debug-lemonadelisted

Diagnose lemonade-server (lemond) problems on the user's Strix Halo box "d" — server not responding on port 13305, models missing from Open WebUI, a model failing to load, slow inference, or ROCm/Vulkan/GPU questions. Use whenever the user invokes /debug-lemonade or reports anything broken or weird with lemonade, lemond, Open WebUI models, llama.cpp backends, GPU page faults, or "models gone" on this machine, even if they don't name lemonade explicitly (e.g. "open webui shows no models", "the model won't load", "inference hangs").
bcmyguest/personal-skills · ★ 0 · AI & Automation · score 58

Install: claude install-skill bcmyguest/personal-skills

# Debug lemonade on Strix Halo This box has a known topology and a few recurring failure modes. Diagnose in this order — each step is cheap and rules out a whole class before touching anything privileged. `sudo` needs a password: when a privileged command is required, ask the user to run it with the `!` prefix instead of attempting it. **Route by symptom:** - Server not responding / Open WebUI shows no models → **Step 1** (is the right lemond up?). - A specific model won't load or errors → **Step 2**. - Slow inference / GPU / ROCm / Vulkan questions → **Step 3**. - Deciding lemond-vs-llama.cpp blame, or whether to escalate → **Step 4**. Read the **Topology** below first — most failures here are a known consequence of it. ## Topology (machine "d") - AMD Strix Halo (gfx1151) APU, 128 GB unified RAM, Ubuntu 24.04. GPU compute runs on **Vulkan** (RADV); ROCm compute has been broken on most kernels (see GPU section). - `lemond` = lemonade-server **system** service: runs as user `lemonade`, port **13305**, config `/var/lib/lemonade/config.json`, models in `/var/lib/lemonade/.cache/lemonade` and `/var/lib/lemonade/.cache/huggingface`. Not readable by user b — journal/API only. - lemond manages its own pinned llama.cpp builds (`<cache>/bin/llamacpp/<backend>`) and spawns `llama-server` children on 127.0.0.1:8001+. It cannot see /home (ProtectHome) and cannot use the user's pip vllm. - Open WebUI on :8080 is just a frontend for lemond; ollama is separate on :11434. - U