#2024 bench: RTX 5090 desktop 1M-agent — 0.1.42 vs 0.1.43 interleaved A/B (overlap default pays +1.2/+2.4% decode; #1712 wash-to-negative on sm_120)

Descrição

Follow-up to our merged 5090 reports (#1592 lineage), same box and config, on the v0.1.43 release.

Six legs, A B A B A B, wrapper paired per tag, version gate passed on all six, 30 s settle. Two probe sets per leg: our classic 200K/decode/needle/repetition suite, and strata-bench `gprobe` short prompts (150/300/450 tokens) — the band the #1712 pinned-upload change touches.

Headlines:
- **ADAPT_OVERLAP pays on this regime**: instruct decode +1.2%, greedy +2.4% (engine log tok/s; every 0.1.43 leg beat every 0.1.42 leg). Prefill/needle flat. All gates clean (needle 6/6, zero repetition 6/6), consistent with the release's byte-identical CUDA claim.
- **#1712 is a small consistent cost here**: all 9 gprobe pairs +0.1–0.4%. The staged walk on sm_120 + native IQ3_S + 91–94% hit rate is expert-bytes-bound; the pageable-copy stall it removes is smaller than the memcpy+event-sync it adds. Not an objection to the default — the gfx1100 fix is worth more than 3 ms — but a measured data point for a per-platform gate.
- **Regime contrast with #2010** (5090 laptop x8, overlap −21%): this box streams only 4–5.5% of routed experts in steady state. A residency/PCIe-fraction-aware default could cover both. Cross-posted there.

**Verdict: 0.1.43 is our daily keeper.**

Raw data in `bench/results/2026-10-11-community-rtx5090-1m-agent-0143/data/` (twelve per-leg JSONs + serve-log extracts).