https://guijiagi.com/categories/%E6%8A%80%E6%9C%AF%E5%8E%9F%E7%90%86/Recent content in 技术原理 on 硅基 AGI · 智能体学习与测评Hugozh-cn本站内容采用 CC BY-NC-SA 4.0 国际许可协议授权Thu, 16 Jul 2026 11:45:00 +0800https://guijiagi.com/posts/b1-2fdfaeb0/Thu, 16 Jul 2026 11:45:00 +0800https://guijiagi.com/posts/b1-2fdfaeb0/深入探讨AI Agent长期记忆的持久化方案，从向量存储到知识图谱的完整技术选型https://guijiagi.com/posts/b1-acff1ec2/Thu, 16 Jul 2026 11:37:00 +0800https://guijiagi.com/posts/b1-acff1ec2/系统梳理思维链推理技术的发展，从基础CoT到Tree of Thoughts再到推理时Scalinghttps://guijiagi.com/posts/b1-6be79662/Thu, 16 Jul 2026 11:35:00 +0800https://guijiagi.com/posts/b1-6be79662/分析大模型Scaling Law的现状、面临的瓶颈与未来AI发展的新范式探索https://guijiagi.com/posts/b1-38dc2b9b/Thu, 16 Jul 2026 11:33:00 +0800https://guijiagi.com/posts/b1-38dc2b9b/深入解析Flash Attention的算法原理、GPU内存层次优化与实际性能表现https://guijiagi.com/posts/b1-77165612/Thu, 16 Jul 2026 11:26:00 +0800https://guijiagi.com/posts/b1-77165612/深入解析知识蒸馏技术在大模型时代的应用，涵盖响应蒸馏、特征蒸馏与Agent蒸馏https://guijiagi.com/posts/b1-91d5b13c/Thu, 16 Jul 2026 11:23:00 +0800https://guijiagi.com/posts/b1-91d5b13c/分析大模型长上下文处理的技术瓶颈，涵盖注意力机制优化、KV Cache管理、评估基准等关键问题https://guijiagi.com/posts/b1-1cb758aa/Thu, 16 Jul 2026 11:20:00 +0800https://guijiagi.com/posts/b1-1cb758aa/系统分析大模型幻觉问题的根源，并给出从模型层到应用层的完整缓解方案https://guijiagi.com/posts/b1-cabd68a1/Thu, 16 Jul 2026 11:19:00 +0800https://guijiagi.com/posts/b1-cabd68a1/深入解析Mixture of Experts架构的技术原理、训练方法和推理优化策略https://guijiagi.com/posts/b1-f85bfd84/Thu, 16 Jul 2026 11:18:00 +0800https://guijiagi.com/posts/b1-f85bfd84/系统讲解大模型训练数据全流程工程实践，涵盖数据采集、清洗、去重、质量评估等关键环节https://guijiagi.com/posts/b1-a7bddd3c/Thu, 16 Jul 2026 11:17:00 +0800https://guijiagi.com/posts/b1-a7bddd3c/深入解析AI Agent工具调用的技术演进，从OpenAI Function Calling到MCP协议的标准化进程https://guijiagi.com/posts/b1-eb414007/Thu, 16 Jul 2026 11:12:00 +0800https://guijiagi.com/posts/b1-eb414007/深度解析AI视频生成技术的核心原理、主流方案、技术挑战与商业化路径https://guijiagi.com/posts/b1-bb5b6ad2/Thu, 16 Jul 2026 11:10:00 +0800https://guijiagi.com/posts/b1-bb5b6ad2/全面分析AI芯片市场格局，涵盖NVIDIA、AMD、华为昇腾、Google TPU等关键玩家的技术路线与竞争态势https://guijiagi.com/posts/b1-4a19c4f9/Thu, 16 Jul 2026 11:09:00 +0800https://guijiagi.com/posts/b1-4a19c4f9/系统讲解多模态大模型的技术演进，从CLIP双塔到GPT-4V统一架构再到原生多模态训练https://guijiagi.com/posts/b1-fd3b570e/Thu, 16 Jul 2026 11:05:00 +0800https://guijiagi.com/posts/b1-fd3b570e/梳理AI对齐的主流技术路线，从RLHF到DPO再到宪法AI的演进逻辑与实现要点https://guijiagi.com/posts/b1-fc96905a/Thu, 16 Jul 2026 11:03:00 +0800https://guijiagi.com/posts/b1-fc96905a/Model Context Protocol协议的架构、传输机制、工具定义格式及生态现状全面解读https://guijiagi.com/posts/b1-f95dfde9/Thu, 16 Jul 2026 11:02:00 +0800https://guijiagi.com/posts/b1-f95dfde9/从基础RAG到高级RAG的完整技术路线，涵盖混合检索、重排序、查询改写、自反思等关键技术https://guijiagi.com/posts/b1-40a7d213/Thu, 16 Jul 2026 11:01:00 +0800https://guijiagi.com/posts/b1-40a7d213/系统梳理大模型推理优化的核心技术路径，涵盖KV Cache、连续批处理、投机解码等前沿方案https://guijiagi.com/posts/b1-f1c00d34/Thu, 16 Jul 2026 11:00:00 +0800https://guijiagi.com/posts/b1-f1c00d34/深入解析AI Agent记忆系统的分层架构设计，涵盖工作记忆、情景记忆、语义记忆的工程实现https://guijiagi.com/posts/b2-ff7086a1/Thu, 16 Jul 2026 10:48:00 +0800https://guijiagi.com/posts/b2-ff7086a1/分析o1之后推理增强技术的最新进展，涵盖隐式CoT、过程奖励模型与推理时搜索的工程实践https://guijiagi.com/posts/b2-743cce1d/Thu, 16 Jul 2026 10:45:00 +0800https://guijiagi.com/posts/b2-743cce1d/系统构建大模型评估框架，涵盖知识能力、推理能力、安全性、对齐度等多维度评估方法https://guijiagi.com/posts/b2-7b01cd1c/Thu, 16 Jul 2026 10:43:00 +0800https://guijiagi.com/posts/b2-7b01cd1c/系统分析大模型幻觉问题的成因机制，涵盖训练数据、解码策略、知识冲突等因素及缓解方案https://guijiagi.com/posts/b2-5c290f1f/Thu, 16 Jul 2026 10:40:00 +0800https://guijiagi.com/posts/b2-5c290f1f/深入探讨大模型预训练阶段的数据配比策略，涵盖语言比例、领域分布、课程学习与动态调整https://guijiagi.com/posts/b2-d60bc32e/Thu, 16 Jul 2026 10:36:00 +0800https://guijiagi.com/posts/b2-d60bc32e/分析大模型处理超长上下文面临的技术挑战，涵盖注意力衰减、Needle-in-Haystack测试与长上下文优化策略https://guijiagi.com/posts/b2-5052092c/Thu, 16 Jul 2026 10:32:00 +0800https://guijiagi.com/posts/b2-5052092c/系统梳理大模型上下文窗口管理的工程方案，涵盖滑动窗口、层次化摘要、选择性记忆与结构化压缩技术https://guijiagi.com/posts/b2-48526210/Thu, 16 Jul 2026 10:30:00 +0800https://guijiagi.com/posts/b2-48526210/深度解析OpenAI o1系列模型的推理时计算扩展技术，涵盖过程奖励模型、搜索算法与思维链优化https://guijiagi.com/posts/b2-703e58c9/Thu, 16 Jul 2026 10:28:00 +0800https://guijiagi.com/posts/b2-703e58c9/系统梳理大模型训练数据治理体系，涵盖数据采集、清洗、去重、质量评估与数据配比优化https://guijiagi.com/posts/b2-e2d3a168/Thu, 16 Jul 2026 10:27:00 +0800https://guijiagi.com/posts/b2-e2d3a168/深度解析AI音乐生成的技术演进，涵盖MIDI符号生成、音频扩散模型与Suno等产品的架构设计https://guijiagi.com/posts/b2-63a9c16d/Thu, 16 Jul 2026 10:25:00 +0800https://guijiagi.com/posts/b2-63a9c16d/全面分析AI生成内容检测技术，涵盖统计检测、水印方案及对抗性攻防的当前态势https://guijiagi.com/posts/b2-50735315/Thu, 16 Jul 2026 10:24:00 +0800https://guijiagi.com/posts/b2-50735315/梳理大模型函数调用技术的标准化历程，分析不同厂商方案差异及MCP协议带来的统一前景https://guijiagi.com/posts/b2-17d2f42e/Thu, 16 Jul 2026 10:22:00 +0800https://guijiagi.com/posts/b2-17d2f42e/系统梳理大模型知识蒸馏技术体系，涵盖logits蒸馏、特征蒸馏、关系蒸馏及多教师蒸馏的实践方法https://guijiagi.com/posts/b2-af6834ab/Thu, 16 Jul 2026 10:21:00 +0800https://guijiagi.com/posts/b2-af6834ab/系统梳理AI Agent规划算法的发展脉络，分析ReAct、Reflexion、LATS等规划方法的原理与适用场景https://guijiagi.com/posts/b2-39472406/Thu, 16 Jul 2026 10:19:00 +0800https://guijiagi.com/posts/b2-39472406/深入分析自适应检索增强生成技术，探讨模型如何根据问题难度动态决定检索策略https://guijiagi.com/posts/b2-97b38611/Thu, 16 Jul 2026 10:18:00 +0800https://guijiagi.com/posts/b2-97b38611/探讨知识图谱与大语言模型融合的技术方案，涵盖知识注入、推理增强与可解释性提升https://guijiagi.com/posts/b2-4eded1bd/Thu, 16 Jul 2026 10:15:00 +0800https://guijiagi.com/posts/b2-4eded1bd/从数学原理到工程实现，全面剖析MoE模型的路由机制、训练挑战与推理优化策略https://guijiagi.com/posts/b2-e74cc835/Thu, 16 Jul 2026 10:14:00 +0800https://guijiagi.com/posts/b2-e74cc835/深度对比2026年主流AI视频生成模型的技术架构、生成质量与商业应用场景https://guijiagi.com/posts/b2-b58082a6/Thu, 16 Jul 2026 10:13:00 +0800https://guijiagi.com/posts/b2-b58082a6/解析数字人技术的完整技术栈，涵盖TTS、唇形同步、表情生成与实时渲染的最新进展https://guijiagi.com/posts/b2-337d8871/Thu, 16 Jul 2026 10:12:00 +0800https://guijiagi.com/posts/b2-337d8871/从CoT到ToT再到GoT，系统梳理提示工程中的推理增强技术栈及其适用场景https://guijiagi.com/posts/b2-19902ccc/Thu, 16 Jul 2026 10:07:00 +0800https://guijiagi.com/posts/b2-19902ccc/梳理多模态大模型从早期CLIP双塔架构到GPT-4o原生多模态的技术演进路径，分析架构设计选择与工程权衡https://guijiagi.com/posts/b2-fd3b570e/Thu, 16 Jul 2026 10:05:00 +0800https://guijiagi.com/posts/b2-fd3b570e/全面梳理大模型安全对齐技术体系，涵盖RLHF、DPO、Constitutional AI及红队测试实践https://guijiagi.com/posts/b2-40a7d213/Thu, 16 Jul 2026 10:04:00 +0800https://guijiagi.com/posts/b2-40a7d213/系统梳理大模型推理优化技术栈，涵盖KV Cache、注意力优化、量化、投机解码等核心技术https://guijiagi.com/posts/b2-77f8d611/Thu, 16 Jul 2026 10:01:00 +0800https://guijiagi.com/posts/b2-77f8d611/详解Model Context Protocol协议的设计理念、核心规范及在AI Agent中的工程实现https://guijiagi.com/posts/b2-cd38df19/Thu, 16 Jul 2026 10:00:00 +0800https://guijiagi.com/posts/b2-cd38df19/深入剖析AI Agent记忆系统的三层架构设计，涵盖工作记忆、情景记忆和语义记忆的实现方案与工程实践https://guijiagi.com/posts/article-97/Mon, 13 Jul 2026 09:00:00 +0800https://guijiagi.com/posts/article-97/系统解析大模型分布式训练的并行策略，涵盖数据并行、张量并行、流水线并行及混合策略https://guijiagi.com/posts/article-95/Mon, 13 Jul 2026 08:40:00 +0800https://guijiagi.com/posts/article-95/梳理AI从纯文本到图文音视频多模态感知的演进历程，解析关键技术突破与未来方向https://guijiagi.com/posts/article-91/Mon, 13 Jul 2026 08:00:00 +0800https://guijiagi.com/posts/article-91/系统梳理大模型压缩的三大技术路线，从原理到工程落地，让大模型跑在每一台设备上https://guijiagi.com/posts/article-89/Mon, 13 Jul 2026 07:40:00 +0800https://guijiagi.com/posts/article-89/回顾从Transformer诞生到2026年的架构演进，梳理每一次关键突破的技术脉络https://guijiagi.com/posts/article-87/Mon, 13 Jul 2026 07:20:00 +0800https://guijiagi.com/posts/article-87/深入解析Prefix Cache的原理、实现与工程优化，大幅降低大模型推理的首token延迟https://guijiagi.com/posts/article-82/Mon, 13 Jul 2026 06:30:00 +0800https://guijiagi.com/posts/article-82/深入解析Continuous Batching的原理、实现与工程优化，理解vLLM/TGI等推理框架的核心技术https://guijiagi.com/posts/article-80/Mon, 13 Jul 2026 06:10:00 +0800https://guijiagi.com/posts/article-80/系统解析Rotary Position Embedding的数学原理、关键变体与工程实现，理解为什么它成为主流位置编码方案https://guijiagi.com/posts/article-71/Mon, 13 Jul 2026 04:40:00 +0800https://guijiagi.com/posts/article-71/深入解析Speculative Decoding的数学原理、工程实现和最新变体，包括投机采样、批量投机等前沿技术https://guijiagi.com/posts/article-65/Mon, 13 Jul 2026 03:40:00 +0800https://guijiagi.com/posts/article-65/回顾从BERT到GPT再到大模型时代的语言模型演进历程，梳理关键技术节点和思想脉络https://guijiagi.com/posts/article-60/Mon, 13 Jul 2026 02:50:00 +0800https://guijiagi.com/posts/article-60/系统梳理从CoT到ToT的推理范式演进，涵盖原理分析、适用场景、工程实现及前沿扩展https://guijiagi.com/posts/article-55/Mon, 13 Jul 2026 02:00:00 +0800https://guijiagi.com/posts/article-55/从标准Self-Attention到Linear Attention、Flash Attention、Sparse Attention，全面梳理注意力机制的演进脉络https://guijiagi.com/posts/article-47/Mon, 13 Jul 2026 00:40:00 +0800https://guijiagi.com/posts/article-47/从数据不平衡到跨语言迁移，多语言大模型的_core技术挑战与前沿解法https://guijiagi.com/posts/b2-42dd7e26/Mon, 13 Jul 2026 00:40:00 +0800https://guijiagi.com/posts/b2-42dd7e26/从数据不平衡到跨语言迁移，多语言大模型的_core技术挑战与前沿解法https://guijiagi.com/posts/article-41/Sun, 12 Jul 2026 23:40:00 +0800https://guijiagi.com/posts/article-41/从PagedAttention到多级缓存，KV Cache优化的技术全栈解析https://guijiagi.com/posts/b2-280c75eb/Sun, 12 Jul 2026 23:40:00 +0800https://guijiagi.com/posts/b2-280c75eb/从PagedAttention到多级缓存，KV Cache优化的技术全栈解析https://guijiagi.com/posts/article-33/Sun, 12 Jul 2026 22:20:00 +0800https://guijiagi.com/posts/article-33/从路由机制到负载均衡，全面拆解Mixture of Experts架构的设计原理https://guijiagi.com/posts/b2-0e8204a0/Sun, 12 Jul 2026 22:20:00 +0800https://guijiagi.com/posts/b2-0e8204a0/从路由机制到负载均衡，全面拆解Mixture of Experts架构的设计原理https://guijiagi.com/posts/article-27/Sun, 12 Jul 2026 21:20:00 +0800https://guijiagi.com/posts/article-27/从32K到10M+，上下文窗口扩展的技术路径、瓶颈分析与未来展望https://guijiagi.com/posts/b2-188174a6/Sun, 12 Jul 2026 21:20:00 +0800https://guijiagi.com/posts/b2-188174a6/从32K到10M+，上下文窗口扩展的技术路径、瓶颈分析与未来展望https://guijiagi.com/posts/article-21/Sun, 12 Jul 2026 20:20:00 +0800https://guijiagi.com/posts/article-21/深入剖析大模型幻觉的成因机制，从训练数据到推理过程的系统性缓解方案https://guijiagi.com/posts/b2-2909bef4/Sun, 12 Jul 2026 20:20:00 +0800https://guijiagi.com/posts/b2-2909bef4/深入剖析大模型幻觉的成因机制，从训练数据到推理过程的系统性缓解方案https://guijiagi.com/posts/article-10/Sun, 12 Jul 2026 18:30:00 +0800https://guijiagi.com/posts/article-10/深入解析Direct Preference Optimization的数学原理和实践优势，探讨它如何简化RLHF流程https://guijiagi.com/posts/b2-c244c3f2/Sun, 12 Jul 2026 18:30:00 +0800https://guijiagi.com/posts/b2-c244c3f2/深入解析Direct Preference Optimization的数学原理和实践优势，探讨它如何简化RLHF流程https://guijiagi.com/posts/article-06/Sun, 12 Jul 2026 17:50:00 +0800https://guijiagi.com/posts/article-06/从原理到实践，全面解析Transformer-XL如何突破标准Transformer的序列长度限制https://guijiagi.com/posts/b2-5c0c8534/Sun, 12 Jul 2026 17:50:00 +0800https://guijiagi.com/posts/b2-5c0c8534/从原理到实践，全面解析Transformer-XL如何突破标准Transformer的序列长度限制https://guijiagi.com/posts/loss-function-llm-training/Thu, 02 Jul 2026 11:14:00 +0800https://guijiagi.com/posts/loss-function-llm-training/从交叉熵到对比学习，系统梳理LLM训练中的各类损失函数及其作用https://guijiagi.com/posts/softmax-variants-2026/Thu, 02 Jul 2026 11:13:00 +0800https://guijiagi.com/posts/softmax-variants-2026/系统梳理Softmax及其变体的数学原理、数值稳定性与在LLM中的应用https://guijiagi.com/posts/embedding-layer-optimization/Thu, 02 Jul 2026 11:12:00 +0800https://guijiagi.com/posts/embedding-layer-optimization/解析LLM嵌入层的优化策略，从查找表到矩阵分解与动态嵌入https://guijiagi.com/posts/tokenizer-comparison-2026/Thu, 02 Jul 2026 11:11:00 +0800https://guijiagi.com/posts/tokenizer-comparison-2026/全面对比BPE、WordPiece、Unigram和SentencePiece分词算法的原理与工程取舍https://guijiagi.com/posts/paged-attention-impl-detail/Thu, 02 Jul 2026 11:10:00 +0800https://guijiagi.com/posts/paged-attention-impl-detail/深入解析vLLM中PagedAttention的实现细节，从块管理到内存共享https://guijiagi.com/posts/continuous-batching-internals/Thu, 02 Jul 2026 11:09:00 +0800https://guijiagi.com/posts/continuous-batching-internals/解析vLLM连续批处理的内部机制，如何动态调度请求提升GPU利用率https://guijiagi.com/posts/mixture-of-resolutions/Thu, 02 Jul 2026 11:08:00 +0800https://guijiagi.com/posts/mixture-of-resolutions/解析多尺度分辨率处理如何提升模型效率与性能，以及在2026年的最新进展https://guijiagi.com/posts/sequence-parallelism-explained/Thu, 02 Jul 2026 11:07:00 +0800https://guijiagi.com/posts/sequence-parallelism-explained/解析序列并行的核心原理、DeepSpeed Ulysses与Megatron SP的区别与联系https://guijiagi.com/posts/pipeline-parallelism-2026/Thu, 02 Jul 2026 11:06:00 +0800https://guijiagi.com/posts/pipeline-parallelism-2026/从GPipe的朴素流水线到PipeDream的1F1B调度，梳理流水线并行的演进与2026年方案https://guijiagi.com/posts/tensor-parallelism-deep/Thu, 02 Jul 2026 11:05:00 +0800https://guijiagi.com/posts/tensor-parallelism-deep/从原理到实现，深入解析Megatron-LM张量并行方案的设计与工程细节https://guijiagi.com/posts/gradient-checkpointing-explained/Thu, 02 Jul 2026 11:04:00 +0800https://guijiagi.com/posts/gradient-checkpointing-explained/解析梯度检查点如何以时间换空间，使超大模型训练成为可能https://guijiagi.com/posts/activation-function-survey-2026/Thu, 02 Jul 2026 11:03:00 +0800https://guijiagi.com/posts/activation-function-survey-2026/从ReLU到GELU再到SwiGLU，全面梳理激活函数的演进脉络与2026年新趋势https://guijiagi.com/posts/neural-network-normalization/Thu, 02 Jul 2026 11:02:00 +0800https://guijiagi.com/posts/neural-network-normalization/深度解析Layer Normalization、RMSNorm等归一化技术的原理、差异与演进https://guijiagi.com/posts/pruning-llm-2026/Thu, 02 Jul 2026 11:01:00 +0800https://guijiagi.com/posts/pruning-llm-2026/从幅度剪枝到稀疏训练，系统对比LLM剪枝的两大路线及2026年最新方案https://guijiagi.com/posts/distillation-llm-2026/Thu, 02 Jul 2026 11:00:00 +0800https://guijiagi.com/posts/distillation-llm-2026/从Logit蒸馏到特征对齐，全面梳理大模型知识蒸馏的方法与实践https://guijiagi.com/posts/quantization-techniques-comparison/Thu, 02 Jul 2026 10:59:00 +0800https://guijiagi.com/posts/quantization-techniques-comparison/系统对比INT4、INT8、FP8量化方案在LLM中的精度、速度与工程权衡https://guijiagi.com/posts/linear-attention-research-2026/Thu, 02 Jul 2026 10:58:00 +0800https://guijiagi.com/posts/linear-attention-research-2026/从核函数近似到状态空间模型，梳理线性注意力在2026年的最新进展https://guijiagi.com/posts/ring-attention-distributed-context/Thu, 02 Jul 2026 10:57:00 +0800https://guijiagi.com/posts/ring-attention-distributed-context/解析Ring Attention如何通过环形通信实现超长序列的分布式注意力计算https://guijiagi.com/posts/mixture-of-depths-explained/Thu, 02 Jul 2026 10:56:00 +0800https://guijiagi.com/posts/mixture-of-depths-explained/解析Mixture-of-Depth如何让模型按需调整计算深度，实现效率与性能的平衡https://guijiagi.com/posts/specul-ive-decoding-2026/Thu, 02 Jul 2026 10:55:00 +0800https://guijiagi.com/posts/specul-ive-decoding-2026/从Draft-Verify到Medusa头，系统梳理推测解码的演进与2026年最新方案https://guijiagi.com/posts/flash-attention-3-explained/Thu, 02 Jul 2026 10:54:00 +0800https://guijiagi.com/posts/flash-attention-3-explained/从Flash Attention 1到3，解析注意力计算的性能极限突破之路https://guijiagi.com/posts/kv-cache-optimization-techniques/Thu, 02 Jul 2026 10:53:00 +0800https://guijiagi.com/posts/kv-cache-optimization-techniques/从PagedAttention到量化缓存，系统梳理KV Cache的各类优化方案https://guijiagi.com/posts/moe-architecture-2026-detail/Thu, 02 Jul 2026 10:52:00 +0800https://guijiagi.com/posts/moe-architecture-2026-detail/从GShard到DeepSeek-V3，全面解析MoE架构的设计演进与工程实践https://guijiagi.com/posts/rope-rotation-position-encoding/Thu, 02 Jul 2026 10:51:00 +0800https://guijiagi.com/posts/rope-rotation-position-encoding/深入解析RoPE旋转位置编码的数学原理、长度外推方法及2026年最新改进https://guijiagi.com/posts/transformer-attention-mechanism-deep/Thu, 02 Jul 2026 10:50:00 +0800https://guijiagi.com/posts/transformer-attention-mechanism-deep/从数学原理到工程实现，全面拆解Transformer注意力机制的核心细节https://guijiagi.com/posts/inference-framework-comparison-2026/Tue, 30 Jun 2026 17:30:00 +0800https://guijiagi.com/posts/inference-framework-comparison-2026/2026年主流大模型推理框架深度对比：vLLM、SGLang、TensorRT-LLM、TGI的性能、功能与选型指南https://guijiagi.com/posts/alignment-evolution-rlhf-to-orpo/Tue, 30 Jun 2026 17:10:00 +0800https://guijiagi.com/posts/alignment-evolution-rlhf-to-orpo/大模型对齐技术演进：RLHF、DPO、ORPO、GRPO等核心算法原理、实现要点与2026年最新进展https://guijiagi.com/posts/agentic-rag-evolution-2026/Tue, 30 Jun 2026 16:50:00 +0800https://guijiagi.com/posts/agentic-rag-evolution-2026/RAG 技术全景：从基础检索增强生成到 Agentic RAG 的架构演进，对比主流方案并给出选型指南https://guijiagi.com/posts/mcp-protocol-fact-standard-2026/Tue, 30 Jun 2026 16:40:00 +0800https://guijiagi.com/posts/mcp-protocol-fact-standard-2026/MCP（Model Context Protocol）协议深度解析：工作原理、生态进展、与A2A协议的对比，以及Agent互操作性的未来https://guijiagi.com/posts/hallucination-root-cause-2026/Tue, 30 Jun 2026 16:30:00 +0800https://guijiagi.com/posts/hallucination-root-cause-2026/深入分析大模型幻觉产生的根本原因，对比2026年主流幻觉检测与缓解技术的原理、效果与适用场景https://guijiagi.com/posts/llm-quantization-int4-2026/Tue, 30 Jun 2026 16:30:00 +0800https://guijiagi.com/posts/llm-quantization-int4-2026/大模型量化技术全景：PTQ、AWQ、GPTQ、LLM.int8() 等主流量化方案原理对比，以及 INT4 推理的精度-效率权衡https://guijiagi.com/posts/agent-memory-architecture-2026/Tue, 30 Jun 2026 16:20:00 +0800https://guijiagi.com/posts/agent-memory-architecture-2026/2026年Agent记忆系统技术全景：向量数据库、知识图谱、长期记忆架构的工程实践与前沿研究https://guijiagi.com/posts/multi-agent-collaboration-2026/Tue, 30 Jun 2026 16:00:00 +0800https://guijiagi.com/posts/multi-agent-collaboration-2026/多智能体协作框架全景对比：LangGraph、CrewAI、AutoGen、MetaGPT的核心架构差异与选型指南https://guijiagi.com/posts/moe-expert-routing-load-balance/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/moe-expert-routing-load-balance/深入解析MoE混合专家模型的路由机制、负载均衡策略以及在2026年的最新进展https://guijiagi.com/posts/rlhf-alternatives-2026-dpo-grpo-simpo/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/rlhf-alternatives-2026-dpo-grpo-simpo/深入对比2026年主流的RLHF替代方案：DPO、GRPO、SimPO的技术原理、优缺点和适用场景https://guijiagi.com/posts/rlhf-alternatives-2026/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/rlhf-alternatives-2026/全面对比2026年主流RLHF替代方案，涵盖DPO、GRPO、SimPO等对齐技术的原理、优缺点和适用场景https://guijiagi.com/posts/transformer-architecture-2026-attention-to-moe-evolution/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/transformer-architecture-2026-attention-to-moe-evolution/深入解析Transformer架构在2026年的最新演进，涵盖注意力机制优化、MoE混合专家、线性注意力等前沿方向https://guijiagi.com/posts/transformer-architecture-2026/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/transformer-architecture-2026/深入解析Transformer架构在2026年的最新演进，涵盖注意力机制优化、混合专家模型、线性注意力等前沿方向https://guijiagi.com/posts/llm-alignment-constitutional-ai/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-alignment-constitutional-ai/系统梳理大模型对齐技术的发展脉络，从RLHF到Constitutional AI再到2026年最新的对齐方案https://guijiagi.com/posts/llm-hallucination-mitigation/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-hallucination-mitigation/深入分析大模型幻觉问题的根因，系统介绍从训练到推理全链路的缓解技术https://guijiagi.com/posts/llm-watermark-technology/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-watermark-technology/全面介绍大模型水印技术的原理、方案和应用，涵盖文本水印、多模态水印及其检测方法https://guijiagi.com/posts/kv-cache-management-2026-frontier/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/kv-cache-management-2026-frontier/深入分析2026年大模型推理优化的核心方案——KV Cache管理，涵盖压缩、卸载、共享等前沿技术https://guijiagi.com/posts/llm-training-data-governance-pipeline/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-training-data-governance-pipeline/全面解析2026年大模型训练数据治理体系，涵盖数据采集、清洗、去重、质量评估和持续维护https://guijiagi.com/posts/llm-training-data-governance/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-training-data-governance/系统介绍大模型训练数据治理的全链路流程，涵盖数据采集、清洗、去重、质量评估和持续优化https://guijiagi.com/posts/llm-training-stability/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-training-stability/深入分析大模型预训练过程中的稳定性问题，包括Loss Spike、梯度爆炸、训练崩溃等，并提供系统性解决方案https://guijiagi.com/posts/llm-distillation-2026/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/llm-distillation-2026/系统介绍2026年大模型知识蒸馏技术的最新进展，涵盖白盒蒸馏、黑盒蒸馏、思维链蒸馏等方法https://guijiagi.com/posts/multi-token-prediction-mtp-inference-acceleration/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/multi-token-prediction-mtp-inference-acceleration/深入解析多Token预测(MTP)技术原理、训练方法和推理加速效果，对比分析主流实现方案https://guijiagi.com/posts/multi-token-prediction/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/multi-token-prediction/深入介绍多Token预测（MTP/Multi-Token Prediction）技术，如何通过将每次生成多个token来加速大模型推理https://guijiagi.com/posts/scaling-laws-2026/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/scaling-laws-2026/深入分析2026年大模型缩放定律的现状，探讨规模收益递减的证据与反证https://guijiagi.com/posts/long-context-evolution-positional-encoding-to-ring-attention/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/long-context-evolution-positional-encoding-to-ring-attention/全面梳理大模型长上下文技术的演进路线，涵盖位置编码外推、分块注意力、环形注意力等前沿方案https://guijiagi.com/posts/long-context-techniques-2026/Tue, 30 Jun 2026 09:20:00 +0800https://guijiagi.com/posts/long-context-techniques-2026/系统梳理大模型长上下文处理技术的演进历程，从位置编码方案到环形注意力机制https://guijiagi.com/posts/ai-inference-optimization-2026/Tue, 30 Jun 2026 09:10:00 +0800https://guijiagi.com/posts/ai-inference-optimization-2026/2026年大模型推理优化全攻略：FP8/INT4量化、结构化/动态剪枝、知识蒸馏、推测解码、投机采样等核心技术的深度解析与实战对比https://guijiagi.com/posts/llm-interpretability-research/Sun, 28 Jun 2026 11:19:00 +0800https://guijiagi.com/posts/llm-interpretability-research/系统解析大模型可解释性研究的方法、工具与2026年最新进展，探讨机械可解释性与注意力分析https://guijiagi.com/posts/world-models-ai-physics-understanding/Sun, 28 Jun 2026 11:18:00 +0800https://guijiagi.com/posts/world-models-ai-physics-understanding/深入探讨AI世界模型的最新研究进展，涵盖物理直觉、因果推理、具身智能与2026年前沿方向https://guijiagi.com/posts/test-time-compute-theory-practice/Sun, 28 Jun 2026 11:17:00 +0800https://guijiagi.com/posts/test-time-compute-theory-practice/深入解析Test-time Compute的原理、方法与实践，探讨推理时计算Scaling的新范式https://guijiagi.com/posts/training-data-cleaning-pipeline/Sun, 28 Jun 2026 11:16:00 +0800https://guijiagi.com/posts/training-data-cleaning-pipeline/系统解析大模型训练数据清洗的全流程：去重、质量过滤、安全过滤、数据配比与2026年最佳实践https://guijiagi.com/posts/scaling-laws-2026-status/Sun, 28 Jun 2026 11:15:00 +0800https://guijiagi.com/posts/scaling-laws-2026-status/深入分析2026年大模型Scaling Laws的最新进展，探讨计算最优、数据墙、突发Scaling等关键问题https://guijiagi.com/posts/emergent-abilities-llm/Sun, 28 Jun 2026 11:14:00 +0800https://guijiagi.com/posts/emergent-abilities-llm/系统分析大模型涌现能力的现象、机制与临界规模，探讨能力突现的数学原理https://guijiagi.com/posts/multimodal-fusion-architectures/Sun, 28 Jun 2026 11:13:00 +0800https://guijiagi.com/posts/multimodal-fusion-architectures/深入对比多模态大模型的三种融合架构：Early Fusion、Late Fusion、Cross-Attention的原理与2026年实践https://guijiagi.com/posts/knowledge-distillation-teacher-student/Sun, 28 Jun 2026 11:12:00 +0800https://guijiagi.com/posts/knowledge-distillation-teacher-student/深入解析大模型知识蒸馏的原理、方法与2026年最佳实践，涵盖logit蒸馏、特征蒸馏、序列级蒸馏https://guijiagi.com/posts/quantization-principles-int4-gptq-awq/Sun, 28 Jun 2026 11:11:00 +0800https://guijiagi.com/posts/quantization-principles-int4-gptq-awq/深入解析大模型量化的数学原理，涵盖INT4/INT8量化、GPTQ、AWQ算法及其工程实现https://guijiagi.com/posts/llm-training-pipeline-pretrain-sft-rlhf-dpo/Sun, 28 Jun 2026 11:10:00 +0800https://guijiagi.com/posts/llm-training-pipeline-pretrain-sft-rlhf-dpo/系统解析大模型训练的四个阶段：预训练、监督微调、RLHF、DPO的原理、流程与2026年最佳实践https://guijiagi.com/posts/ring-attention-million-token-context/Sun, 28 Jun 2026 11:09:00 +0800https://guijiagi.com/posts/ring-attention-million-token-context/深入解析Ring Attention的分布式注意力计算原理、Flash Attention集成与超长上下文实现https://guijiagi.com/posts/continuous-batching-vllm/Sun, 28 Jun 2026 11:08:00 +0800https://guijiagi.com/posts/continuous-batching-vllm/深入解析连续批处理（Continuous Batching）的原理、vLLM实现细节与2026年高吞吐推理系统设计https://guijiagi.com/posts/speculative-decoding-principles/Sun, 28 Jun 2026 11:07:00 +0800https://guijiagi.com/posts/speculative-decoding-principles/深入解析投机解码的算法原理、Draft Model 设计、接受率优化与2026年最新工程实践https://guijiagi.com/posts/moe-architecture-evolution/Sun, 28 Jun 2026 11:06:00 +0800https://guijiagi.com/posts/moe-architecture-evolution/系统梳理MoE混合专家架构从Mixtral到DeepSeek V4的演进历程，解析路由机制、负载均衡与工程优化https://guijiagi.com/posts/flash-attention-3-principles/Sun, 28 Jun 2026 11:05:00 +0800https://guijiagi.com/posts/flash-attention-3-principles/深入解析Flash Attention 3的算法原理、GPU硬件协同优化与2026年最新性能基准https://guijiagi.com/posts/kv-cache-principles/Sun, 28 Jun 2026 11:04:00 +0800https://guijiagi.com/posts/kv-cache-principles/深入解析KV Cache的工作原理、内存计算、优化策略与2026年最新技术进展https://guijiagi.com/posts/tokenizer-principles-and-practice/Sun, 28 Jun 2026 11:03:00 +0800https://guijiagi.com/posts/tokenizer-principles-and-practice/深入解析BPE、WordPiece、Unigram、SentencePiece四种分词算法的原理、优缺点与2026年工程实践https://guijiagi.com/posts/positional-encoding-comparison/Sun, 28 Jun 2026 11:02:00 +0800https://guijiagi.com/posts/positional-encoding-comparison/深入对比RoPE旋转位置编码、ALiBi线性偏置、NoPE无位置编码三种方案的数学原理、长文本外推能力与实测性能https://guijiagi.com/posts/attention-mechanisms-panorama/Sun, 28 Jun 2026 11:01:00 +0800https://guijiagi.com/posts/attention-mechanisms-panorama/系统解析2026年主流注意力机制变体：Self-Attention、Cross-Attention、Multi-Query、Grouped-Query、Latent Attention的原理、数学推导与工程实现https://guijiagi.com/posts/transformer-architecture-2026-evolution/Sun, 28 Jun 2026 11:00:00 +0800https://guijiagi.com/posts/transformer-architecture-2026-evolution/深入解析Transformer架构在2026年的最新演进，涵盖Attention机制优化、MoE混合专家架构、Mamba状态空间模型三大方向https://guijiagi.com/posts/edge-ai-2026-mobile-iot-automotive-large-models/Sun, 28 Jun 2026 10:18:00 +0800https://guijiagi.com/posts/edge-ai-2026-mobile-iot-automotive-large-models/2026 年边缘 AI 发展全景：大模型在手机、IoT 设备、汽车上的部署技术、应用场景和产业格局https://guijiagi.com/posts/kv-cache-optimization-pagedattention-to-mla/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/kv-cache-optimization-pagedattention-to-mla/大模型KV Cache优化技术全景解析：PagedAttention、GQA、MQA、MLA原理与实测对比https://guijiagi.com/posts/speculative-decoding-practical-3x-speedup-guide/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/speculative-decoding-practical-3x-speedup-guide/Speculative Decoding投机解码实战配置指南：原理详解、Draft模型选择与3倍加速实测https://guijiagi.com/posts/model-quantization-2026-int4-int8-awq-gptq-comparison/Sun, 28 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/model-quantization-2026-int4-int8-awq-gptq-comparison/2026年主流模型量化技术全面实测对比：INT4/INT8/AWQ/GPTQ质量损失与推理效率分析https://guijiagi.com/posts/kv-cache-optimization-2026/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/kv-cache-optimization-2026/全面解析KV Cache优化策略，从内存管理到压缩淘汰的完整技术方案https://guijiagi.com/posts/kv-cache-optimization/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/kv-cache-optimization/KV Cache优化策略详解https://guijiagi.com/posts/transformer-alternatives-survey/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/transformer-alternatives-survey/Transformer替代架构Surveyhttps://guijiagi.com/posts/llm-safety-guardrails/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/llm-safety-guardrails/大模型安全护栏实现机制https://guijiagi.com/posts/hallucination-root-cause/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/hallucination-root-cause/大模型幻觉问题根因分析https://guijiagi.com/posts/llm-inference-acceleration-2026/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/llm-inference-acceleration-2026/大模型推理加速技术全景解析，从KV Cache到投机解码的完整技术栈https://guijiagi.com/posts/llm-inference-acceleration-survey/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/llm-inference-acceleration-survey/大模型推理加速技术全景https://guijiagi.com/posts/emergent-ability-mechanism/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/emergent-ability-mechanism/大模型涌现能力机理研究https://guijiagi.com/posts/pretraining-data-cleaning/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/pretraining-data-cleaning/大模型预训练数据清洗方法论https://guijiagi.com/posts/alignment-tax-analysis/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/alignment-tax-analysis/对齐税Alignment Tax原理分析https://guijiagi.com/posts/multimodal-fusion-architecture-2026/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/multimodal-fusion-architecture-2026/深度解析多模态融合架构原理，从早期融合到晚期融合再到混合架构的技术演进https://guijiagi.com/posts/multimodal-fusion-architecture/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/multimodal-fusion-architecture/多模态融合架构原理深度解析https://guijiagi.com/posts/function-calling-internals/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/function-calling-internals/函数调用Function Calling底层原理https://guijiagi.com/posts/moe-architecture-analysis/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/moe-architecture-analysis/混合专家模型MoE架构剖析https://guijiagi.com/posts/model-quantization-principles/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/model-quantization-principles/模型量化技术原理与实践https://guijiagi.com/posts/model-quantization-techniques-2026/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/model-quantization-techniques-2026/深入解析大模型量化技术原理，从对称量化到混合精度的完整实践指南https://guijiagi.com/posts/model-distillation-survey/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/model-distillation-survey/模型蒸馏技术全景对比https://guijiagi.com/posts/rlhf-technique-deep-dive/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/rlhf-technique-deep-dive/强化学习RLHF技术原理详解https://guijiagi.com/posts/inference-scaling-laws/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/inference-scaling-laws/推理时计算Scaling Lawshttps://guijiagi.com/posts/long-context-evolution/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/long-context-evolution/长上下文窗口技术演进路线https://guijiagi.com/posts/attention-mechanism-variants-2026/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/attention-mechanism-variants-2026/系统对比分析主流注意力机制变体，从标准Self-Attention到Flash Attention的技术演进https://guijiagi.com/posts/attention-variant-comparison/Sat, 27 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/attention-variant-comparison/注意力机制变体全面对比分析https://guijiagi.com/posts/mcp-protocol-deep-dive/Fri, 26 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/mcp-protocol-deep-dive/深入剖析 Model Context Protocol (MCP) 的架构设计、通信机制、工具调用流程，并通过代码示例展示如何从零实现一个 MCP Server 与 Client。https://guijiagi.com/posts/react-vs-plan-execute/Fri, 26 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/react-vs-plan-execute/深入对比 ReAct 与 Plan-and-Execute 两种主流智能体推理范式的工作原理、优劣势及适用场景，指导工程选型。https://guijiagi.com/posts/agent-memory-vector-db/Fri, 26 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/agent-memory-vector-db/从实际 Agent 记忆系统需求出发，深入对比 Pinecone、Milvus、Qdrant 等主流向量数据库的架构、性能与适用场景。https://guijiagi.com/posts/agent-streaming-response/Fri, 26 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/agent-streaming-response/深入探讨 AGI 智能体的流式响应架构，从 SSE 到 WebSocket，从模型骨架层到用户交互层的完整设计实录。https://guijiagi.com/posts/flash-attention-3/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/flash-attention-3/深度解析 Flash Attention 3 如何通过异步执行、warp 专用化和 FP8 支持，在 H100 GPU 上实现比 Flash Attention 2 快 1.5-2 倍的性能。https://guijiagi.com/posts/llm-context-length/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/llm-context-length/系统梳理大语言模型上下文窗口扩展的技术方案，涵盖位置插值、NTK-aware、YaRN、LongRoPE 等方法及其工程实现。https://guijiagi.com/posts/mixture-of-depths/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/mixture-of-depths/深度解析 Mixture of Depths (MoD) 如何通过动态跳过冗余 Transformer 层，在保持性能的同时大幅提升推理效率。https://guijiagi.com/posts/moe-architecture-deep/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/moe-architecture-deep/深入剖析 Mixture of Experts 架构的原理、路由机制、训练策略及工程实现，涵盖 GShard、Switch Transformer、Mixtral 等代表性工作。https://guijiagi.com/posts/ring-attention-long-context/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/ring-attention-long-context/深度解析 Ring Attention 如何通过环形通信将注意力计算分布到多 GPU，实现百万级 token 上下文窗口。https://guijiagi.com/posts/rlhf-dpo-grpo/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/rlhf-dpo-grpo/从 RLHF 的 PPO 到 DPO 的无奖励模型，再到 GRPO 的群组相对策略优化，深度解析三种主流 LLM 对齐算法的原理与优劣。https://guijiagi.com/posts/tokenizer-comparison/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/tokenizer-comparison/深入对比四种主流分词算法的原理、训练方法、优缺点及工程实现，涵盖 BPE、WordPiece、Unigram Language Model 和 SentencePiece。https://guijiagi.com/posts/transformer-alternatives-2026/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/transformer-alternatives-2026/深度解析 Mamba、S4/S6、RWKV 等替代 Transformer 的新一代架构，从状态空间模型到线性注意力，全面对比优劣。https://guijiagi.com/posts/quantization-techniques/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/quantization-techniques/系统梳理大语言模型量化技术，涵盖 PTQ 与 QAT、对称与非对称量化、GPTQ、AWQ、SmoothQuant 等方法的原理与实现。https://guijiagi.com/posts/training-stability/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/training-stability/深入分析大语言模型训练中的稳定性问题，涵盖梯度爆炸/消失、Loss Spike、数值不稳定等原因诊断与解决方案。https://guijiagi.com/posts/continuous-batching/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/continuous-batching/深度解析 vLLM 连续批处理（Continuous Batching）如何通过动态请求调度和 PagedAttention 实现高吞吐 LLM 推理服务。https://guijiagi.com/posts/tokenizer-special-tokens/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/tokenizer-special-tokens/梳理大语言模型特殊标记（Special Tokens）的设计演进，从 BERT 到 ChatML，再到现代模型的统一 Prompt 格式。https://guijiagi.com/posts/speculative-decoding-explained/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/speculative-decoding-explained/从 Speculative Decoding 到 Medusa 和 EAGLE，全面解析投机解码如何在不损失质量的前提下将 LLM 推理速度提升 2-4 倍。https://guijiagi.com/posts/positional-encoding/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/positional-encoding/系统梳理 Transformer 位置编码的演进历程，深入分析 Sinusoidal、Learned、Relative、RoPE、ALiBi 等方案的数学原理与工程实现。https://guijiagi.com/posts/attention-mechanism-evolution/Thu, 25 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/attention-mechanism-evolution/全面梳理注意力机制从 2014 年到 2025 年的演进历程，涵盖 Bahdanau Attention、Self-Attention、Multi-Head、Linear Attention、Flash Attention 等关键技术。https://guijiagi.com/posts/attention-mechanism-explained/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/attention-mechanism-explained/深入解析 Transformer Attention 机制的数学原理、变体演进与复杂度分析https://guijiagi.com/posts/flash-attention-3-guide/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/flash-attention-3-guide/全面解析 Flash Attention 1/2/3 的演进、IO 复杂度分析及性能数据https://guijiagi.com/posts/layer-normalization-deep/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/layer-normalization-deep/深入对比 LayerNorm 与 RMSNorm 的原理、数值稳定性与性能差异https://guijiagi.com/posts/mixture-of-experts-internals/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/mixture-of-experts-internals/深入解析 Mixture-of-Experts 架构的路由机制、负载均衡策略与工程优化https://guijiagi.com/posts/ring-attention-explained/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/ring-attention-explained/深入解析 Ring Attention 如何实现百万级 Token 的长上下文训练与推理https://guijiagi.com/posts/tokenizer-internals/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/tokenizer-internals/深入解析主流分词算法的原理、训练过程与工程实现https://guijiagi.com/posts/mixture-of-experts-deep-dive/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/mixture-of-experts-deep-dive/全面解析 MoE 架构原理、路由算法、负载均衡及演进历程https://guijiagi.com/posts/activation-function-evolution/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/activation-function-evolution/梳理 LLM 激活函数的演进历程、数学原理与性能对比https://guijiagi.com/posts/context-window-extension/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/context-window-extension/系统梳理 LLM 上下文窗口扩展的核心技术路线与工程实践https://guijiagi.com/posts/speculative-decoding-deep/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/speculative-decoding-deep/深入解析投机解码原理、Medusa/EAGLE/Lookahead 等加速方案https://guijiagi.com/posts/positional-encoding-guide/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/positional-encoding-guide/系统解析 Transformer 位置编码的演进脉络与长度外推技术https://guijiagi.com/posts/rope-rotation-embedding/Thu, 25 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/rope-rotation-embedding/深入解析 RoPE 数学原理及主流长度外推方法 NTK/YaRN/Longropehttps://guijiagi.com/posts/kv-cache-guide/Wed, 24 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/kv-cache-guide/深入解析 KV Cache 的工作原理、内存计算与优化技术，包括 PagedAttention、量化 KV Cache 等https://guijiagi.com/posts/transformer-architecture-deep/Wed, 24 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/transformer-architecture-deep/深入剖析 Transformer 架构的每个组件，从自注意力机制到 GPT 解码器-only 架构的演进https://guijiagi.com/posts/tokenization-guide/Wed, 24 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/tokenization-guide/从字符级到子词级分词，深入解析 BPE、WordPiece、SentencePiece 算法及工程实践https://guijiagi.com/posts/speculative-decoding/Wed, 24 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/speculative-decoding/深入解析推测解码如何通过 Draft Model 并行验证加速 LLM 自回归推理https://guijiagi.com/posts/attention-mechanism-guide/Wed, 24 Jun 2026 15:00:00 +0800https://guijiagi.com/posts/attention-mechanism-guide/全面解析注意力机制的数学原理与工程优化，从基础公式到 Flash Attention 的 IO 感知设计https://guijiagi.com/posts/llm-inference-optimization/Wed, 24 Jun 2026 11:30:00 +0800https://guijiagi.com/posts/llm-inference-optimization/LLM 推理性能优化的完整技术栈，涵盖显存、计算和通信优化https://guijiagi.com/posts/agent-tool-use-guide/Wed, 24 Jun 2026 11:00:00 +0800https://guijiagi.com/posts/agent-tool-use-guide/系统梳理 Agent 工具使用的演进路径、设计模式和工程实践https://guijiagi.com/posts/agent-loop-design/Tue, 23 Jun 2026 14:20:00 +0800https://guijiagi.com/posts/agent-loop-design/深入分析 Agent 循环设计模式，从经典 ReAct 到现代 Plan-and-Executehttps://guijiagi.com/posts/react-pattern/Thu, 18 Jun 2026 14:00:00 +0800https://guijiagi.com/posts/react-pattern/ReAct（Reasoning + Acting）是 Agent 最经典的推理范式。本文从原理拆解到代码实现，深入分析 ReAct 的优势、局限和进化方向。https://guijiagi.com/posts/tool-use-evolution/Mon, 15 Jun 2026 10:00:00 +0800https://guijiagi.com/posts/tool-use-evolution/LLM 本身只能生成文本。工具调用赋予了 AI 操作真实世界的能力。本文梳理从 Function Calling 到 MCP 的完整技术演进。https://guijiagi.com/posts/multimodal-agent/Sat, 13 Jun 2026 12:00:00 +0800https://guijiagi.com/posts/multimodal-agent/2026 年的多模态 Agent 不再是「文本+图片」的简单组合，而是文本、图像、音频、视频、代码的统一理解与生成。https://guijiagi.com/posts/mcp-protocol/Fri, 12 Jun 2026 16:00:00 +0800https://guijiagi.com/posts/mcp-protocol/Model Context Protocol 正在统一 AI 工具调用的接口标准。本文深入解析 MCP 的设计理念、协议细节和生态现状。