JupyterLab

This is a simple guide to launch a JupyterLab instance from a GPU node.

To execute JupyterLab instances on a GPU node you have to first create a Conda environment following these instructions. Note: you only have to do this once.

# Create a conda environment
module load anaconda
conda create --name lab python=3.12
conda activate lab

# Install dependencies
conda install -y conda-forge::tensorflow-gpu conda-forge::cuda conda-forge::cudnn
pip install jupyterlab ipykernel

To launch your JupyterLab instance first allocate a GPU job:

# Launch a Slurm job on a GPU node
srun -A <slurm_account> -p gpu -N 1 --gpus=1 --pty bash

Now launch your JupyterLab instance:

# Launch Jupyter
module load anaconda
conda activate lab
jupyter lab --no-browser --ip=$(host $(hostname) | awk '/10.119.12/ {print $4}')

You should see something like this.

    To access the server, open this file in a browser:
        file:///home/c3_username/.local/share/jupyter/runtime/jpserver-619173-open.html
    Or copy and paste one of these URLs:
        http://10.119.12.71:8888/lab?token=58e3faf95c177d7e5e3f357b79ec7334b343930222067b81
        http://127.0.0.1:8888/lab?token=58e3faf95c177d7e5e3f357b79ec7334b343930222067b81

Now you can access your Jupyter instance from a web browser by following the first http URL.

Allowed Ports

Whenever you launch Jupyter without specifying a port it will try to use port 8888 by default. However, if 8888 is already in use it will use something else. Since we only allow ports 8888 through 8897, you might have to specify it manually, otherwise you won't be able to connect to your Jupyter instance easily:

jupyter lab --no-browser --ip=$(host $(hostname) | awk '/10.119.12/ {print $4}') --port=<port>