#2654 Community benchmark: Tesla V100-PCIE-32GB (sm_70), Flash-Next IQ2_XS and IQ3_XXS

Beschreibung

Three community benchmark reports from one box, one settings set and one suite, so they read against each other: the two quants in the mode this machine serves, plus IQ3_XXS in the other expert mode as a mode comparison.

**Hardware:** Tesla V100-PCIE-32GB (sm_70, PCIe 3.0 x16), Ryzen 5 3600, 31.2 GiB RAM + 23 GiB swap, ADATA NVMe, Linux 6.8.0-147, driver 580.178.04, engine built with CUDA 12.8.

**Engine:** local v0.1.42 source build for sm_70 (`STRATA_EXPERIMENTAL_SM60=ON`), sha256 in the reports. It carries one unpushed, opt-in local patch (a conversation park that would hit the RAM floor is paged out instead of refused). The flag was passed but no parking or page-out line appears in any run's engine log and the conversations tested stay far below the 16 GiB / 20-slot limits, so the patch's path was never exercised and the numbers are what an unpatched v0.1.42 build does with the same settings. Every README states this. I have no clean release-build run to compare against; say the word and I will make one.

**Measured** (`tools/strata_bench.py --long`, 3 runs per test, medians, 262,144-token context, `--kv k8v4`):

| | IQ2_XS `--resident-experts` | IQ3_XXS `--resident-experts` | IQ3_XXS `--mmap-experts` |
| --- | --- | --- | --- |
| chat decode | 76.1 tok/s | 66.6 tok/s | 62.5 tok/s |
| prompt tok/s, 4,019 / 16,058 / 32,049 tokens | 1251.6 / 1373.8 / 1344.9 | 1176.9 / 1355.4 / 1275.1 | 930.7 / 1180.1 / 1206.8 |
| first token, same lengths | 3.23 / 11.72 / 23.87 s | 3.44 / 11.88 / 25.18 s | 4.34 / 13.64 / 26.60 s |
| agent session decode / cache hit | 74.4 tok/s / 68.7 % | 76.1 tok/s / 68.8 % | 75.4 tok/s / 68.7 % |

The denser quant costs about 6 % on chat decode and little on prompt throughput, both in resident mode. Resident IQ3_XXS pins 20.64 GiB page-locked and its start log reports 2076 of 2076 lendable prompt-path slots resident, so answering reads no expert from disk. In mapped mode the first 4K prompt is a cold file-cache read (438 tok/s against 1015 warm) - the range is in the report rather than hidden.

**Not tested:** recall/needle checks, any coding task, concurrency (`--clients` was not used and the server was not started with `--batch`), power draw and throttling.

Results-only; no engine changes. `docs/COMMUNITY_BENCHMARKS.md` carries one entry covering the three reports.