[kgmkm]

Summary

Community benchmark for engine 0.1.45 on an Intel Arc Pro B70 32 GB
(Ryzen 7 3700X, 121 GB RAM, native Ubuntu 26.04, SYCL AOT build
bmg-g31), serving the official ISTA-DASLab GSQ-RCO IQ3_S at 32,768
context, single-GPU (a B580 drives the display only), with
experimental speed projection enabled. Results only: no engine, build or
test changes.

One run per condition (stated in the README, no median/range), all prompts
unique with 0 reused tokens:

Configuration	Prompt tok/s	Decode tok/s
decode-story (48 prompt, 1400 gen)	56.6	27.8
decode-explan (47 prompt, 1299 gen)	68.8	38.0
prompt-4k (4021 prompt)	421.3	n/a
prompt-20k (19880 prompt)	1004.5	n/a

Draft acceptance 48.9%/51.4%. The README also includes an ESP refusal A/B
on the same prompt: this time both sides generated 500 tokens
(finish=length, no refusal either way), unlike the 0.1.41 RTX 5060 Ti
report where ESP off refused at 106 tokens.

What changed

• Add bench/results/2026-10-17-community-arc-pro-b70-iq3_s-0145/:
  the README (hardware, settings, method, tables, limitations),
  per-condition medians (measurements.csv), one row per request
  (requests.csv), a /metrics snapshot, the sanitized server config and
  engine start lines, and the client script.
  Local paths are generalized and no credentials are included.

• Add bench/results/2026-10-14-community-arc-pro-b70-wsl2-iq3_s-0143/
  as a bonus: the same card's older WSL2 GPU-PV numbers (engine 0.1.43.1
  plus four uncommitted local WSL2 patches), summary values transcribed
  from a 2026-10-14 migration memo — README and measurements.csv only,
  no per-request data preserved (stated in that README).

• Add nothing else: no code, docs, or index changes (index left to
  the maintainers).

• Not attached: the full engine log and the model/pack files. The main
  README lists what is missing.

Extra Notes

• Native Linux on Intel (oneAPI 2026.2, xe driver, XFCE desktop on
  the B580), so these numbers describe that stack.

• VRAM reserve 500 MiB sits below the engine's own recommendation: a LOW
  warning was active (211 MiB free, engine suggests 801), so the main
  numbers describe that pressured configuration.

• The story decode (27.8 tok/s) trails the explanation decode
  (38.0 tok/s) on lower draft acceptance (48.9% vs 51.4%); outputs
  differ, so this is not a controlled decode comparison.

• Not tested: needle recall, TTFT, vision, soak, other quants, larger
  contexts, a layer split across the B580, calibration sweeps,
  concurrent requests.

[Niko1221]

Thanks for the pull request.

We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to review properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected.

To reopen it takes about two minutes:

1. Rebase on main if it no longer applies, and re-run your checks on 0.2.0-beta.
2. Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other].
3. Fill in the pull request template: the problem and result, and the validation (what you ran, on which revision, what you saw).
4. Comment /reopen. A check runs and either reopens the pull request or lists what is still missing.

The rules are in CONTRIBUTING.md.