jupyter-pytorch-full

⚠️ WARNING ⚠️ This is very much alpha software! This has yet to be tested in a real deployment and serves as a proof-of-concept

This ROCK is a recreation of the kubeflownotebookswg/jupyter-pytorch-full image using rockcraft.

This image is used as part of the Charmed Kubeflow product. It is deployed in a user namespace, and enables users to perform CPU-accelerated tasks using PyTorch from within a Jupyter Notebook.

Both PyTorch and Jupyter are installed in Conda environment, which is automatically activated.

Quick start

To get started with this image, you can:

docker run --rm ghcr.io/canonical/jupyter-pytorch-full:latest

Build references

Using the Makefile for the original image, I was able to recreate the image in a few stages. The upstream image is built from the following layers:

• base
• jupyter
• jupyter-pytorch
• jupyter-pytorch-full

In this version, the last three layers are combined into a single part in rockcraft.yaml named conda-jupyter.

How to build

# Install some deps
$ sudo snap install rockcraft --classic --edge
$ sudo snap install skopeo --edge --devmode
$ sudo snap install yq

# Pack the ROCK (optionally add --verbose)
$ rockcraft pack
$ version=$(yq -r '.version' rockcraft.yaml)

# For the following commands, you must either use `sudo` or ensure your user is a member of the
# `docker` group, or the commands will fail.

# Import the ROCK into the local docker image cache
$ sudo skopeo --insecure-policy copy \
    oci-archive:"jupyter-pytorch-full_${version}_amd64.rock" \
    docker-daemon:"canonical/jupyter-pytorch-full:${version}"

# Run the image, invoking `jupyter` through pebble
$ sudo docker run --rm "canonical/jupyter-pytorch-full:${version}"

TODO

• Test with a real Charmed Kubeflow
• Upgrade to Ubuntu 22.04