Community result: 2.5-3.1x output at 256K context on a 16 GB card (IQ3_S) - calibration, --spec 6, learned profile #775

Sharing a measured result and a writeup: docs/TUNING-256K-16GB.md on my fork ->
https://github.com/JiuYue0820/Strata/blob/docs-256k-tuning/docs/TUNING-256K-16GB.md

PC: RTX 5070 Ti 16 GB, i7-14700KF (8P+12E), 96 GB DDR5-5600, Windows, engine 0.1.38, Qwen3.8-Flash-Next IQ3_S (GSQ-RCO), vision encoder on GPU, context 262,144, KV int8 + streaming (--kv-resident 32768).

Changes: START-HERE.bat --calibrate (kept --pool-workers 13, --spec-min-p 0.70, --pcie-frac 0.00), --spec 6, and the learned expert profile (--expert-profile-save).

Benchmark: one 257,630-token prompt (this repo's own source/docs) through the HTTP API, greedy, reasoning_effort: none. "Warm" = prefix reused, which is how agent sessions actually run. Single runs, so the usual ±20% noise applies.

Prompt reading went from 1,468 to 1,744 tok/s.

The calibration internals on this machine, in case they are useful data points:

• --pool-workers: 19 → 36.9 tok/s, 13 → 54.4, 10 → 41.3 (E-cores stall every verify window; the defaults were measured on a 6-core Ryzen with no E-cores)
• --spec-min-p: 0.3 → 36.6, 0.5 → 41.7, 0.7 → 44.6
• --pcie-frac: 0.0 → 42.1, 0.2 → 37.9, 0.35 → 27.9, 0.55 → 23.9, 0.75 → 20.7 (PCIe 5 x16 + fast CPU: computing a missed expert beats copying it)

Also fills in the IQ3_S 262K row that DETAILS.md leaves unmeasured: on 96 GB of RAM it runs with room to spare, 38.8-53.5 tok/s across all post-tuning runs.

One documentation gap this exposed: a --setup re-run rewrites the config and silently drops a manually added --expert-profile/--expert-profile-save pair (the three calibrated settings survive via the settings file). Worth either preserving or printing a note.