Verify
Checks amdgpu, KFD and render access, gfx1151 topology, build tools, and a complete ROCm SDK.
Radeon 8060S · gfx1151
Run a serious 27B model
on the machine beside you.
A reproducible Qwen3.8-27B stack for Strix Halo: calibrated IQ4_XS weights, an IQ4_XS DFlash2 drafter, the UMA scheduler ring, and retained-PM4 dispatch. Built from source, installed in your home directory.
- No system ROCm replacement
- SHA-256 verified models
- Exact tested Git pins
Qwen3.8-27B
ROCm / HIP
Retained PM4
DFlash2
IQ4_XS
OpenAI-compatible API
01
One command
From source to serving.
bash <(curl -fsSL https://raw.githubusercontent.com/pwilkin/strix-halo/main/install.sh)
Process substitution keeps stdin attached to your terminal, so the installer can ask where to store the models. Prefer to inspect scripts before executing them: open the installer source or download it first.
Build
Compiles pinned ROCr, HIP, and llama.cpp revisions into isolated, user-owned prefixes.
Fetch
Downloads the selected target, DFlash2 draft, and Bartowski BF16 projector through resumable Xet transfers.
Serve
Installs optimized and generic launchers in ~/.local/bin, then adds that directory to your PATH.
02
Before you begin
Prerequisites, checked—not guessed.
AMD Strix Halo
Ryzen AI Max / Max+ with Radeon 8050S or 8060S exposed through KFD as gfx1151.
64-bit Linux
The amdgpu kernel module, /dev/kfd, and a writable DRM render node must be available to your user.
ROCm development stack
HIP 6.1+, ROCm Clang, hipBLAS, rocBLAS, AMD COMGR, and rocprofiler-register. The installer builds the modified userspace runtimes, not the compiler stack.
30 GiB free
Enough room for source, build artifacts, two model GGUFs, and the BF16 vision projector. A separate model disk is supported.
03
The controlled leaderboard
Reproduced first. Reported second.
We reran the alternative stacks on this Radeon 8060S with the same 31,497-token source-file prompt, 256-token output, fixed-width DFlash2 policy, warmup, and palindrome ordering. In that reproduction, our custom ROCm stack is the fastest on both prefill and decode.
Matched long-prompt reproduction
Our stack leads by 13–25%.
| Rank | Reproduced stack | Prompt | Decode | Acceptance | Two-pass range | Our margin |
|---|---|---|---|---|---|---|
| 01 | Our stackSelected IQ4_XS + IQ4_XS DFlash2 custom ROCm · retained PM4 |
256.84 t/s | 26.256 t/s | 60.74% | 26.232–26.279 | Fastest |
| 02 | ReproducedROCmFP4 FAST + FP4 DFlash2 Nathan-derived Vulkan kernel stack |
215.64 t/s | 23.168 t/s | 57.50% | 23.143–23.193 | +13.33% |
| 03 | ReproducedSame selected IQ4_XS models Nathan-derived Vulkan kernel stack |
219.67 t/s | 20.942 t/s | 60.74% | 20.933–20.950 | +25.38% |
“Nathan-derived” means the measured Vulkan fork at 5e085d12, which ports and credits Nathanw1014’s Strix Halo Vulkan gates.
The ROCmFP4 arm changes target and draft together; the same-model Vulkan arm isolates the engine/backend. All six generations were coherent,
every arm reproduced within 0.25%, and temperatures remained 37–54 °C.
Broader context
Community headlines and our shallow workload matrix
| Stack | Backend | Target + speculation | Reported decode | Protocol / context | Evidence |
|---|---|---|---|---|---|
| Measured hereIQ4_XS + DFlash2, width 6 | Custom ROCm retained PM4 |
All-calibrated IQ4_XS/Q8 IQ4_XS DFlash2 |
25.67 prose 39.14 reasoning 58.53 JSON |
128 generated tokens; deterministic shallow prompts; two counterbalanced servers per cell | Model card |
| Measured hereIQ4_XS + DFlash2, width 3 | Custom ROCm retained PM4 |
All-calibrated IQ4_XS/Q8 IQ4_XS DFlash2 |
25.74 prose 32.48 reasoning 41.03 JSON |
Same 128-token palindrome bracket; lower draft width benefits low-acceptance prose | Model card |
| ExternalROCmFP4 FAST + adaptive DFlash2 | Custom ROCmFPX / HIP | ROCmFP4 FAST ROCmFP4 DFlash2 |
45.44 code 48.26 prose · 46.85 reasoning · 46.03 JSON |
Adaptive width 3–7; 0.95+ reported acceptance; different target, engine, and prompts | ROCmFP4 card |
| ExternalUD-Q5_K_XL + DFlash2 | Vulkan / RADV | Q5_K_XL Q4_K_M DFlash2, width 7 |
31.4 t/s @ 80 W 30.2 t/s @ 70 W |
Four-task 1K-token average; 55.9% acceptance; per-task range 21.1–39.8 t/s | LocalLLaMA report |
| ExternalROCmFP4 STRIX + MTP | ROCm 7.2.4 | ROCmFP4 STRIX Q4_0 MTP, width 4 |
30.30 t/s | 8K decode; 256 tokens; greedy; median of 3–5; 92.6% MTP acceptance | ROCmFP4 card |
| ExternalUD-Q4_K_XL + DFlash2 | Vulkan / RADV | Q4_K_XL Q4_K_M DFlash2, width 7 |
28.5 t/s @ 80 W 27.3 t/s @ 70 W |
Same four-task 1K-token suite as Q5 result; 44.9% acceptance | LocalLLaMA report |
| HistoricalEarlier DFlash2 Vulkan result | Vulkan / RADV | UD-Q4_K_XL DFlash2, width 4 |
26.39 t/s | Zero-context-depth historical result; superseded by newer Q5/Q4 four-task measurements | Nathan toolbox |
| ExternalPieBru evidence-tracked recipes | Vulkan / RADV | Q5 / Q6_K_XL / Q8 DFlash2 |
~23 Q5 ~17–21 Q6 · ~15–18 Q8 |
Sustained recipe bands; prompt and context dependent; approximately 330 t/s prefill | Evidence repository |
| ExternalOfficial Ollama route | Vulkan / RADV | Official qwen3.8:27b route | 20.42 t/s | Nine warm repeats on a 128 GB GTR9 Pro; 292.49 prompt t/s; different runtime/API | Route matrix |
| HistoricalEarly MTP write-up | Vulkan / RADV | Q6_K target MTP width 2 |
18 t/s 9.6 t/s baseline |
Early report; width 2 beat width 5 by 19%; older runtime generation | daily.dev summary |
| Measured hereSelected target, no speculation | Custom ROCm retained PM4 |
All-calibrated IQ4_XS/Q8 | 14.10 t/s | llama-bench tg128; retained PM4; zero-context synthetic decode |
Model card |
| ExternalROCmFP4 STRIX, no speculation | ROCm 7.2.4 | ROCmFP4 STRIX | 13.46 t/s @ 8K 12.58 t/s @ 32K |
Greedy 256-token decode; context-filled server measurement | ROCmFP4 card |
Our protocol
Counterbalanced by construction
Three deterministic workloads, 128 output tokens, one discarded warmup per server, two independently loaded servers per configuration, and palindrome ordering to expose drift.
Correctness
The verifier stayed authoritative
Q8_0 and IQ4_XS DFlash2 produced identical target-output hashes for every workload. Acceptance changed slightly; final deterministic output did not.
Read the numbers
Acceptance is the hidden axis
Our width-6 run ranges from 25.67 to 58.53 t/s solely because prompt-dependent draft acceptance ranges from 29.3% to 95.6%.
04
What changes
A faster path through the same hardware.
ModelIQ4_XS / Q8Calibrated target
→
DraftDFlash2 IQ4_XSHalf-size sidecar
→
EngineUMA ring + TOP_Kstrix-halo branch
→
RuntimeRetained PM4Custom HIP / ROCr
Keep the experiments moving
Useful? Help fund the next weird driver idea.
The models, patches, installer, and benchmark notes are public. Support helps pay for hardware, power, storage, and the occasional very long compile.