#955 bench: Arc Pro B65 Gen4 results on patched v0.1.40

描述

Reports the latest **2026-10-06 Intel Arc Pro B65 Gen4 x16** measurements using the Strata **v0.1.40 source tag**, original full 512-expert Flash-Next IQ2_XS, 8K context, INT8 KV, confidence-limited MTP window 4, prefill 512, reserve 3072 MiB and four workers. `STRATA_DBG_NAN` is explicitly unset.

The included Intel engine still self-reports `0.1.39-sycl`; the exact v0.1.40 commit (`1735d6471df29b42c26170efaac1f1446a58640f`), binary hash and patches identify this build. It required two additional shared-interface repair patches to compile/link, alongside four retained local safety/test patches. Their scope, unavailable optional modes and reproduction steps are disclosed in BUILD.md; this does not claim a full port of the new CUDA/HIP optimizations.

Three fresh native processes each run five tasks at exact 512/7000 inputs with 640 outputs: **30 measured requests**. A separate three-process campaign runs public benchy-v1 20/2185-token fixtures with 256 outputs: **six measured requests**. All three repetitions, medians/ranges, actual token counts, separate native prompt/decode timings, TTFT, cache state, draft acceptance and numeric telemetry are included.

Five-task decode medians span **40.37–41.28 tok/s at 512 inputs** and **39.81–40.90 at 7000**. At 7000 inputs, prompt medians span **292.99–307.62 tok/s**, with **22.79–23.93 s TTFT**. Public 20/2185-input decode medians are **47.71/47.58 tok/s**; per-shape ranges are included. Different workloads/output lengths are reported separately.

All measured outputs match their qualified reference hashes and all six benchmark processes exit 0. Independent Gen4/AER/GPU/memory/swap/OOM/thermal/lifecycle guards remained enforced. Selected kernel checks, IQ/native-expert checks, 12 supplemental GEMM cases and 24 isolated candidate API/lifecycle checks passed. Retained serving/private API and model recovery passed afterward. The candidate was not installed into normal serving; no candidate reboot test is claimed.

Results only, confined to `bench/results/2026-10-05-community-arc-b65/`; engine fixes are reproduction attachments inside that folder. The report presents the latest measurements and brief build/configuration notes. No generated text, model payloads, credentials or private endpoints are published.

Scope: greedy single-request operation within 8K, retained OS page cache, fixed expert residency and no prompt reuse/adaptation. Public fixtures use the native serving protocol. Larger contexts, broad sampled/RP quality and other optional engine modes are not established. Additional Arc setup/source observations are noted in the report.