How hi compute works

The ideas behind hi compute - providers, runs and instances, hardware names, names, and the limits that keep costs in check.

hi compute rents a machine from a provider, lets you use it as if it were local, and gives it back. It hides each provider’s own tool behind one set of commands, but it does not pretend they are the same: hardware names and prices stay the provider’s own.

Runs and instances

hi compute handles two kinds of work with the same commands:

hi compute serve is an instance that runs a model server for you. Runs and instances share ls, status, logs, stop, and the same limits.

Providers

hi compute providers    # which providers are signed in
hi compute hardware     # every provider's hardware, memory, and price

Which provider is used

hi picks the provider in this order:

1. --on colab, --on hf, or --on runpod on the command.
2. The HI_COMPUTE_PROVIDER environment variable.
3. The hardware name: --gpu G4 only exists on Colab, --gpu a10g-small only on Hugging Face, --gpu rtx-4090 only on RunPod.
4. The only provider you are signed in to.

If none of these decides, hi asks you to add --on.

Hardware names

Use the names from hi compute hardware. hi checks the name before calling the provider, because an unknown name can otherwise fall back to other hardware: the Colab CLI once turned an unknown GPU into an A100.

Names

Every run and instance has a name. Choose one with --name, or hi makes one from the hardware, such as t4-3fa1. Names use lowercase letters, digits, and hyphens. Refer to an instance by its name or by provider/name, and to a Hugging Face job also by its ID:

hi compute ssh box
hi compute ssh colab/box
hi compute logs hf/6ab97cba6b030d633f69a5ea

Limits and confirmation

Every run and instance has a maximum lifetime, --max, in hours (2) or with a unit (30m, 2d). Instances default to 4 hours and runs to 1 hour; Colab allows at most 24 hours. --max none means no limit, after a warning you must confirm. Paid hardware asks you to confirm before it starts. Costs, limits, and billing explains how each provider enforces the limit.

Add --dry-run to any up, run, or serve to see the exact provider command or API request without starting anything.

The guided menu

Run hi compute with no arguments in a terminal for a guided menu:

• A Running panel at the top shows your machines and how long each has left; no time limit is highlighted.
• Every provider is listed; choosing one you are not signed in to signs you in first (for RunPod, by asking for the API key).
• Choose with the arrow keys and Enter; the list stays put and only the marker moves. Long lists such as Hugging Face’s hardware scroll only at the edge. Press / and type to filter, for example /a10, and use number keys to jump in short lists.
• Text answers are checked as you type them, so a lifetime like soon is flagged before anything happens.
• Each answer stays on screen as a ✔ line, so you can see the choices you made.
• Before anything costs money, a summary card shows the provider, hardware, price, who pays, the time limit, and the equivalent command. The default answer is No. Machines without a time limit get an amber warning.
• Esc goes back a step; Ctrl+C leaves the menu.

Every action prints the command it runs, so you learn the flags as you go. When the input is not a terminal, the menu falls back to numbered questions.

Commands at a glance

The command reference lists every option.