[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f3us734apsv1va":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":18,"bodyZh":22,"bodyEn":23,"sourceName":24,"sourceUrl":25,"updatedAt":26},"dual3090-qengine1343-ab-memory-headroom-2026-11-05","HARDWARE","2026-11-04T23:05:17.133Z","双 3090 新 A\u002FB 自报：长提示处理更快，主机内存余量却更紧","New dual-3090 A\u002FB report: faster long-prompt processing, tighter host RAM headroom","社区同机对照同时报告吞吐与主机内存增长；两臂是不同模型加引擎组合，不能据此计算单一引擎增益。仅一份自报，QEngine 另受非商业许可证限制。","A same-rig community comparison reports throughput alongside growing host memory use. Its arms use different model\u002Fengine stacks, so it does not isolate engine gains. This is one self-report; QEngine also carries a noncommercial license.",[12,15],{"id":13,"name":14},5,"Qwen3.8-27B",{"id":16,"name":17},11,"Qwen3.8-Flash-Next",[19],{"id":20,"name":21},8,"2x NVIDIA RTX 3090","11 月 4 日 20:41 UTC，club-3090 #1944 发布一份双 RTX 3090 同机 A\u002FB：QEngine 1.3.4.3 Flash-Next 对照日常使用的 vLLM 27B dual-fast。作者按 27B、QEngine、27B 的顺序复测，值得看的不只是快了多少，还有长时间负载后的资源余量。它是一份社区自报，不是 AraoAI 独立复现。\n\n1. 条件必须一起读。两卡功耗上限为 260\u002F280W，平台为 Ryzen 7 7700X、124GiB 内存，驱动 615.71.10，驱动 P2P 被拒绝；vLLM 镜像为 v0.32.0。报告称两臂启用思考、effort 为 low，吞吐测试使用三次预热和五次测量。但两臂模型与运行时都不同，不能称为“同模型换引擎”的受控实验。\n\n2. 收益并不一致。原文叙事生成 decode 为 100.16 对 151.56 tok\u002Fs，代码生成则为 132.72 对 139.07；10K 预填充为 2520 对 2328，90K 为 1738 对 2949，后者仅一次测量。数字说明特定工作负载的变化，不支持“一切更快”。任务集每臂只跑一轮，132\u002F150 与 128\u002F150 也不足以宣布模型能力排名。\n\n3. 能启动不等于能长期共存。作者报告 QEngine 就绪时主机内存使用 98GiB，任务及跑分后到 117GiB，可用仅 7.5GiB；维护者确认缓存归档会增长，并提醒内存余量机制保护引擎，不保证其他服务有空间。原文也明确未做完整 soak，不能据此宣布长时稳定。\n\n商业部署还有硬门槛：核对到的 QEngine LICENSE 是 PolyForm Noncommercial 1.0.0，不是 Apache-2.0，不能默认可用于商业服务。选型时把模型、引擎、任务形态、持续资源和许可一起看，比只盯最高 decode 数字更有用。\n\n来源：[报告及维护者回复](https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1944)；[QEngine LICENSE](https:\u002F\u002Fgithub.com\u002Fbpx03\u002FQEngine\u002Fblob\u002Fmain\u002FLICENSE)。","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.\n\n1. Read the conditions together. The two RTX 3090s were capped at 260\u002F280 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.\n\n2. Gains were workload-dependent. Narrative decode was 100.16 versus 151.56 tok\u002Fs; 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\u002F150 versus 128\u002F150 cannot establish a model capability ranking.\n\n3. 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.\n\nCommercial 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.\n\nSources: [report and maintainer response](https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1944); [QEngine LICENSE](https:\u002F\u002Fgithub.com\u002Fbpx03\u002FQEngine\u002Fblob\u002Fmain\u002FLICENSE).","club-3090 社区自报及维护者回复","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1944","2026-11-04T23:05:17.134Z"]