Skip to main content
The control plane handles authentication, scheduling, metadata, and the console. It needs no GPU or Docker socket. Compute resources are registered as Fleets, independently of where the control plane runs. A single compute machine also uses a Node Fleet. Install the Node daemon on that machine and enroll it using a Join Token. There is no control-plane local executor.

​
Serving

Node and Kubernetes Fleets support both temporary Playground sessions and Model Deployments. Node reserves host ports and returns the selected machine’s workload address; ensure the control plane can reach the serving port. Kubernetes uses its API proxy. Slurm does not provide interactive serving.

​
Setup

Node

Enroll one or more compute machines.

Kubernetes

Connect an existing Kubernetes cluster.

Slurm

Connect an HPC cluster.
Deploy the control plane with Docker Compose or Kubernetes, then register Fleets.