Building an Apptainer Container with Python 3.9 and PyTorch (pip)

This guide explains how to create an Apptainer container using Python 3.9 and install PyTorch using pip.

Requirements

• Apptainer installed on your system (apptainer --version)
• Internet access
• No sudo rights required (supports --fakeroot or --sandbox builds)

Create the Definition File

Create a file called pytorch.def with the following content:

Bootstrap: docker
From: python:3.9-slim

%post
    # Update and install dependencies
    apt-get update && apt-get install -y --no-install-recommends \
        build-essential \
        libglib2.0-0 libxext6 libsm6 libxrender1 \
        git wget curl ca-certificates && \
        apt-get clean && rm -rf /var/lib/apt/lists/*

    # Install PyTorch
    pip install --upgrade pip
    # GPU version
    pip install torch torchvision torchaudio
    # ...or CPU version
    pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

    # Optional: add other Python tools
    pip install jupyterlab matplotlib scikit-learn

%environment
    export PYTHONUNBUFFERED=1

%runscript
    exec python3 "$@"

Note: select either the GPU version of the CPU version and remove the other line in the script above

Build the Container

apptainer build --fakeroot pytorch.sif pytorch.def

Use the Container

Use the --nv flag when running or entering the container to enable the use of the GPU in the container. Make sure you activate the CUDA enviroment on the node before using the GPU container:

module load cuda

Enter a shell

apptainer shell --nv pytorch.sif

Note: you can leave out the --nv option if you are not using the GPU

Run Python

python3

Or execute scripts directly with apptainer

apptainer exec --nv pytorch.sif python3 myscript.py

Test PyTorch

Inside the container:

CPU and GPU

python3 -c "import torch; print(torch.__version__); print(torch.rand(2,3))"

GPU

python3 -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"

If everything is set up correctly, this should return True and the name of your GPU.