RunPod

Set up RunPod for hi compute - an API key, your SSH key, and how hi keeps RunPod pods from running past their time limit.

hi uses RunPod’s REST API v2 directly to start pods on its secure cloud, with SSH access and prices from RunPod’s catalog.

RunPod’s secure cloud often has no free units of its cheapest GPUs. The NOW column shows what is free; if a start fails, hi suggests free GPUs with at least as much memory.

Set it up

1. Create an API key

1. Open console.runpod.io/user/credentials and choose API Keys.
2. Create a key with read and write access; a read-only key can list hardware but not start pods.
3. Make sure the account has a balance; pods stop when it reaches zero.

2. Give the key to hi

hi login runpod

Paste the key when asked; nothing is shown as you type. hi checks the key with RunPod, then saves it in ~/.runpod/config.toml, readable only by you. That is where RunPod’s own runpodctl keeps it too, so both tools share one key and you do not need runpodctl installed.

The guided menu does the same: choose RunPod as the provider in hi compute, and it asks for the key the first time.

In scripts, set the key in the environment instead; it takes precedence over the saved one:

export RUNPOD_API_KEY=...

3. Your SSH key

hi passes your public key, ~/.ssh/id_ed25519.pub, to each pod it starts, so hi compute ssh and tunnel work without registering anything on RunPod. Create a key with ssh-keygen -t ed25519 if you have none.

4. Check

hi compute providers
hi compute hardware --on runpod

Hardware

RunPod’s GPU names are shortened: NVIDIA GeForce RTX 4090 becomes rtx-4090, A100 PCIe becomes a100-pcie. hi compute hardware --on runpod lists CPU pods with 2 vCPUs (cpu3c, cpu5g, …), then every GPU in RunPod’s secure cloud with its price per hour. GPUs with free units come first, marked few free; those with none are listed last as none free:

PROVIDER  HARDWARE      KIND  MEMORY             RATE     NOW
runpod    cpu3c         CPU   2 vCPU, 4 GB RAM   $0.06/h
runpod    rtx-4000-ada  GPU   20 GB VRAM         $0.28/h  few free
runpod    a40           GPU   48 GB VRAM         $0.49/h  few free
...
runpod    rtx-a4000     GPU   16 GB VRAM         $0.25/h  none free

Availability changes by the minute, so treat it as a hint.

hi compute up --gpu rtx-4090 --name box --max 2

Community Cloud

Never put API tokens, passwords, SSH private keys, cloud credentials, or sensitive data on a Community Cloud machine: not in files, environment variables, notebooks, or git remotes. It runs on a third-party host that RunPod does not own. Use it for public code and data only.

By default hi uses RunPod’s Secure Cloud, in RunPod’s own datacenters. Community Cloud rents the same GPUs from third-party hosts, 25–54% cheaper, and is still on-demand: nobody can take the machine back. Choose it by adding @community to the hardware name:

hi compute hardware --on runpod --community
hi compute up --on runpod --gpu rtx-4090@community --max 2h

hi shows the warning above before every Community start, even with --yes, and again on every hi compute ssh, tunnel, logs, and serve to the machine. Hosts vary: a host going offline ends the machine, so keep your work checkpointed.

Through a hi server, the Slack request carries the same warning. An approved Secure start is never moved to Community Cloud when its hardware is sold out, and a Community start is only replaced on Community Cloud. A group’s hardware list in the policy must name @community hardware explicitly for the group to use it.

What hi starts

up waits until the pod is running and reachable over SSH, which can take a few minutes while RunPod pulls the image.

Time limits

RunPod pods have no built-in time limit, so hi enforces --max twice:

• A small watchdog on the pod terminates it at the limit, using the pod’s own RunPod key and the runpodctl that RunPod puts in every pod. This works even when your laptop is off.
• The same local watcher as for Colab stops it from your laptop, and every hi compute ls checks too.

hi compute stop terminates the pod rather than pausing it, because a paused RunPod pod still bills for its disk.

Not yet supported on RunPod

• hi compute run: RunPod has no run-to-completion jobs; start a pod and run your script over hi compute ssh.
• Model recipes: hi compute serve works with --gpu and --quant, building llama.cpp on the pod over SSH as on Colab, but has not been tested on RunPod.

Troubleshooting