JupyterLab on a rented GPU

Start a GPU machine, run JupyterLab on it, and open it in your laptop's browser through a private tunnel.

This starts a machine, installs and starts JupyterLab on it, and forwards it to your browser. JupyterLab listens only on the machine’s localhost and is only reachable through your tunnel.

1. Start a machine

hi compute up --gpu T4 --name lab --max 3h

On Hugging Face, use --gpu t4-small and make sure your SSH key is on the Hub.

2. Start JupyterLab on it

hi compute ssh lab -- 'pip install -q jupyterlab &&
  (setsid nohup jupyter lab --no-browser --allow-root \
     --ip 127.0.0.1 --port 8888 --IdentityProvider.token="" \
     > /tmp/jupyter.log 2>&1 < /dev/null &) && sleep 5 && tail -3 /tmp/jupyter.log'

• setsid nohup … & keeps JupyterLab running after the SSH command returns.
• --ip 127.0.0.1 keeps it private to the machine; the tunnel is the only way in, so it runs without a token.

3. Open it

hi compute tunnel lab 8888:18888

Open http://127.0.0.1:18888/lab. Using a local port other than 8888 avoids clashing with a Jupyter already running on your laptop.

Check the GPU from a notebook cell:

!nvidia-smi

4. Keep your work

Files live on the machine’s disk and disappear when it stops. Download notebooks from JupyterLab’s file browser, or copy them with scp through the SSH config entry on Colab.

5. Stop

Press Ctrl+C in the tunnel terminal and answer y, or:

hi compute stop lab

The tunnel closes by itself when the machine stops.