The issue is that the 'cloud', 'region', and 'instance_type' parameters must be set by the optimizer. To solve this, you can follow the recommended way of utilizing on-prem resources by installing Kubernetes on your on-prem cluster and using SkyPilot's Kubernetes support to submit tasks. If you're on GCP, you can refer to the provided guide to set up a Kubernetes cluster on GCP VMs and submit tasks to it using SkyPilot. Note that GPU support is currently in-progress and only single-node CPU jobs are supported in the master branch.