Install

Install from PyPi

To get the latest release of Earth2Studio, install from the Python index:

pip install earth2studio

Install from Source

To install the latest main branch version of Earth2Studio:

git clone https://github.com/NVIDIA/earth2studio.git

cd earth2-inference-studio

pip install .

Verify installation using:

python

>>> import earth2studio
>>> earth2studio.__version__

Base Install Limitations

The base pip install does not guarantee all functionality and/or examples are operational due to optional dependencies. We encourage users that face package issues to familiarize themselves with the optional model installs and suggested environment set up for the most complete experience.

Datasource Dependencies

Some data sources require additional dependencies, libraries or specific Python versions to install. To install all dependencies

pip install earth2studio[data]

Model Dependencies

Some models require additional dependencies which are not installed by default. Use the optional install commands to add these dependencies.

CorrDiff

Notes: Additional dependencies for all CorrDiff models.

pip install earth2studio[corrdiff]

FengWu

Notes: Requires ONNX GPU Runtime. May need manual install depending on CUDA version.

pip install earth2studio[fengwu]

FuXi

Notes: Requires ONNX GPU Runtime. May need manual install depending on CUDA version.

pip install earth2studio[fuxi]

Pangu

Notes: Requires ONNX GPU Runtime. May need manual install depending on CUDA version.

pip install earth2studio[pangu]

SFNO

Notes: Requires Modulus-Makani to be installed manually.

pip install "makani[all] @ git+https://github.com/NVIDIA/modulus-makani.git@v0.1.0"
pip install earth2studio[sfno]

Using pip install earth2studio[all] will install all optional functionality dependencies.

Environments

For the best experience we recommend creating a fresh environment whether that be using a Docker container or a Conda environment. Below are some recipes for creating a handful of environments that we recommend for setting up Earth2Studio to run all build in models.

Modulus Docker Container

The recommended environment to run Earth2Studio in is the Modulus docker container. This is the environment the team develops with and is the primary test bed. You can install Earth2Studio in a running container directly:

docker run -i -t nvcr.io/nvidia/modulus/modulus:24.12

>>> pip install "makani[all] @ git+https://github.com/NVIDIA/modulus-makani.git@v0.1.0"

>>> pip install earth2studio[all]

or build your own Earth2Studio container using a Dockerfile:

FROM nvcr.io/nvidia/modulus/modulus

RUN pip install "makani[all] @ git+https://github.com/NVIDIA/modulus-makani.git@v0.1.0"

RUN pip install earth2studio[all]

PyTorch Docker Container

Modulus docker container is shipped with some packages that are not directly needed by Earth2Studio. Thus, some may prefer to install from the PyTorch container. Note that for ONNX models to work we will need a specific install:

docker run -i -t nvcr.io/nvidia/pytorch:24.11-py3

>>> pip install "makani[all] @ git+https://github.com/NVIDIA/modulus-makani.git@v0.1.0"

>>> pip install earth2studio[all]

Conda Environment

For instances where Docker is not an option, we recommend creating a Conda environment. Ensuring the PyTorch is running on your GPU is essential, make sure you are installing the correct PyTorch for your hardware and CUDA is accessible.

conda create -n earth2studio python=3.11
conda activate earth2studio

conda install pytorch pytorch-cuda=12.1 -c pytorch -c nvidia
conda install eccodes python-eccodes -c conda-forge
pip install "makani[all] @ git+https://github.com/NVIDIA/modulus-makani.git@v0.1.0"

pip install earth2studio[all]

Configuration

Earth2Studio uses a few environment variables to configure various parts of the package. The import ones are:

• EARTH2STUDIO_CACHE: The location of the cache used for Earth2Studio. This is a file path where things like models and cached data from data sources will be stored.

• EARTH2STUDIO_PACKAGE_TIMEOUT: The max number of seconds for a download operation of a model package file from a remote store such as NGC, Huggingface or S3.