Radeon Stack Lab

These notes are the written weekly tracker. Radeon Stack Lab is the Ubuntu companion app: same living catalog, pickable profiles, run on the box that has the R9700.

It is an independent local app (Origin repo / local install). There is no public product domain to click. If you have the repo, run it next to this site. If you do not, this page is still the map of what the profiles mean.

What it is for

The app does not replace ROCm, llama.cpp, or Strata. It selects a profile and points you at a serve path those tools already understand.

Profiles (conceptual)

qwen-reliable — daily driver

• Model: Qwen/Qwen3.6-35B-A3B
• Quant / weights: unsloth/Qwen3.6-35B-A3B-GGUF · UD-Q5_K_XL
• Serve: llama.cpp Vulkan (HIP gfx1201 if it builds)
• Use when: overnight chat, anything you need to trust tomorrow morning
• TODO: exact command line / measured VRAM from rig owner.

qwen-fast-agent — fast agent

• Model: Qwen 3.8 Flash (Qwen/Qwen3.8-Flash-Next)
• Quant / weights: unsloth/Qwen3.8-Flash-Next-GGUF · UD-Q6_K_XL
• Serve: Strata v0.1.42; llama.cpp Q6 fallback
• Use when: agent loops that need the higher tok/s we measured (~32.6 vs ~16)
• TODO: exact command line / measured VRAM+RAM from rig owner.

Abliterated checkpoints are not a Stack Lab default profile. If you experiment, do it after the daily driver works — Abliterated.

How it relates to this site

Radeon Stack Lab (Ubuntu app, local)
    → profile: qwen-reliable | qwen-fast-agent
        → same pins as src/data/*.yaml
            → llama.cpp / Strata / vLLM-try on the R9700
                → optional harness on /v1

Weekday scrape updates the YAML here. The companion catalog should follow. Maintainer steps: weekday scrape. Public cards: Recommended stacks. First-party tok/s: benchmarks.

Benchmark my rig (upcoming)

Stack Lab will grow a Benchmark my rig flow that records model, quant, context, engine, gfx, and tok/s — no invented numbers. It is not shipping in this snapshot. Until it lands, reproduce from llama.cpp or Strata and read the published set on benchmarks.

What to do if you only have these notes

1. Follow Getting started.
2. Serve the daily driver (tutorial).
3. Attach a harness only after /v1 answers.
4. If you later clone the Origin repo for Stack Lab, keep its catalog aligned with this YAML.