Use Conda on Any PU HPC System

Initial Setup

The initial setup need only be run once to begin using Conda environments on a PU HPC system.

You may also repeat these steps if you’d like to start over as if it is your first time using a Conda environment on a PU HPC system.

1. Open a Terminal on the remote system and use the module command to load Anaconda. The module command activates one of many software installations available on a given system.

    module load anaconda3/2024.2
   

2. Remove any existing conda settings, environments, and related data from the Terminal. With this example command, we move the .condarc file and the .conda folder to condarc-backup and conda-backup in $HOME. Dispose of them to free up quota once you have a working environment.

    PU_TIMESTMP=(`date '+%b_%d_%Y_%H%M%S'`); [ -d $HOME/.conda ] && mv $HOME/.conda $HOME/conda-backup-$PU_TIMESTMP; [ -f $HOME/.condarc ] && mv $HOME/.condarc $HOME/condarc-backup-$PU_TIMESTMP
   

3. Configure conda to store both its package cache and environments in a location other than your $HOME.¹ In addition, prioritize the conda-forge channel for software installations by default. With this example command, we write the necessary configuration changes to a new $HOME/.condarc file.

    PU_HPC_SYS=(`echo $HOSTNAME | tr -d '0123456789'`); if [ $PU_HPC_SYS = 'adroit' ]; then SCRATCH_LOC='network' ; else SCRATCH_LOC='gpfs' ; fi; echo -e "pkgs_dirs:\n - /scratch/$SCRATCH_LOC/$USER/.conda/pkgs\nenvs_dirs:\n - /scratch/$SCRATCH_LOC/$USER/.conda/envs" > $HOME/.condarc ; conda config --add channels conda-forge ; mkdir -p $HOME/.conda /scratch/$SCRATCH_LOC/$USER/.conda/{pkgs,envs} ; touch $HOME/.conda/environments.txt ; ln -s /scratch/$SCRATCH_LOC/$USER/.conda/{pkgs,envs} $HOME/.conda
   

Creating a New Environment

1. Use the conda command to create a new environment. For this example, we will name the environment “torch-env”, indicative of PyTorch development. We’ll install ipykernel inside for now in order to be able to use the custom environment with Jupyter.

    conda create -n torch-env ipykernel -c conda-forge -y
   

   While there are good references for getting started with Pytorch on the HPC Clusters, the process can be problematic. To both demonstrate the flexibility of the conda command and avoid some trouble, we will create the environment, activate the environment, and install additional packages via both conda and pip safely and reproducibly.

2. Activate the environment you just created. If you’ve done so successfully, the active environment name will be displayed before your prompt.

    conda activate torch-env
   

3. Use the conda command to install PyTorch. Note that the command recommended for the HPC environment is prefixed by manually setting a value for the CONDA_OVERRIDE_CUDA variable, and uses the constraint pytorch==2.0*==cuda11*. This is to allow the installation to properly complete in the absence of a GPU on the head node, but still work in the presence of a GPU on the nodes themselves.

    CONDA_OVERRIDE_CUDA="11.2" conda install "pytorch==2.0*=cuda11*" torchvision -c conda-forge -y
   

4. Install other useful packages via pip. When pip is managed by conda, as it is in our torch-env environment, conda is aware of any pip-installed packages.

    pip install scikit-learn
    pip install quantile_forest
    pip install matplotlib
    pip install seaborn
   

Removing an Environment

The Initial Setup can be used to re-initialize your entire conda configuration, but that may be too aggressive. To remove an existing environment, for instance, torch-env, pass its name to the conda command:

conda env remove --name torch-env

Installing New Packages into Existing Environments

When working with environments via the Terminal, always load the Anaconda module, which then makes the conda command available.

module load anaconda3/2024.2

Once the conda command is available, you can proceed to activate the environment by name.

conda activate torch-env

Add new software to the active environment using either conda or pip. When in doubt, favor conda, but don’t discount pip.

conda install pandas