guide

Your first remote GPU

Five minutes from a fresh install to a shell on a rented GPU, and back again without leaving anything running.

This walk-through rents a small GPU, runs a command on it, and gives it back. It assumes you have installed hi and signed in to at least one provider.

1. Check your providers

hi compute providers

colab    ready
hf       ready

A provider that is not ready says what to do next, such as sign in with hi login hf.

2. Pick hardware

hi compute hardware

PROVIDER  HARDWARE    KIND  MEMORY      RATE
colab     cpu         CPU   12 GB RAM   ~0.08 units/h
colab     T4          GPU   15 GB VRAM  ~1.07 units/h
...
hf        t4-small    GPU   16 GB VRAM  $0.40/h
hf        a10g-small  GPU   24 GB VRAM  $1.00/h
...

Hardware names are the provider’s own. A T4 is T4 on Colab and t4-small on Hugging Face. The name you choose also picks the provider.

3. Preview, then start

--dry-run shows exactly what would happen and starts nothing:

hi compute up --gpu T4 --name first --max 30m --dry-run

Start colab/first on T4 (~1.07 units/h), stopping after 30m at 14:05.
Would run: colab new -s first --gpu T4

When it looks right, run it without --dry-run. hi asks you to confirm paid hardware:

hi compute up --gpu T4 --name first --max 30m

Start colab/first on T4 (~1.07 units/h), stopping after 30m at 14:05.
Start it? [y/N] y
...
first is up. Next:
  hi compute ssh first
  hi compute tunnel first 8000
  hi compute stop first

On Hugging Face, the same step is --gpu t4-small, and the start line also names who pays, as in ($0.40/h, billed to hifinab).

4. Use it

hi compute ssh first -- nvidia-smi     # run one command
hi compute ssh first                   # or open a shell; exit when done

On Hugging Face, shells need your SSH key registered on the Hub first. See Hugging Face Jobs.

5. Give it back

hi compute stop first
hi compute ls

Stopped first.
No instances are running.

Even if you forget, the machine stops itself at --max (30 minutes here).

Next steps

  • Run a training script instead of a shell: Run a job to completion
  • Reach a web app or notebook on the machine: Interactive machines
  • Try a large language model: Serve a model
  • Prefer menus? Run hi compute with no arguments for a guided menu that prints the equivalent command at every step.