and install some other dependencies

git clone https://github.com/NVlabs/cBottle
cd cBottle
pip install -e .