Interactive machines

Start a machine that stays up, open shells on it, forward its ports to your laptop, connect VS Code, and stop it.

An instance is a machine that stays up until you stop it or it reaches its maximum lifetime. Use one for shells, notebooks, debugging, and anything you want to reach from your laptop.

Start one

hi compute up --gpu T4 --name box --max 2h

Option	Meaning
`--gpu <hardware>`	Hardware from `hi compute hardware`; default is the cheapest CPU
`--name <name>`	Name to use in later commands; generated if omitted
`--max <duration>`	Stop after this long: hours (`2`, `1.5`), `30m`, `2d`, or `none`; default `4h`, at most `24h` on Colab
`--image <image>`	Container image (Hugging Face); default `python:3.12` on CPU and `pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel` on GPU
`--high-mem`	High-RAM machine for `cpu`, `T4`, or `A100` (Colab Pro)
`--namespace <ns>`	Bill this Hugging Face account
`--data <org>/<name>`	Your team’s data on the machine through the hi server (RunPod, Shadeform); see On a machine you SSH into

Open a shell

hi compute ssh box                       # interactive shell
hi compute ssh box -- nvidia-smi         # run one command and return
hi compute ssh box -- 'cd /tmp && ls'    # quote commands with shell syntax

Shells use your SSH key, ~/.ssh/id_ed25519 (or id_ecdsa). Create one with ssh-keygen -t ed25519 if you have none. On Hugging Face, the public key must also be registered on the Hub.

Forward a port

tunnel makes a port on the machine reachable on your laptop, on 127.0.0.1 only:

hi compute tunnel box 8000          # the machine's 8000 on 127.0.0.1:8000
hi compute tunnel box 8888:18888    # the machine's 8888 on 127.0.0.1:18888

Forwarding http://127.0.0.1:8000 to box port 8000. Press Ctrl+C to close.

It runs in the foreground until you press Ctrl+C. In a terminal, hi then asks whether to stop the machine too.
If the machine stops while the tunnel is open, the tunnel closes itself within about 30 seconds.
If a local port is taken, the tunnel fails with Address already in use; pick another local port with remote:local.
Colab uses port 8080 on its machines for itself, so hi refuses to tunnel it. Serve on another port, such as 8000.

JupyterLab on a rented GPU is a complete example.

Connect VS Code, rsync, or plain ssh

hi compute proxy <name> is the connection hi compute ssh uses. Put it in ~/.ssh/config and any SSH tool can reach a Colab machine by name, including VS Code Remote-SSH, scp, and rsync:

Host hi-box
  User root
  ProxyCommand hi compute proxy box
  StrictHostKeyChecking no
  UserKnownHostsFile /dev/null

ssh hi-box
rsync -av ./data/ hi-box:/content/data/

hi compute proxy is for Colab. Hugging Face machines have a direct SSH address, <job-id>@ssh.hf.jobs; use hi compute ssh <name>, which finds it for you.

See what is running

hi compute ls

NAME  PROVIDER  HARDWARE    UP   STOPS IN
box   colab     T4          12m  1h48m
job   hf        a10g-small  3m   57m

Colab balance: 1795.89 compute units, currently using 1.07 units/h

ls also shows machines that hi did not start, marked (not started by hi), and stops any machine of yours past its limit. hi compute status <name> shows one machine in detail.

Stop

hi compute stop              # pick from what is running, then confirm
hi compute stop box
hi compute stop --all        # asks first; add --yes in scripts

Without a name, stop lists everything running with its hardware and time left, plus an option to stop all of them, and asks before stopping. In scripts, name the instance or use --all.

Stopping releases the machine and everything on its disk.

Setup logs

Commands such as serve run setup on the machine and keep logs there. hi compute logs <name> shows them, and --follow keeps printing new lines:

hi compute logs box --follow