-
Running Ollama on a $40 Radeon
The RX 580 is ancient by GPU standards, but with the Vulkan backend it does genuinely useful local inference. The gist of getting a model running: Tokens per second will not win any benchmarks, but for a card that costs less than a nice dinner, it is hard to complain. Use ollama ps to watch…
-
A Raspberry Pi status board for the rack
Actual blinkenlights: a small panel driven by a Raspberry Pi that shows what is up and what is not. Green is good, amber is thinking about it, dark is asleep. It reads from the same Prometheus the rest of the homelab uses, so the lights never disagree with the dashboards.