#3481 [Benchmark] RX 9060 XT 16GB / Qwen3.8-Flash-Next / IQ2_XS / strata-qwen38-thinking-v1-draft.1 (custom track)

Descripción

## Problem and result
A standard-benchmark result (`strata bench run`, suite `standard-v1`, protocol `strata-qwen38-thinking-v1-draft.1`, draft)
for an AMD RX 9060 XT 16 GB (gfx1200) with the original Qwen3.8-Flash-Next GSQ-RCO **IQ2_XS**, **custom track**.
Status **complete**: 21 of 21 required requests, 0 failed, 0 skipped, sanity 12/12, optional concurrency stage included.

| case | prompt tok/s | decode tok/s | TTFT s |
|---|---|---|---|
| short | 69.1 (54.7-75.5) | 45.9 (45.4-47.0) | 1.16 (1.06-1.45) |
| prompt-4k | 1219.5 (1213.2-1234.2) | 41.6 (40.9-46.2) | 3.39 (3.36-3.41) |
| prompt-16k | 1507.8 (1506.2-1508.1) | 42.5 (41.9-49.0) | 10.93 (10.92-10.93) |

Four requests at once (server capacity 2): 48.2 / 50.7 / 48.2 tok/s total per round.
Report: `bench/results/2026-10-20-standard-rx9060xt-20261020T163249Z-19c44879/README.md`.

## Related discussion
#3480 (the gfx1200 prompt-path settings this configuration uses: draft side stream on, queue priming off, prompt ring 384).

## Approach
Results only, no engine change. The engine is upstream 0.2.1 (`9ae93f7b`) plus ten commits, published at
[shenmintao/Strata `nas/gfx1200-021` @ 690e4f6f](https://github.com/shenmintao/Strata/tree/nas/gfx1200-021); the image
carries that commit in `/opt/strata/COMMIT`, which the runner records. The full run config (LAN address replaced),
the engine log of the run and SHA256SUMS of every file are in the folder. Required stages: warm-up, 3x short / 4K / 16K,
12 sanity checks; optional: concurrency (3 rounds of 4).

## Validation
**Passed:** `strata.py bench run --url http://127.0.0.1:8080 --suite standard-v1 --optional-concurrency --track custom ...`
(the exact flags are in `results.json`), engine 0.1.46.1 at 690e4f6f, fresh start, 3 min 36 s after load; the runner's
validator reported status `complete`.
**Failed:** none in this run. Two earlier runs of the same configuration are kept in `provenance/earlier-runs/` and in the
README table: run `153736Z` had no COMMIT file in the image (nonconforming; also R4 answered `{"north": -124, "south": -875}`,
11/12), run `162259Z` read the 16K prompts back from the engine's conversation disk tier filled by the first run
(nonconforming "unexpected reuse"); their session files were removed before this run.
**Not run:** the default track (an unmodified release with setup's settings) on this box; a gfx1201 or Windows run.

## Limits
Custom track: not comparable with default-track results. The engine also serves a household voice assistant (prefix
warm-ups about every 20 minutes), declared as background. The runner could not read the GPU name, VRAM, driver and ROCm
versions inside the container; the README's added notes give them. Prepared with an AI assistant; all numbers come from
the files in the folder.