This is an old revision of the document!
How-to create a simple apptainer container
Consider the case where you need to create a working python3 distrib with pytorch installed for running a script using GPU resources on ChaCha or DISCO. On those machines, we use Apptainer for the containers and this documentation describes how to create, customize and run images using Apptainer. Launching Apptainer tasks using Slurm is defined elsewhere.
Defining the content of the container
Start by defining a container definition, such as in rules.def (this correponds to a Dockerfile file):
Bootstrap: docker From: nvidia/cuda:12.6.2-base-ubuntu22.04 %help This container provides a CUDA environment, with Python and Jupyternotebook. %labels AUTHOR_NAME Pierre-André Mudry VERSION 1.0 %environment %post -c /bin/bash export DEBIAN_FRONTEND=noninteractive apt-get -y update apt-get -y install git python3-pip python3-dev python3-opencv libglib2.0-0 %files requirements.txt requirements.txt %post -c /bin/bash python3 -m pip install -r requirements.txt python3 -m pip install --upgrade pip pip3 install torch torchvision torchaudio -f https://download.pytorch.org/whl/cu111/torch_stable.html
This definition files is of a Debian distrib with CUDA, with python3 and pytorch installed and has other requirements as well. Those requirements are copied from you current directory requirements.txt file using the %files directive. In this example, we use for requirements.txt:
warp-lang usd-core matplotlib pyglet jupyter python-language-server
Of course, you can tailor this to fit your needs.
Building the image
Once the definition is made, you can build the test.sif image with apptainer build test.sif rules.def. The image is created in your own directory. By default, when you run the image, it has access to the host filesystem. To run the image interactively (or not), you then simply run
Running the image
apptainer run --nv ./test.sif
The --nv flag must be used to use NVidia resources. If your run this script interactively, you then have access to the virtual environment. You can launch for instance nvidia-smi or jupyter notebook --ip=0.0.0.0 --port=8888 --no-browser --allow-root to create a Juypter notebook system.
