brightlab.Made for curious minds
Marimo labs / H100 setup

Optional extension · Teacher setup

Take a lab to an H100.

Use the same interactive notebook on a GPU machine you control.

For students: the hosted labs already work on your browser CPU. Start there. An H100 is useful for exploring larger workloads, and requires a teacher’s or institution’s GPU access.

For teachers: Brightlab supplies the notebook and launch bundle. It does not provision an H100 or start a paid session. Use your school’s existing server or approved cloud provider.

  1. Prepare the GPU host

    Use a Linux H100 host with SSH access, Docker, the NVIDIA Container Toolkit and a compatible NVIDIA driver. Ask the administrator to confirm support for the NVIDIA PyTorch 25.11 container. Run nvidia-smi and check that the GPU name contains H100.

    NVIDIA container requirements ↗ · Container Toolkit setup ↗

  2. Download and upload a lab

    Open any Marimo lab and choose H100 bundle. Unzip it and upload the extracted directory to the GPU machine. Each bundle contains a self-contained notebook, launch script and detailed README. Replace USER and GPU_HOST below with the details supplied by your administrator.

    scp -r extracted-lab USER@GPU_HOST:~/brightlab-lab

  3. Start the notebook

    SSH to the GPU host, enter the uploaded directory, and start the container. Keep this terminal open. Marimo prints an access token; authentication remains enabled.

    ssh USER@GPU_HOST
cd ~/brightlab-lab
bash run-h100.sh

  4. Connect from your computer

    Open a second terminal on your computer and forward the notebook port. Include any SSH identity file or custom port your provider requires. Open http://localhost:2718 in your browser and enter the token from the GPU terminal.

    ssh -N -L 2718:127.0.0.1:2718 USER@GPU_HOST

    The notebook port stays private to the host. For a whole class, ask your administrator for authenticated, individual sessions and GPU quotas.

  5. Run and verify

    The top of the notebook is the small CPU investigation. Scroll to Optional: run on an H100, select H100 CUDA, keep workload scale at 1, then choose Run the larger experiment.

    The notebook checks CUDA availability, confirms the device is an H100 and reads back a known matrix result. A successful run shows the actual GPU name, elapsed time, peak CUDA memory, a chart, a table and an evidence download. It reports an error if the requested hardware is unavailable.

    The larger experiment has its own parameter and workload scale. The second classroom control does not carry over. Repeated fixtures measure computation; they do not create new independent evidence. Small workloads can be faster on a CPU.

  6. Save and stop

    Download your evidence. Press Ctrl+C in the GPU terminal to stop the notebook, close the SSH tunnel, then stop or delete a rented GPU instance in the provider’s console. Closing a notebook does not stop cloud billing.

Already have a Python environment or a notebook service?

Download the Python notebook from the lab page. On a host with CUDA-compatible PyTorch and Marimo 0.24.0 installed, run:

marimo edit notebook.py --host 127.0.0.1 --port 2718 --headless

Use the same SSH tunnel. A service that supports Marimo Python imports can open this self-contained file; choose its H100 runtime if one is offered. The notebook checks the actual device. GPU availability, accounts and pricing depend on your provider.

The classroom calculations and larger CPU experiments have been checked locally. H100 execution must be verified on your host; an H100 was not available during authoring. Marimo deployment reference ↗