Google Colab

Set up Colab for hi compute - the Colab CLI, sign-in, plan requirements, SSH keys - and the Colab-specific rules worth knowing.

hi drives Colab through Google’s Colab CLI, because Colab has no public API. You pay with the compute units of your Colab plan.

Set it up

1. Install the Colab CLI

hi install installs it. On a machine where you only installed hi:

uv tool install google-colab-cli

2. Sign in

hi login colab

This opens a browser for Google sign-in and asks you to paste a code back, then shows your balance. Sign in with the Google account that has the Colab plan. The token is kept by the Colab CLI in ~/.config/colab-cli/.

3. Check your plan

• SSH, and so hi compute ssh, tunnel, and serve, needs Colab Pro or Pro+. The free tier forbids SSH.
• Which GPUs you can get depends on your plan and on availability. An H100 is often unavailable.

4. Make sure you have an SSH key

hi compute ssh and tunnel use ~/.ssh/id_ed25519 or ~/.ssh/id_ecdsa. Colab accepts the key automatically; there is nothing to register.

ls ~/.ssh/id_ed25519 || ssh-keygen -t ed25519

5. Check

hi compute providers
hi compute billing

colab    ready

Colab balance: 1795.89 compute units, currently using 0.00 units/h

Hardware

--high-mem requests a high-RAM shape for cpu, T4, and A100.

Rules worth knowing

• 24 hours at most. Colab ends sessions after 24 hours; --max cannot be longer.
• Port 8080 is Colab’s. Colab’s own proxy uses it on every machine, so serve on another port, such as 8000, and tunnel that.
• No public web services. Colab’s terms forbid them, so hi only ever forwards ports to 127.0.0.1 on your laptop.
• The disk is slow for random reads. It is network storage. Keep lookup-heavy data in RAM; a G4 has 176 GB of it.
• No custom images. A Colab machine is a notebook runtime with CUDA 12.8 and Python preinstalled. serve builds llama.cpp on it the first time, which takes about two minutes on a G4.

How limits work on Colab

A Colab runtime cannot stop itself, so hi compute up starts a small background process on your laptop (hi compute __watch) that stops the machine at --max. It checks the wall clock every 30 seconds, so it also works after your laptop wakes from sleep. After a reboot it is gone, which is why every hi compute ls also stops machines past their limit.

Runs on Colab

hi compute run on Colab uses colab run: a fresh runtime runs one Python script and is released. Runs cannot detach, and there is no secret store, so --detach and --secret are refused. For long work, start an instance and run the script inside it over hi compute ssh.

Troubleshooting