A resident model holds the passively cooled Tesla T4 in P0 at about
30 W, which keeps the poweredge-t640 fans near 4,800 RPM. Unloading
idle models lets the card return to P8 at about 15 W.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bump the localai role to the commit that keeps agent, KB, config and backend
state on the models PVC and patches hostNetwork and the p2p token secret.
Set the playbook to p2p/federated off to match the live deployment: Eric
chose T4-only after p2p workers broke model loads.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>