New dual-3090 A/B report: faster long-prompt processing, tighter host RAM headroom
A same-rig community comparison reports throughput alongside growing host memory use. Its arms use different model/engine stacks, so it does not isolate engine gains. This is one self-report; QEngine also carries a noncommercial license.
Source: club-3090 社区自报及维护者回复
At 20:41 UTC on November 4, club-3090 issue #1944 published a same-rig comparison of QEngine 1.3.4.3 Flash-Next and the vLLM 27B dual-fast stack used on that machine. The author ran 27B, QEngine, then a 27B control re-bench. Sustained resource headroom matters alongside throughput. This is one community self-report, not an independent AraoAI reproduction.
-
Read the conditions together. The two RTX 3090s were capped at 260/280 W, paired with a Ryzen 7 7700X and 124 GiB of RAM. Driver 615.71.10 refused driver P2P. The vLLM image was v0.32.0. Both arms used thinking at low effort; throughput tests used three warmups and five measured runs. They nevertheless used different models and runtimes, so this does not isolate the effect of switching engines for one model.
-
Gains were workload-dependent. Narrative decode was 100.16 versus 151.56 tok/s; code decode was 132.72 versus 139.07. Prefill at 10K was 2520 versus 2328, while 90K was 1738 versus 2949, with just one measurement at 90K. These figures do not support a universal speedup. Each arm also received only one task-pack draw: 132/150 versus 128/150 cannot establish a model capability ranking.
-
Starting successfully is not proof of safe co-location. The author reported QEngine host memory use growing from 98 GiB at ready to 117 GiB after tasks and benchmarks, leaving 7.5 GiB available. The maintainer confirmed archive growth and warned that the margin protects the engine, not other services. No full soak was performed; long-term stability remains unproven.
Commercial use has a separate gate. QEngine’s checked LICENSE is PolyForm Noncommercial 1.0.0, not Apache-2.0. Commercial serving must not be assumed permitted. Model identity, runtime, workload, sustained resources and licensing all belong in the decision—not just peak decode.
Sources: report and maintainer response; QEngine LICENSE.