Launch ipykernel in VSCode¶

VSCode can start ipykernel (and run Jupyter notebook) from a Singularity container.

Here is a simple example setup.

Build or pull container that has ipykernel installed. Note: You do not need to install jupyter into the Python environment you want to use. Only the IPyKernel package is required to launch a Python process as a kernel and execute code against your notebook (pip install ipykernel). Visit the Jupyter extension wiki to learn more.

apptainer pull docker://nvcr.io/nvidia/pytorch:22.03-py3

2. Create ~/.local/share/jupyter/kernels/singularity-kernel01/kernel.json file with content like this on the computer where the container is. Make sure you match the location of the container. Below is a working example for Pelle.

{
  "argv": [
    "apptainer", "exec",
    "--bind", "/crex,/gorilla,/proj,/sw,/run/user","--nv",
    "/proj/sinapp/nobackup/00.workshop/pytorch_22.03-py3.sif",
    "python", "-m", "ipykernel_launcher", "-f", "{connection_file}"
  ],
  "display_name": "Pytorch_22.03",
  "language": "python"
}

3. From VSCode connect via remote shell to Pelle.
4. Open a Jupyter notebook or create a new one.
5. "Select Kernel", then "Jupyter Kernel"... vscode-01

6. Select from the predefined Jupyter kernels.. vscode-02

7. Run...

Note

On Pelle, users are not allowed to ssh to the compute node where the job is running, so this approach will not work to start kernel on a compute node. Instead,
1. Start jupyter with jupyter --ip 0.0.0.0 --no-browser in an interactive session
2. From VSCode remote session on Pelle, chose "Existing Jupyter Server" when selecting a kernel.
Make sure, you point to the correct address which includes the node address, for example http://p2033.uppmax.uu.se:8888/lab?token=28026d8d41adccf863f01c471260c7fc5b20b6441872a79a