[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fbPBcLLe933n1_E35I8yfxcTH0FTrmERS23cuX1ubwT4":3},{"item":4,"related":44},{"id":5,"type":6,"title":7,"slug":8,"summary":9,"body":10,"coverUrl":11,"productScreenshots":12,"productLinks":13,"authorName":14,"authorUrl":15,"authorSubject":16,"category":17,"tags":22,"sourceLabel":35,"sourceName":35,"sourceUrl":36,"status":37,"seoTitle":7,"seoDescription":9,"canonicalUrl":38,"isFeatured":39,"sno":40,"sortOrder":41,"publishedAt":42,"updatedAt":43,"createdAt":43},"6bac9743-eceb-4902-be32-55f1b2973ebf","article","Speculative Decoding：小模型先猜几个词，大模型再逐个验？","speculative-decoding-llm-inference","投机解码让较小的草稿模型先提出一串候选 token，再由目标大模型并行验证，减少逐 token 等待的空档。本文讲清接受与拒绝机制、何时可能加速，以及它并不保证每次都更快或改变目标模型分布。","大模型通常按 token 一个接一个生成：先算出下一个 token，再把它加入上下文，继续算下一个。即使 GPU 很强，每一步也可能受到内存读写和串行依赖影响。Speculative Decoding（投机解码）尝试让一个较小的草稿模型先“猜”一串候选，再由目标大模型并行检查，减少串行等待。\n\n## 先草拟，再批量核对\n\n草稿模型依据当前上下文快速生成若干候选 token。目标大模型随后一次处理这段候选序列，判断每个位置是否接受草稿结果；如果某个位置不合适，就按算法规则从目标模型分布中重新采样，并以新的上下文继续。接受的前缀可以直接保留，避免目标模型每次只前进一步。\n\n这像是速记员先把演讲内容打一版，主讲人再快速校对整段：若前几句都对，就一次确认多句；碰到不匹配的位置，再从那里修订。大模型本身仍然是最终裁判，草稿模型并不拥有最终决定权。\n\n## 为什么可以不牺牲目标分布\n\n论文中的 speculative sampling 使用接受与拒绝采样规则，使输出分布在算法假设下与直接从目标模型采样相匹配。这里的“匹配”指统计分布层面的性质，并不意味着每次运行都会给出相同句子，也不表示草稿模型可以随便猜、结果永远不会偏差。实现必须遵循正确的采样与校正步骤。\n\n速度提升也不是保证。若草稿模型提出的 token 经常被拒绝、两模型在硬件上调度不佳，额外的草拟与验证开销可能抵消收益。序列长度、温度、模型大小、批处理、GPU 利用率和内存带宽都会影响结果。小模型与大模型越容易预测出相同候选，投机越可能划算。\n\n投机解码主要减少生成路径上的等待，并不减少用户问题本身，也不会自动让答案更准确。它改变的是“如何高效计算目标模型的下一个 token”。评估时应同时看延迟、吞吐、接受率和输出分布，而不是只看草稿模型的速度。\n\n来源：[论文：Accelerating Large Language Model Decoding with Speculative Sampling](https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.01318)","\u002Fuploads\u002F2026-10-17\u002F11154c61-dafc-47bb-87e3-fa2e14fedec4.jpg",[],[],"Foundit","https:\u002F\u002Ffoundit.cn","foundit-ai-editorial",{"id":18,"name":19,"slug":20,"description":21},"6179d3b6-dc34-4483-9ded-3cd9f1b37a47","科普","abbreviation","介绍各领域新兴概念",[23,27,31],{"id":24,"name":25,"slug":26},"a2ccffe0-49b2-458b-baf6-a83a1b20443d","大语言模型","llm",{"id":28,"name":29,"slug":30},"88d2bc27-0e0f-468a-b907-2991cb97b87b","人工智能","ai",{"id":32,"name":33,"slug":34},"4c2bbea6-eab7-40a8-8447-1de478ff7749","分析","analyse","Research paper: Speculative Sampling","https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.01318","published",null,false,53,0,"2026-10-17T00:00:00.000Z","2026-10-17T03:08:56.162Z",[45,54,64],{"id":46,"type":6,"title":47,"slug":48,"summary":49,"coverUrl":50,"authorName":14,"sno":51,"publishedAt":52,"createdAt":53},"afc03b50-8146-4a47-9596-9ef29a87ec7d","Diffusion LLM：语言模型也能从噪声里逐步生成文字吗？","diffusion-llm-masked-language-generation","Diffusion LLM 不按单一的从左到右顺序生成文字，而是通过多轮掩码恢复逐步确定 token。本文解释它与自回归模型的区别、并行生成潜力、训练目标和现实限制。","\u002Fuploads\u002F2026-09-12\u002F117f7406-bd6c-4bad-861c-5ab835ae8e91.jpg",45,"2026-09-12T00:00:00.000Z","2026-09-12T03:56:51.479Z",{"id":55,"type":6,"title":56,"slug":57,"summary":58,"coverUrl":59,"authorName":60,"sno":61,"publishedAt":62,"createdAt":63},"35a3e5ea-b651-45b6-9b82-d0c4aac1006a","如何让内容更容易被国内大模型引用","geo-in-china","通过可抓取的页面、结构化的答案、可信的证据和多平台权威分发，让大模型更容易发现、理解、信任并引用内容。","https:\u002F\u002Foxqtewbrpuiouqqjrvdv.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fpublic-media\u002F2026-07-17\u002F7d9c4c0e-eb28-4011-9b18-f7e7212eaa9f.jpg","龙家轩",49,"2026-07-17T00:00:00.000Z","2026-07-17T04:40:14.581Z",{"id":65,"type":6,"title":66,"slug":67,"summary":68,"coverUrl":69,"authorName":14,"sno":70,"publishedAt":52,"createdAt":71},"7645a21b-41c6-4aa4-a221-202846cdb37f","不做 OCR 也能搜 PDF？视觉文档检索怎么工作","colpali-visual-document-retrieval-explained","ColPali 直接把 PDF 页面作为图像进行多向量检索，保留表格、图表和版式信息。本文解释视觉文档检索、late interaction、RAG 接入方式和索引成本。","\u002Fuploads\u002F2026-09-12\u002F1adbb11c-8c10-4034-95e1-b488e3188b9b.jpg",52,"2026-09-12T03:56:48.776Z"]