[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fkvem5b8ee516":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-vllm-reasoning-stream-timing-correction-2026-11-22","NEWS","2026-11-21T23:04:43.674Z","社区跑分工具漏读reasoning字段：首token计时错误，也能造出异常高吞吐","A missing reasoning field can distort TTFT and decode rates in community benchmarks","club-3090的#2132报告vLLM推理流字段适配缺口，可能把思考耗时算进prefill。提案仍开放；同次启动前后跑分不是引擎加速对照。","club-3090 PR #2132 reports a reasoning-stream compatibility gap affecting TTFT and derived rates on vLLM. The proposal remains open; successive benches on one boot are not an engine speedup study.",[12],{"id":13,"name":14},100,"Inkling-2-Mini",[],"11月21日提交的club-3090 PR #2132指出：跑分工具如果只读reasoning_content、漏读reasoning，就可能错过推理模型的首个流式事件，把思考耗时记进prefill，再把剩余很短的回答阶段算成异常高的decode。这里变的是计时边界，不是显卡突然加速。提案目前仍开放，本站未独立复现。\n\n1. 字段兼容也是测量协议的一部分。该报告称vLLM自0.34起使用reasoning作为输出字段，而工具若干读取点只识别reasoning_content。修补范围包括bench、长上下文探测和压力测试。不能把这些端点都叫OpenAI兼容，就默认流式扩展字段完全一致；适配器版本应与引擎版本一起留档。\n\n2. 先有思考，后有答案，不应把前者吞掉。作者在双3090、Inkling-2-Mini与vLLM的同次启动上顺序跑了修补前后工具，报告旧工具读出数秒TTFT及异常decode，新工具的计时与端到端速率更接近。作者明确说这不是速度提升实验。本文不将其中任一吞吐写入硬件选型榜，也不从工具修补幅度推导引擎性能收益。\n\n3. 关闭开关也要检查行为。报告同时指出，模型可以接受enable_thinking:false并最终回答OK，却依然先输出推理。只看HTTP成功或最终内容，会让预检误判“已关思考”，继而把过小的输出预算导致的失败归咎于模型。开关探测需同时检查推理字段、答案字段和实际请求形状。\n\n修补还包含模拟服务测试：这些测试证明的是工具在设定场景下如何处理流，不是新增真机吞吐。作者明确没有运行SGLang和llama.cpp的现场回归，不能据此写成跨框架独立验证。对已有跑分，优先保留原始流和测量脚本版本；无法重建计时边界的异常数字先降级，不靠“看起来合理”补救。\n\n来源：https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F2132","club-3090 PR #2132, opened November 21, describes a benchmark compatibility gap: a reader accepting reasoning_content but not reasoning can miss the first streamed reasoning event. Reasoning time is then charged to prefill, while a short remaining answer window can produce an implausibly high derived decode rate. This changes the timing boundary, not the GPU. The proposal remains open and AraoAI has not independently reproduced it.\n\n1. Stream fields belong to the measurement protocol. The report says vLLM 0.34 uses reasoning for output and identifies readers that recognized only reasoning_content. Proposed changes cover the benchmark, context probe and stress checks. An OpenAI-compatible endpoint does not guarantee identical extension fields; preserve the adapter version as well as the engine build.\n\n2. Reasoning preceding an answer must not disappear from timing. On Inkling-2-Mini with two RTX 3090s and vLLM, the author sequentially ran the original and patched benchmark on the same boot. Old timings showed multi-second TTFT and anomalous decode estimates; patched timings agreed more closely with wall-clock rates. The author explicitly disclaims a speedup comparison. These figures are not being imported as hardware-selection records.\n\n3. An accepted off-switch is not an honored off-switch. A request using enable_thinking:false may return OK only after reasoning first. Checking HTTP success or final content alone can misclassify the model as non-reasoning, impose an insufficient output budget and blame the resulting failures on the model. Validate reasoning fields, answer fields and the actual request together.\n\nMock-server tests in the proposal exercise tool behavior; they are not hardware benchmarks. The author also explicitly did not run live SGLang or llama.cpp regressions, so this is not cross-framework reproduction. Preserve raw streams and timing-script versions. If a suspicious result's timing boundaries cannot be reconstructed, downgrade it rather than infer correctness from plausibility.\n\nSource: https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F2132","club-3090社区测量工具修补提案","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F2132"]