Skip to main content
Run JupyterLab on a dedicated cloud GPU. Train models, explore datasets, and prototype backed by an RTX 4090, A100, or H100: without installing CUDA locally.

1. Deploy a devbox

2. Install and start JupyterLab

JupyterLab listens on the instance’s loopback address only, so nothing on the internet can reach it. It asks for a token, which it makes when it starts.
Print its address with the token:
It prints a line like this:
Copy the address before ::.

3. Open it through an SSH tunnel

Forward the port to your computer and leave this running:
Open http://127.0.0.1:8888/?token=<TOKEN> in your browser. Press Ctrl+C to close the tunnel; JupyterLab keeps running on the instance. If you launched the instance with your own SSH key, the SSH command on its page works too: add -N -L 8888:localhost:8888 to it, as in port forwarding. If port 8888 is busy on your computer, forward another one, such as -L 8889:localhost:8888, and open http://127.0.0.1:8889/?token=<TOKEN>.

4. Verify GPU access

In a notebook cell:

5. Persist notebooks with a volume

Notebooks saved to /workspace/ persist across deploys.

Upload data

Install additional packages

Tips

  • Use the ubuntu-devbox template: it includes CUDA, cuDNN, and Python.
  • Save notebooks to /workspace/ when using a volume for persistence.
  • For long training runs, use nohup in a terminal tab so the job survives browser disconnects.
  • To use a password instead of a token, open a shell with runcrate ssh jupyter, run jupyter lab password, then pkill -f jupyter-lab, then exit. From your computer, start JupyterLab again with the command in step 2. Keep --ip=127.0.0.1.

Cleanup