Installing

Python Version

Python (3.9 to 3.12)

We require at least Python 3.9, and currently do not support >3.12. This is due to an issue with one of the required dependencies.

Installation

Environments

We currently do not provide a conda build of anemoi-training so the suggested installation is through Python virtual environments.

For linux the process to make and use a venv is as follows,

python -m venv /path/to/my/venv
source /path/to/my/venv/bin/activate

Instructions

To install the package, you can use the following command:

python -m pip install anemoi-training

We also maintain other dependency sets for different subsets of functionality:

python -m pip install "anemoi-training[profile]" # Install optional dependencies for profiling gpu usage
python -m pip install "anemoi-training[docs]" # Install optional dependencies for generating docs

Warning

After installation, verify that your PyTorch build has CUDA or ROCm support:

If this returns False, GPU training will not work. A common cause is installing the default torch wheel, which on aarch64 systems does not include CUDA support. In that case, reinstall PyTorch with the appropriate index URL, for example:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126

For full instructions on installing PyTorch with CUDA support, see the PyTorch installation guide.

Development versions

To install the most recent development version, install from github:

$ python -m pip install git+https://github.com/ecmwf/anemoi-core.git#subdirectory=training

Testing

To run the test suite after installing anemoi-training, install (via pypi) py.test and run pytest in the training directory of the anemoi-core repository.