A NUT Server in an LXC, and a Grafana Dashboard to Watch It
Running Network UPS Tools inside a Proxmox container, passing the USB cable through, and turning battery runtime into a graph I actually look at.
Homelab write-ups, practical notes, and drafts that are still being shaped.
vLLM is faster than Ollama and vastly harder to configure. Pair-debugging the memory math with a good model collapsed a weekend of flag-guessing into an evening.
What actually fits on a low-end GPU in 2026, why context length eats your VRAM faster than weights do, and the jobs an 8GB card is genuinely the right tool for.
Three nodes, a quorum device, ZFS replication instead of Ceph, and the fencing behavior nobody warns you about until your lights flicker.