#1947 bench: ROG Flow Z13 Windows HIP, Huihui IQ4_XS 256K tuning and placement

本文

## Summary

Records a reproducible configuration search on a 128 GiB ASUS ROG Flow Z13 (Ryzen AI Max+ 395 / Radeon 8060S gfx1151, 96 GiB GPU allocation), Windows 11, Strata 0.1.42 at 61b3fb5, with Huihui Qwen3.8-Flash-Next unsloth-UD-IQ4_XS and the official Qwen MTP draft layer.

The final same-machine six-case comparison improved weighted harmonic decode throughput from 46.41 to 48.87 tok/s (+5.31%). It uses two repetitions per case and 512-token output prefixes. The report distinguishes these from completed xhigh answers and near-256K input tests.

## What changed

Adds one community-results folder with 18 files:
- Portable original, selected text, and later vision configuration snapshots; exact per-profile configurations and selection decisions.
- 809 per-request records: 780 formal throughput requests, 10 complete quality answers, four full-context requests, 12 coding conversation turns, and three fresh-process startup checks.
- Exact deduplicated prompts, full response/timing/grade/memory records, a CSV index, model/pack/engine hashes, matching Windows hipBLASLt tuning table, replay/analysis script and checksums.
- Audited placement log: all 24,576 experts in a 55.43 GiB GPU cache, effective host expert residency zero, mapped embeddings, MTP/head/verify allocations, worker affinity and startup free-memory counters.

The recorded selected text profile uses spec 6 / floor 0.4, prefill 16,384, int8 KV, MTP window 8,192, eight pool workers, and 262,144 total context capacity. Daily prompt cache is six; scored comparisons disable prompt reuse. Server fallback is high; xhigh tests explicitly override it. Thinking budget/cap behavior is documented.

## Validation and limitations

- Offline report verification passes: checksums, exact message references, 809 unique records, 780 formal requests, and independently recomputed final medians/score. Each uploaded Git blob matches the local file. Diff whitespace check and config-render smoke check pass.
- Completed coding checks: 94 per quality profile; full-context BFS 41; 12-turn conversation 376. Three fresh-process starts: 92.531 / 91.407 / 91.281 seconds, each with a complete 26-test check.
- Exact-count reasoning failed under both profiles (expected 601380693). Failed allocations, an earlier unbounded-thinking run and explicit stored-reply regrades are retained.
- TTFT, sustained power/temperature and aggregate peak VRAM were not measured. Windows commit headroom fell to about 0.16 GiB. Two repetitions and this small quality set do not establish broad equivalence, a global optimum or a win over different online conditions.

Related vision warm-up change #1946 remains separate. The later vision config is labeled as outside this text search's measurement coverage. No engine defaults, kernels or headline performance claims change in this PR. No model weights or private user paths are included.