#2737 bench: community report - GTX 1080 Ti (sm_61): eight-knob screen, long context 32K-250K with needles, 1h soak

本文

A results-only report from the same rig as #2642 / #2684 (GTX 1080 Ti 11 GB, 2x Xeon E5-2696 v4, IQ3_S, engine 0.1.44.1, CUDA 12.8): eight opt-in knobs screened one at a time (base = #2695's `STRATA_DENSE_F32=1`), the winners stacked, then long-context behavior with needle checks and a one-hour soak. It also answers the TEST_REQUESTS row "a long soak (an hour or more of back-to-back requests) ... on a resident setup" (see below).

## What it adds

`bench/results/2026-10-16-community-gtx-1080-ti-tuning-soak-longctx/`: pair transcripts with the engine's own lines and quality probes, `needles-lg-grow2.json`, `soak-results.json` (237 rows), the runner scripts (incl. the root-level cpuset cgroup wrapper this rig needs for node binding), and a README in the report format. One list entry in `docs/COMMUNITY_BENCHMARKS.md`.

## What it leaves alone

No engine, setup or tool change; no default is flipped. `docs/OLDER_GPUS.md` is untouched here (#2695 already edits it).

## The screen (medians of 3 requests per arm; decode verdicts only when all 3 moved together)

| Knob | prompt short/long | decode short/long | verdict |
| --- | --- | --- | --- |
| `--mtp-hnorm stream` | unchanged | **+28% / +28%** | adopted (#2227 replicated on sm_61, bigger than sm_89) |
| `--ple-io ram` | unchanged | **+9% / +12%** | adopted (28.8 GB PLE table; mlock refused by ulimit, pages faulted in) |
| no `--pcie-frac` (auto picks 0.32) | unchanged | **+12% / +20%** | adopted (#2498's 0.00 recommendation measured slower here too) |
| `--adapt-async 1` | unchanged | +7% / +2% | not adopted (weak) |
| `--kv k8v4` | unchanged | unchanged | not adopted |
| `--kv-grow` + `--expert-profile-save` | unchanged | unchanged | profile-save kept; `--kv-grow` dropped (negative at long context too) |
| `--aux-cpus auto` | unchanged | **−23% / −15%** | rejected |
| `--batch 2 --batch-mtp` | **−45%** | −20% | rejected (2-client burst: 40.6 s queued vs 71.4 s batched; #2126's startup hang did not reproduce) |
| the three adopted together | short +3% | **+10% / +16%** | confirmed |

## Long context and soak

32K/128K/250K fresh prompts: 415 / 372 / 333 prompt tok/s, decode flat with depth (28-32). KV streaming at 128K served 88.27% of 431,340 block reads from VRAM (203.9 MiB from RAM). Needles at depth 50: **FOUND at 124,782 and at 251,546 tokens** — first correctness evidence at the far end of the 262K window on an 11 GB Pascal card with `--kv int8 --kv-resident 32768`. `--kv-grow` was negative at every depth for these fresh prompts (its #2650 benefit is drift inside one growing session). Soak: 60 min, 237 requests, **0 failures, no drift**, end temp 73 C without throttling.

## Limitations

One round per knob (decode is noisy on this rig, ±20% between runs; prompt numbers are steady ±3%); 250K is one run per arm (the needle is the anchor); correctness evidence is two needles plus one arithmetic probe identical in all 20 arms — no quality suite; single model, single card, Linux only.

Cross-refs: #2227 (mtp-hnorm), #2498 (pcie-frac), #2650 (kv-grow), #2642 (same rig's suite report — it never got the 32K prompts), #2684 (the NUMA binding used), #2695 (the same day's dense-F32 A/B).