local-inference

name	local-inference
description	Use when running inference probes on the mesh, routing tasks to pi or hermes harnesses, or configuring model loading across LM Studio (adeck:1234), vLLM (zrrh:8000), or llama-server (zrrh direct). Covers gateway config, ctx ceiling procedure, tool-call probing, KV quant config, dense vs MoE tradeoffs, and harness architecture.
compatibility	Designed for the daemonturgy mesh (nxiz/zrrh/adeck). Requires nix develop flake shell and Tailscale mesh access.

local-inference skill

Reference docs for the daemonturgy mesh inference lab.

Contents

• references/process.md — probe procedure, gateway config, proven ceilings, gotchas
• references/harness.md — pi vs hermes architecture, vLLM, llama-server, provider wiring
• references/links.md — LM Studio SDK, llama.cpp, pi, hermes documentation links

Quick reference

LM Studio gateway: adeck:1234 — all SDK and OpenAI endpoint calls for LMS-served models.

vLLM (NVFP4/modelopt models): http://zrrh:8000 — not routed through adeck. Direct only.

llama-server (direct/benchmarking): zrrh via scripts/launch_llamacpp_zrrh.sh

Run a ctx ceiling probe:

nix develop --command python3 scripts/ctx_optimize_probe.py <model> \
  --gateway adeck:1234 --out-dir runs/fit/<run-id>

Run a tool call probe:

nix develop --command python3 scripts/tool_call_probe.py <model> \
  --gateway http://adeck:1234/v1 --ctx 65536 --out-dir runs/fit/<run-id>

Check what's loaded / unload before new probe:

ssh zk@adeck lms ps
ssh zk@adeck "lms unload <model-id>"

See references/process.md for full procedure, proven ceilings table, and gotchas.