#3026 bench: community report, RTX 4060 Laptop 8 GiB (Windows 11, 32 GB RAM), Q2_0, engine 0.1.45

说明

bench: community report, RTX 4060 Laptop 8 GiB (Windows 11, 32 GB RAM), Q2_0, engine 0.1.45

Adds `bench/results/2026-10-18-community-rtx-4060-laptop-gpu-qwen3-8-flash-next-q2-0/`: the README the
suite wrote, `results.json` per request and the engine log excerpt. I added a "Notes from the author"
section for the build's provenance, what was not tested, the long-context figures below, and one
measurement the suite does not cover. Nothing else in the tree is touched.

Why it may be useful: **this is the smallest VRAM tier in `bench/results` so far** — there is no single
8 GiB card (the smallest existing entry is a 12 GiB RTX 3060, then 16 and 24 GiB), and 8 GiB is the tier
where the expert cache holds about 9% of the (layer, expert) pairs. `docs/TEST_REQUESTS.md` also asks for
"a host with 32 GB of RAM or less" because "the expert file tier and the resident mode depend on how much
of the model fits in RAM"; this is such a host (31.7 GiB), on Windows rather than Linux.

Suite medians (3 runs each, temperature 0, thinking off, the install's own configuration):

| | |
| --- | --- |
| short chat decode | **30.5 tok/s** (22.9-44.1, n=9), first token 1.10 s |
| 4K fresh prompt | **220.1 tok/s**, first token 18.33 s |
| 16K fresh prompt | **220.4 tok/s**, first token 72.97 s — no decay from 4K to 16K |
| agent session, follow-up turns | decode **23.1 tok/s** (7.3-33.7, n=15), first token 16.21 s, conversation-cache hit 68.6% |

**Long context, from daily use rather than the suite** (the suite stops at 16K). This install also runs an
agent client against the same Q2_0 server; its engine log covers ten days and 1,932 requests with thinking
on and the client's own sampling, so it is not the suite's protocol:

- the longest prompt carried **113,621 tokens** (112,781 reused from the conversation cache + 840 read) at
  99.5 prompt tok/s, and generated 3,175 tokens at 24.5 tok/s decode
- 88 requests with a 64K-120K prompt: prompt median 47.0 tok/s, decode median 24.0 tok/s
- 27 requests above 96K: decode median 23.9 tok/s, i.e. no decay against the 21.2 tok/s median of the
  1,072 requests above 32K
- the longest single generation was 64,879 tokens (from a 649-token prompt) at 16.2 tok/s

So context is not what limits this machine: the KV cache keeps 20,480 of 131,072 cells per QSA layer in
VRAM with about 0.84 GiB in pinned RAM, and the conversation cache reuses most of a long prompt.

The one thing I would look at: decode here tracks how much of the expert set fits in RAM when the server
starts. Same box, same model (Flash-Next IQ2_XS), same settings, two runs on the same day: a RAM
complement of 10.53 GiB of the 31.03 GiB the VRAM cache does not hold (34%) gave a decode median of
15.5 tok/s; 16.93 GiB (55%) gave 21.6 tok/s. Only the free RAM at startup differed (a clean boot versus a
session with other work loaded).

Not tested in the suite: prompts above 16K (covered above from daily use), `--clients` concurrency,
vision, a layer split, a long soak, and any non-Windows host.

Note: the suite needed two local fixes to `tools/strata_bench.py` to run here (`port` never assigned; the
`CMakeLists.txt` read without an encoding on a Windows whose ANSI code page is not UTF-8). Those are the
sibling pull request; the measurements are unaffected.