WumboJetsII

Headless Fedora 44 AI compute node for local-model evaluation, LLMGauge runs, and WumboLabs evidence.

WumboJetsII is the WumboLabs AI compute node: a headless Fedora 44 system dedicated to local-model evaluation and compute work rather than daily desktop use.

It runs Fedora 44 without a desktop session and supports LLMGauge testing, deeper local-model evaluation campaigns, and local execution work for the wider lab.

Headless AI compute / evaluation hardware

At a Glance

OS Fedora 44

Form Factor Headless, no desktop session

GPU AMD Radeon AI PRO R9700 32GB

CPU / Memory Ryzen 7 9800X3D / 32GB DDR5

AI Stack llama.cpp ROCm/HIP and Vulkan builds

Local AI R9700 evaluation hardware

Role Headless AI compute node

Role in WumboLabs

WumboJetsII is where WumboLabs local AI compute happens: LLMGauge runs, deeper model-evaluation campaigns, and supporting execution work for the other systems in the lab.

The AMD Radeon AI PRO R9700 provides 32GB of physical GPU VRAM for local AI compute. ROCm/HIP and Vulkan development and testing are hardware-scoped; runtime qualification applies to the tested model, artifact, build, backend, and configuration, not to every workload or to a preferred runtime.

The Qwen3.8-27B deep-evaluation campaign ran on the RTX 5070 with llama.cpp/CUDA; that GPU has since been removed, and the campaign stands as historical evidence for that hardware. Read the historical campaign report for its runtime-specific findings and limitations. The Qwen3.8 H1 canonical profile from that era is documented on the canonical Qwen3.8 evaluation page.

Focus areas

Headless Fedora

Fedora 44 operated without a desktop session, managed remotely, and kept to a stable baseline for compute work.

Local LLM Evaluation

LLMGauge workflows and WELP scientific evaluation on the R9700, with model/runtime compatibility, context limits, generation speed, VRAM behavior, and reliability recorded for each tested stack.

Lab Compute Infrastructure

Execution node for WumboLabs evaluation campaigns and local runs that support the other systems in the environment.

ROCm and Vulkan Work

Hardware-scoped AMD compute testing on RDNA4 (gfx1201), including llama.cpp HIP and Vulkan builds, with claim boundaries kept explicit.

Operating goal

The goal is to keep the system stable, reliable, and useful as dedicated evaluation hardware.

Experiments are welcome, but the baseline should remain dependable for compute work.