Skip to content
Contact Support

Jupyter kernels - Tool-assisted management

Warning

NeSI OnDemand is in development and accessible to early access users only. If you are interested in helping us test it please contact us.

Introduction

Jupyter can execute code in different computing environments using kernels. Some kernels are provided by default (Python, R, etc.) but you may want to register your computing environment to use it in notebooks. For example, you may want to load a specific environment module in your kernel or use a Conda environment.

To register a Jupyter kernel, you can follow the steps highlighted in the Jupyter kernels - Manual management or use the nesi-add-kernel tool provided within the Jupyter on NeSI service. This page details the latter option, which we recommend.

Getting started

First you need to open a terminal. It can be from a session on Jupyter on NeSI or from a regular ssh connection on Mahuika login node. If you use the ssh option, make sure to load the JupyterLab module to have access to the nesi-add-kernel tool:

module purge  # remove all previously loaded modules
module load JupyterLab/2024.08.2-foss-2023a-4.2.4

Then, to list all available options, use the -h or --help options as follows:

nesi-add-kernel --help

Here is an example to add a TensorFlow kernel, using NeSI’s module:

nesi-add-kernel tf_kernel TensorFlow/2.8.2-gimkl-2022a-Python-3.10.5

and to share the kernel with other members of your NeSI project:

nesi-add-kernel --shared tf_kernel_shared TensorFlow/2.8.2-gimkl-2022a-Python-3.10.5 

To list all the installed kernels, use the following command:

jupyter-kernelspec list

and to delete a specific kernel:

jupyter-kernelspec remove <kernel_name>

where <kernel_name> stands for the name of the kernel to delete.

Conda environment

First, make sure the JupyterLab module is loaded:

module purge
module load JupyterLab/2024.08.2-foss-2023a-4.2.4

To add a Conda environment created using conda create -p <conda_env_path>, use:

nesi-add-kernel my_conda_env -p <conda_env_path>

otherwise if created using conda create -n <conda_env_name>, use:

nesi-add-kernel my_conda_env -n <conda_env_name>

Virtual environment

If you want to use a Python virtual environment, don’t forget to specify which Python module you used to create it.

For example, if we create a virtual environment named my_test_venv using Python 3.10.5:

module purge
module load Python/3.10.5-gimkl-2022a
python -m venv my_test_venv

to create the corresponding my_test_kernel kernel, we need to use the command:

module purge
module load JupyterLab/2024.08.2-foss-2023a-4.2.4
nesi-add-kernel my_test_kernel Python/3.10.5-gimkl-2022a --venv my_test_venv

Singularity container

Danger

This section has not been tested on NeSI OnDemand

To use a Singularity container, use the -c or --container options as follows:

module purge
module load JupyterLab/2024.08.2-foss-2023a-4.2.4
nesi-add-kernel my_test_kernel -c <container_image.sif>

where <container_image.sif> is a path to your container image.

Note that your container must have the ipykernel Python package installed in it to be able to work as a Jupyter kernel.

Additionally, you can use the --container-args option to pass more arguments to the singularity exec command used to instantiate the kernel.

Here is an example instantiating a NVIDIA NGC container as a kernel. First, we need to pull the container:

module purge
module load Singularity/3.11.3
singularity pull nvidia_tf.sif docker://nvcr.io/nvidia/tensorflow:21.07-tf2-py3

then we can instantiate the kernel, using the --nv singularity flag to ensure that the GPU will be found at runtime (assuming our Jupyter session has access to a GPU):

module purge
module load JupyterLab
nesi-add-kernel nvidia_tf -c nvidia_tf.sif --container-args "'--nv'"

Note that the double-quoting of --nv is needed to properly pass the options to singularity exec.