iconium is a headless LLM inference server built on a Minisforum AI MAX+ 395 (Strix Halo APU) running LM Studio with ROCm GPU acceleration.
This repo documents the current state of the machine: hardware, OS, ROCm configuration, LM Studio setup, loaded models, and running services.
| File | Description |
|---|---|
| hardware.md | CPU, GPU, RAM, NPU, and storage specs |
| os-and-rocm.md | Ubuntu 25.10, ROCm 7.2.1, kernel details |
| lmstudio.md | LM Studio headless setup, backend transplant, ROCm fix |
| models.md | All installed models, quantizations, and GPU offload status |
| services.md | systemd services: lmstudio and openclaw-gateway |
| api.md | OpenAI-compatible API endpoint reference |
| history.md | Journey from vLLM to LM Studio — what was tried and why |
| Item | Value |
|---|---|
| Hostname | iconium |
| LM Studio server | 0.0.0.0:1234 |
| Active model | google_gemma-4-31b-it (IQ4_XS, 18.35 GB) |
| GPU offload | Full — ROCm0 (Radeon 8060S) |
| Autostart | lmstudio.service (systemd, enabled) |
| OpenClaw gateway | openclaw-gateway.service port 18789 |