#2771 bench: RTX 5090 Laptop GPU follow-up for 0.1.44.1

说明

Follow-up to the RTX 5090 Laptop entries (0.1.40.1, 0.1.40.3, 0.1.41 on main; 0.1.43 in #1976): same laptop, model files and scripts, Strata version changed. Results only, no engine changes.

New folder: `bench/results/2026-10-16-community-rtx-5090-laptop-engine-0.1.44.1/`

**Median decode / prompt tok/s (Q2_0, 131K context, greedy, reasoning off, 256-token cap):**

| Prompt tokens | 0.1.43 | 0.1.44.1 |
| ---: | ---: | ---: |
| 4,096 | 140.5 / 1,906.9 | 128.4 / 1,804.1 |
| 32,768 | 141.5 / 2,871.2 | 137.9 / 2,745.6 |
| 128,000 | 131.6 / 2,767.5 | 124.6 / 2,660.9 |

- The 3–9% drop is the laptop, not the engine: its vendor power profile went from "Extreme" to "Normal" on 2026-10-13 (CPU too hot in daily use). A before/after run on the same afternoon with identical configuration (0.1.43 vs 0.1.44.1, 1K–100K prompts) landed within ±2%. The README says so up front.
- Second difference to the 0.1.43 entry: conversation cache 12 GiB / 8 slots instead of 6 GiB / 2 (host RAM only; zero reused tokens in the speed requests).
- Needles 6/6. Coding check 6/10 (the Roman-numeral function got the subtractive forms wrong; 0.1.43 passed 10/10 on the same greedy prompt). Treated as one data point: our private agent task suite (44 tasks x 3 runs, production configuration) went the other way, 126/132 with 0.1.44.1 against 122 with 0.1.43, the best result on this machine so far.
- Restoring a parked 61,775-token conversation after a foreign request: 0.7 s (61,928 tokens reused), unchanged.
- The source was downloaded by commit (`archive/<sha>.zip`), every file was checked against the git tree, and it was built locally for sm_120 with CUDA 13.3.

(#1462, #1608 and #1976 are still open; this one is independent of them, the links to the earlier READMEs resolve once they are in.)

🤖 Generated with [Claude Code](https://claude.com/claude-code)