”Spin up a Jupyter Lab environment on an RTX 4090.”
The agent handles the full setup:list_gpu_types: confirms RTX 4090 availability and its live hourly ratecreate_instance: deploysjupyter-devwith RTX 4090instance_status: polls until runningssh_execute: installs JupyterLab and common ML packages:ssh_execute: starts JupyterLab on the instance’s loopback address, with a token:ssh_execute:sleep 5; jupyter server listto read the address and its token
JupyterLab is running on jupyter-dev. It only listens inside the instance, so open it through an SSH tunnel. Run this on your computer and leave it open:Then open:Specs: RTX 4090 (24 GB VRAM), billed by the second on most GPUs while running Pre-installed packages: PyTorch, torchvision, NumPy, Pandas, Matplotlib, scikit-learn.
JupyterLab listens on 127.0.0.1 and asks for its token, so nobody on the internet can reach it. The tunnel uses the Runcrate CLI. If the instance has your own SSH key, add
-N -L 8888:localhost:8888 to the SSH command on its page instead.”Install some extra packages on my Jupyter instance: I need HuggingFace and Weights & Biases.”
The agent installs without interrupting the running server:ssh_execute:pip install transformers datasets accelerate wandbssh_execute:python -c "import transformers; print(transformers.__version__)"to verify
Installed on jupyter-dev:All available in your running JupyterLab session. Restart the kernel if you had a notebook open.
- transformers 4.44.0
- datasets 3.0.1
- accelerate 0.34.0
- wandb 0.18.0
”I’m done with the Jupyter instance. Download my notebook and shut it down.”
The agent saves your work and cleans up:ssh_execute:ls /root/*.ipynb /workspace/*.ipynb 2>/dev/nullto find notebooksfile_download: downloads/workspace/experiment.ipynbbilling_usage: checks how much the session costdelete_instance: terminatesjupyter-dev
Downloaded experiment.ipynb (142 KB). Session summary:Instance jupyter-dev terminated. Billing stopped.
- Duration: 3.5 hours
- GPU: RTX 4090 (the exact charge is in
billing_usage)