[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1i64tcpt0jiry":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":15,"bodyZh":16,"bodyEn":17,"sourceName":18,"sourceUrl":19,"updatedAt":6},"club3090-single-output-chunk-decode-metric-fix-2026-10-31","ARENA","2026-10-30T23:08:06.090Z","社区测量修正：工具调用一次吐完，不能算成每秒几十万 token","Community benchmark fix: a buffered tool call has no observable decode window","club-3090 在10月30日合并两项测量修正：一次输出完整工具调用的响应不再报告 decode，客户端与引擎速度按同类请求、同测量窗口对齐。异常数字来自计时口径，不是硬件突破。","Two club-3090 fixes merged on October 30 remove decode rates for whole-reply output chunks and align client\u002Fserver comparisons by request shape and measurement window. The extreme rates were timing artifacts, not hardware gains.",[12],{"id":13,"name":14},5,"Qwen3.8-27B",[],"几十万 token\u002Fs，先别庆祝。10月30日，club-3090 合并 PR #1864：当工具调用整段缓冲后一次发出时，bench-agentic 不再把它算成 decode 速度。对应社区报告使用 QEngine 和 Qwen3.8-27B；这次变化修的是测量器，不是模型或显卡突然变快。\n\n原文探测先收到 role 空内容块，随后是完整 tool_calls、结束块和 usage。旧检查只数 choices 块，未区分实际输出，便把完整调用与结束标记间的极短间隔当生成窗口，用整段 token 数相除。那些 token 并非在这段间隔内才生成。\n\n修正后，脚本统计实际承载输出的块。若整个回复只在一个输出块里出现，decode 标为不可测，并明确说明 TTFT 包含了整段生成时间；墙钟吞吐仍可另报，但不能换个名字冒充纯解码。PR 作者报告自己的同配方复查也出现同类现象。这是社区报告及维护者复查，不是 AraoAI 真机复现，也不足以外推全部引擎。\n\n同日合并的 PR #1865 修正另一类误报：客户端把 prose 与 code 的均值混合，引擎侧却取包含预热、深上下文探针的全窗口中位数。两边比较的并不是同一组请求。新逻辑按任务形状分别比较，限定各自正式测量的日志窗口，并对两侧使用一致的中位数口径。\n\n对选型者，结论不是“这张卡不可信”，而是先审计数字的分子与分母：输出是否逐步可见，TTFT 到底止于首个可见输出还是完整回复，两侧是否包含相同请求。原始硬件报告还明确是 RTX 3090 与 RTX 3090 Ti 混卡、限功耗、不同 PCIe 宽度，不能简写成标准双3090成绩。\n\n许可也要分开：club-3090 标为 Apache-2.0，而本次核对的 QEngine LICENSE 为 PolyForm Noncommercial 1.0.0，不可当成任意商用的开源引擎。\n\n来源：#1864 https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1864 ；#1865 https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1865 ；原报告 https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1862 ；许可 https:\u002F\u002Fgithub.com\u002Fbpx03\u002FQEngine\u002Fblob\u002Fmain\u002FLICENSE 。AI 辅助整理。","An apparent rate of hundreds of thousands of tokens per second calls for a timing audit, not celebration. On October 30, club-3090 merged PR #1864 so bench-agentic no longer reports decode throughput when an entire tool call arrives in one output-bearing chunk. The associated community report used QEngine and Qwen3.8-27B. This changes the measurement harness, not model or GPU performance.\n\nA direct probe in the PR shows an empty role chunk, a complete tool_calls chunk, then a finish chunk and usage. The old guard counted choices chunks without distinguishing actual output. It therefore treated the tiny interval from receiving the complete call to receiving the finish marker as a generation window, and divided the full token count by that interval. Those tokens were not necessarily generated inside that window.\n\nThe fix counts output-bearing chunks. A whole reply delivered in one chunk has no observable decode window, so its decode rate is marked unavailable. The warning also explains that its reported TTFT includes generation of the whole reply. Wall-clock throughput can still be reported separately, but is not pure decode throughput. The maintainer reports reproducing the same measurement shape on another rig using the recipe. This is community reporting and maintainer checking, not an AraoAI hardware reproduction or a conclusion about every engine.\n\nPR #1865, merged the same day, addresses a different false warning. The client combined prose and code means while the engine comparison used a pooled median including warm-ups and deep-context probes. The requests did not match. The new comparison separates workload shapes, uses their measured-run log windows and compares medians on both sides.\n\nBefore using a number for selection, check whether output is incrementally visible, what the TTFT endpoint represents and whether both measurements include the same requests. The source rig is explicitly a power-capped RTX 3090 plus RTX 3090 Ti with different PCIe widths, not a standard pair of RTX 3090s.\n\nLicenses are separate too: club-3090 is Apache-2.0, whereas the QEngine LICENSE retrieved for this report is PolyForm Noncommercial 1.0.0. Do not treat the latter as unrestricted commercial open-source software.\n\nSources: https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1864 ; https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1865 ; https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1862 ; https:\u002F\u002Fgithub.com\u002Fbpx03\u002FQEngine\u002Fblob\u002Fmain\u002FLICENSE . AI-assisted editorial summary.","club-3090 community PRs","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1864"]