Skip to content

Krieger (GPU host)

The home GPU machine — Windows, on the tailnet as krieger (100.90.162.24). Normally asleep; woken on demand rather than left running 24/7.

Hardware: ASUS B860M MAX GAMING AX board, RTX 5060 with 8 GB VRAM, 240 mm AIO, Fractal North case. The 8 GB is the binding constraint — Ollama, ComfyUI and the transcription service all compete for it, which is why they’re arbitrated rather than run side by side. (UniFi fingerprints the machine as a Z490-E; that’s wrong, ignore it.)

  • Ollama — local LLM inference. Models currently proxied through LiteLLM include Qwen 2.5 7B, a Foundation-Sec 8B fine-tune, Moondream, Qwen 3.5 9B, and two uncensored models for Open WebUI: Dolphin 3 (Llama 3.1 8B) and an abliterated Qwen3 8B (huihui_ai/qwen3-abliterated:8b-v2). A new model needs three things: an ollama pull on Krieger, an entry in litellm/config.yaml (then restart the litellm container, since a config-only change doesn’t recreate it), and adding it to the open-webui LiteLLM key’s model list, which is an allow-list. Dolphin 3 has no tool support in Ollama, which rejects any request carrying tools (Open WebUI sends them when tools or native function calling are on), so its LiteLLM entry has additional_drop_params: ["tools", "tool_choice"].
  • ComfyUI — image generation.
  • Apollo — game streaming to the RP5 handheld, Apple TV, iPhone and the Skull Canyon NUC, as two instances with separate game lists that a scheduled script keeps in step with the installed Steam games. Not part of this repo: documented in the knowledge repo, home/krieger.md.
  • Presumably a transcription service too, given the transcribe-wake-proxy described below — worth confirming/documenting what specifically runs there.

Lana never talks to Krieger directly. Each service has its own small “wake-proxy” container on Lana (ollama-wake-proxy, comfyui-wake-proxy, transcribe-wake-proxy, all part of the litellm stack in docker-infra) that:

  1. Receives the request meant for Krieger.
  2. Sends a Wake-on-LAN packet via Home Assistant if Krieger looks asleep.
  3. Waits for the target service (Ollama, ComfyUI, etc.) to come up.
  4. Forwards the original request over Tailscale.

This is why LiteLLM’s ollama_chat/* models can have a longer first-request latency — that’s Krieger waking up, not a slow model.

There’s also a separate, more direct Wake-on-LAN call from Home Assistant that predates the wake-proxies and may power an in-app “wake now” button rather than the transparent per-request wake above. Worth checking litellm/ollama-wake-proxy in docker-infra before assuming which code path is actually live.

When a wake “works” but nothing responds

Section titled “When a wake “works” but nothing responds”

Krieger’s Realtek 2.5GbE NIC has a failure mode where the link comes up and transmit works, but it receives zero frames — so it ends up on an APIPA address and is unreachable, while Home Assistant reports the Wake-on-LAN packet as sent. Wake-on-LAN keeping the NIC on standby power is itself the trigger, so the wake-proxies make it more likely, not less. Restart-NetAdapter sometimes clears it; otherwise it needs a cold power cycle with the AC removed. Symptoms, diagnosis and the adapter settings involved are written up in the personal knowledge repo (home/krieger.md), served to Claude by knowledge-mcp.

Besides interactive use, the nightly update-agent run has the local Qwen model judge each pending image update, which wakes Krieger. That only happens on nights when something actually has a new version. Qwen is the agent’s only model, so if Krieger is gaming or won’t wake, that night’s updates are all reported as needing review.

This page is a starting point, not exhaustive — worth adding:

  • Exactly what the transcription service is and how it’s invoked
  • How sleep/wake is triggered on the Windows side (Home Assistant integration, WoL config)