The config that worked, captured the day it worked — drop in the files, hand the runbook to an agent, rebuild it anywhere.
Versioned snapshots of working local-LLM / GPU serving configurations across the fleet.
Each dated folder is a self-contained, drop-in snapshot of one machine's setup: the actual
config files, the benchmarks and charts that justify it, captured system logs, and an
IMPLEMENTATION.md runbook a Claude Code agent can execute to rebuild it from scratch.
| Date | Host | Summary |
|---|---|---|
| 2026-06-11 - PLUTO | PLUTO (Rocky 9.8) | Uncensored Qwen3-30B-A3B served on dual GPUs (RTX 3060 + GTX 1080) via llama.cpp, wired to Hermes. ~10× faster than single-GPU; tool calls 15 min → 3 s. |
- Open the folder's
README.mdfor the what/why/results. - Hand
IMPLEMENTATION.md+ theconfigs/files to a Claude Code agent: "implement this configuration on , verifying paths/UUIDs against the live machine." benchmarks.md+charts/are the acceptance criteria — reproduce them to confirm success.