Welcome to LLM Systems Architecture, a technical experimental node observing the rapid evolution of Artificial Intelligence and Large Language Model architectures.
🌐 Live Blog: blog.llmgo.top
🌐 English Portal: blog.llmgo.top/en/
📊 2026 AI Models Benchmark Matrix: Frontier Models
以底层硬件第一性原理贯穿现代大模型推理系统：从 Roofline 模型推导、算子加速，到 PagedAttention、MLA、MoE 并行与 P/D 解耦池化。
- Chapter 01 · Roofline 模型与 Prefill/Decode 物理撕裂
👉 English: Roofline Modeling of Prefill vs. Decode Disruption - Chapter 02 · PagedAttention 显存池管理与分页虚拟化
👉 English: PagedAttention Memory Virtualization & Block Management - Chapter 03 · Continuous Batching 与 Chunked Prefill 消除排头阻塞
👉 English: Continuous Batching & Chunked Prefill Deep Dive - Chapter 04 · 前缀缓存演进：从 Hash 寻址到 SGLang RadixAttention 树状缓存
👉 English: Prefix Caching & RadixAttention Internals - Chapter 05 · Attention 算子加速史：FlashAttention-1/2/3 到 FlashInfer 统一异构核心
👉 English: FlashAttention to FlashInfer Kernel Evolution - Chapter 06 · 模型架构反哺推理系统：MHA/GQA 到 DeepSeek MLA 矩阵吸收
👉 English: MHA/GQA to DeepSeek MLA Matrix Absorption - Chapter 07 · 稀疏大模型 MoE 推理内核：专家并行 (EP) 与 All-to-All 通信重叠
👉 English: MoE Expert Parallelism & All-to-All Overlap - Chapter 08 · P/D 分离架构：计算与访存解耦、RDMA 分布式 KV Cache 传输与集群池化
👉 English: Prefill-Decode Disaggregation Architecture & RDMA KV Cache - Chapter 09 · 四大主流生产级推理引擎架构横评与调度器源码解密：vLLM v1 vs SGLang vs TRT-LLM vs llama.cpp
👉 English: Production Inference Engines Teardown: vLLM v1 vs SGLang vs TRT-LLM vs llama.cpp - Chapter 10 (终局) · 长思维链 (Reasoning / Test-Time Compute) 与大集群推理系统的未来终局
👉 English: Test-Time Compute, Reasoning Scheduling, and Tiered Cluster Storage
系统掌握 2026 生产级 AI Agent 核心架构：从执行循环、状态机，到 MCP 协议、Skills 扩展、多智能体协作与全链路可观测性。
- Agent 执行循环与状态机设计：从 ReAct 到确定性控制流 (👉 English)
- MCP 协议完全指南：连接 LLM 与本地世界的统一标准 (👉 English)
- 突破 10 万 Star 的 Browser-use 架构深度剖析 (👉 English)
- AI Agent 记忆系统设计：从工作记忆到分层 Graph RAG (👉 English)
- 对抗复合误差的分布式 Agent 编排与量化 Evals 体系 (👉 English)
- AI 编程驾驭指南：从「帮我写个 XX」到架构编排者 (👉 English)
- Test-Time Compute 与 GRPO 强化学习全景拆解：从自我反思到长思维链落地 (👉 English)
- SGLang vs vLLM 架构全景深度对比：RadixAttention、PageAttention 与吞吐极限评测 (👉 English)
- 投机解码 (Speculative Decoding) 与 EAGLE 工业级落地实践 (👉 English)
- 生产级 RAG 检索增强全景演进：向量数据库、混合检索与重排序优化 (👉 English)
I am ifnodoraemon, a developer and AI researcher passionate about exploring the capabilities and boundaries of modern LLMs.
- 🌐 Blog: blog.llmgo.top
- 🐙 GitHub: @ifnodoraemon
- 𝕏 X (Twitter): @ifnodoraemon
- ✉️ Email: ifnodoraemon@gmail.com
This site is built with a focus on speed, minimalism, and a sci-fi cyber aesthetic:
- Vite for ultra-fast generation and bundling.
- Handlebars for lightweight static templating.
- Marked & KaTeX for high-speed Markdown parsing and LaTeX math formula rendering.
- Native CSS variables for consistent, scalable theming without heavy frameworks.
- Automated Technical SEO & Sitemap Auditing with Google Search Console API integration.
This repository is licensed under the MIT License.