Skip to content

Argonne Leadership Computing Facility

Python

We provide prebuilt conda environments containing GPU-supported builds of torch, tensorflow (both with horovod support for multi-node calculations), jax, and many other commonly-used Python modules.

Users can activate this environment by first loading the conda module, and then activating the base environment.

Explicitly (either from an interactive job, or inside a job script):

module load conda ; conda activate base

This will load and activate the base environment.

Virtual environments via venv

To install additional packages that are missing from the base environment, we can build a venv on top of it.

Conda base environment + venv

If you need a package that is not already installed in the base environment, this is generally the recommended approach.

We can create a venv on top of the base Anaconda environment (with --system-site-packaes to inherit the base packaes):

module load conda; conda activate
VENV_DIR="venvs/polaris"
mkdir -p "${VENV_DIR}"
python -m venv "${VENV_DIR}" --system-site-packages
source "${VENV_DIR}/bin/activate"

You can always retroactively change the --system-site-packages flag state for this virtual environment by editing ${VENV_DIR}/pyvenv.cfg and changing the value of the line include-system-site-packages=false.

To install a different version of a package that is already installed in the base environment, you can use:

python3 pip install --ignore-installed <package> # or -I

The shared base environment is not writable, so it is impossible to remove or uninstall packages from it. The packages installed with the above pip command should shadow those installed in the base environment.

Cloning the base Anaconda environment

Warning

This approach is generally not recommended as it can be quite slow and can use significant storage space.

If you need more flexibility, you can clone the conda environment into a custom path, which would then allow for root-like installations via conda install<module> or pip install <module>.

Unlike the venv approach, using a cloned Anaconda environment requires you to copy the entirety of the base environment, which can use significant storage space.

To clone the base environment:

module load conda ; conda activate base
conda create --clone base --prefix /path/to/envs/base-clone
conda activate /path/to/envs/base-clone

where, path/to/envs/base-clone should be replaced by a suitably chosen path.

Note: The cloning process can be quite slow.

Danger

This is typically not recommended.

With the conda environment setup, one can install common Python modules using python3 pip install --users '<module-name>' which will install packages in $PYTHONUSERBASE/lib/pythonX.Y/site-packages.

The $PYTHONUSERBASE environment variable is automatically set when you load the base conda module, and is equal to /home/$USER/.local/polaris/conda/YYYY-MM-DD.

Note, Python modules installed this way that contain command line binaries will not have those binaries automatically added to the shell's $PATH. To manually add the path:

export PATH="$PYTHONUSERBASE/bin:$PATH"

Be sure to remove this location from $PATH if you deactivate the base Anaconda environment or unload the module.

Cloning the Anaconda environment, or using venv are both more flexible and transparent when compared to --user installs.