Skip to main content
SkyPilot runs AI workloads on any infrastructure from one task YAML. Native support for Runcrate (sky launch --infra runcrate) is in review with the SkyPilot project. Until it ships, you can run SkyPilot on Runcrate machines as an SSH Node Pool: you launch the machines, SkyPilot installs its runtime on them, and every sky launch, sky exec and sky jobs launch runs there.

​
What you need

  • SkyPilot installed on your computer (pip install "skypilot[kubernetes]"), plus kubectl, socat and GNU netcat (on macOS: brew install kubectl socat netcat).
  • An SSH key pair. Add the public key on the SSH keys page or with the API.
  • Credit in your workspace. The machine bills by the hour until you terminate it.

​
Set it up

1

Launch an instance with your SSH key

In the dashboard, launch an instance and pick your SSH key. Or with the API (ssh_key_id comes from GET /api/v1/ssh-keys, instance_type_id from GET /api/v1/instances/types):
Wait until its status is running and note its IP. Pick an offer with port_mapped: false: SkyPilot needs SSH on port 22.
2

Describe the pool

Add the machine to ~/.sky/ssh_node_pools.yaml:
3

Install SkyPilot on it

sky check ssh lists runcrate under SSH Node Pools.
4

Run your work

SkyPilot asks for at least 4 vCPUs by default and k3s keeps about one, so pass --cpus and --memory on machines with 8 vCPUs or fewer.Or set resources.infra: ssh/runcrate in a task YAML.

​
When you are done

Remove the SkyPilot runtime, then terminate the instance. SkyPilot does not stop Runcrate billing: only terminating the instance does.

​
Good to know

  • A pool of several machines needs port 6443 open between them. Add a firewall rule for it in the environment (Settings > Environments), limited to the machines’ addresses.
  • With native support, SkyPilot will launch and terminate Runcrate machines itself, pick the region and GPU from Runcrate’s live prices, and tear clusters down with sky down.