#2571 bench: community report — 2x RTX 3090 layer split, UD-Q4_K_XL at 262K (Strata 0.1.44.1)

Beschreibung

## What this adds

One community benchmark report folder, plus its entry in the `Community reports` list. No engine or tool changes.

## Hardware

- 2x RTX 3090 24 GB (sm_86), GPUs 0 and 1, 350 W each, PCIe 3.0 x16 (measured 7.9 GB/s host-to-device — the link the expert tier streams over)
- 2x Xeon Platinum 8163 @ 2.50 GHz, 48 cores / 96 threads, 2 NUMA nodes, 160 GiB RAM
- ZFS on NVMe SSD for the model; driver 615.71.09, CUDA 13.4; Linux 7.0.14-23-pve, service inside an LXC container with `LimitMEMLOCK=infinity`
- A third card (Quadro RTX 4000) runs a parsing service on another container and is not used by this run

## Model and configuration tested

`unsloth/Qwen3.8-Flash-Next` UD-Q4_K_XL (4 GGUF shards, 103.69 GiB), 262,144-token context, 1 slot, layer split across both cards:

```text
--expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 --mtp .../rt
--max-context 262144 --kv k8v4 --kv-resident 32768 --ple-io ram --no-ple-prefetch
```

Strata 0.1.44.1, source build from the release tree (no git checkout, so no commit hash recorded).

## Results (medians, 3 runs per test)

| Test | Prompt tok/s | Decode tok/s | TTFT |
|---|---:|---:|---:|
| short chat | — | 84.9 (76.2–106.8) | 0.39 s |
| 4K fresh | 741.1 (740.5–741.7) | — | 5.46 s |
| 16K fresh | 1440.1 (1413.6–1440.7) | — | 11.20 s |
| 32K fresh | 1521.2 (1508.9–1521.4) | — | 21.12 s |
| agent session | see report | 93.1 (74.3–121.5) | 5.57 s |

Agent-session cache hit 68.7 % (68.5–68.7). Suite total 239 s.

## Notes for the reviewer

- **Run against the live server** with `--url`, so nothing was restarted and the service kept serving real traffic during the suite. `engine-log-excerpt.txt` was therefore taken from the service log by hand rather than by the tool.
- **Engine-reported basis verified per request**: all 15 requests carry `decode_tps_source` and `prompt_tps_source` = `"engine"` in `results.json`.
- **`--no-ple-prefetch` is in the configuration and is load-bearing on Linux.** The report says so plainly and links #2281 for the A/B and the mechanism; it is stated as configuration, not argued here, so the measurements can be read on their own.
- **The 4K end is the weak spot** (741 tok/s against 1440 / 1521 at 16K / 32K) and we have not isolated why.
- **Not a controlled comparison** — one production configuration. The only cross-version figure quoted (v0.1.43 → 0.1.44.1 on the same host) is flagged in the report as two single cold-start measurements, not a suite.
- Same-day functional checks (OpenAI, Anthropic, a control token quoted as literal text, an 8K prompt) are recorded under *Correctness and limitations*; the speed suite itself ran no correctness check.
- Hardware rows the tool could not read (PCIe width and measured bandwidth, storage type, the unused third card) were filled in by hand, as `docs/COMMUNITY_BENCHMARKS.md` asks authors to do. Home folders, host names and credentials were checked to be absent.