Skip to content
UMKIMDocs
Search limited to 0.1.0. Switch to the latest version for up-to-date results.

GPU host prep

This content is for 0.1.0. Switch to the latest version for up-to-date documentation.

GPU spokes need host drivers, device plugins, and capacity labels before DCM placement can target them. Follow the umkim repo guide docs/components/gpu-host-prep.md for the current playbook tags and verification scripts.

Typical flow:

  1. Prepare the host (drivers / toolkit) before or during cluster bootstrap.
  2. Join the cluster via FleetClusterJoin (or lab Ansible).
  3. Confirm node labels / allocatable GPUs.
  4. Deploy inference catalog items and point FleetLLM backends as needed.