[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-92394":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":15,"stars30d":16,"stars90d":15,"forks30d":15,"starsTrendScore":15,"compositeScore":17,"rankGlobal":10,"rankLanguage":10,"license":10,"archived":18,"fork":18,"defaultBranch":19,"hasWiki":20,"hasPages":18,"topics":21,"createdAt":10,"pushedAt":10,"updatedAt":22,"readmeContent":23,"aiSummary":24,"trendingCount":15,"starSnapshotCount":15,"syncStatus":25,"lastSyncTime":26,"discoverSource":27},92394,"AI-wiki","lvy010\u002FAI-wiki","lvy010","AI Full Stack: Data, Algorithms, Models, Hardware, Architecture","",null,"Rust",113,8,72,0,41,43.96,false,"main",true,[],"2026-07-22 04:02:06","#  AI-wiki\n\n- 资料整理和学习笔记\n- Interstellar: **数据、算法、模型、硬件、架构**\n\n搜索了一下，目前市面上没有比较符合思路的整理，计划9月构建完本仓库[roadmap](.\u002Froadmap.md), 好的资料\u002F笔记欢迎提交 pr，可以私信我加aiwiki交流群(｀・ω・´)\n\n\n- [Course](#course)\n  - [LLM](#llm)\n  - [Agent](#agent)\n- [Blog](#blog)\n  - [FlashAttention](#flashattention)\n  - [Harness Engineering](#harness-engineering)\n  - [Quantization](#quantization)\n  - [Speculative Decoding Blog](#speculative-decoding-blog)\n  - [CUDA](#cuda)\n- [Book](#book)\n- [Paper](#paper)\n  - [Base Model](#base-model)\n  - [Fine-tuning](#fine-tuning)\n  - [Attention Optimization](#attention-optimization)\n  - [Speculative Decoding](#speculative-decoding)\n  - [KV Cache & Inference Storage](#kv-cache--inference-storage)\n  - [Prefill-Decode Disaggregation](#prefill-decode-disaggregation)\n  - [LLM Application & Prompt Engineering](#llm-application--prompt-engineering)\n  - [MoE Mixture of Experts](#moe-mixture-of-experts)\n  - [Scheduling & Batching](#scheduling--batching)\n  - [Training Optimization & Scaling](#training-optimization--scaling)\n\n## 一、Course\n### LLM\n- [CS224n](#)\n- [CS336: Language Modeling from Scratch (Stanford \u002F Spring 2026)](#)\n- [CSCI 1390, Spring 2025: Systems for Machine Learning](#)\n\n### Agent\n- [从零开始理解 Agent](#)\n- [Learn Claude Code](#)\n\n## 2、Paper\n### 底座\n[PaLM](#)，[OPT](#)，[BLOOM](#)，[LLaMA](#)\n\n### 微调\n* 对齐微调: [InstructGPT (RLHF)](#)，[Constitutional AI](#)，[Self-Instruct](#)，[Direct Preference Optimization (DPO)](#)，[ORPO](#)，[GRPO](#)\n\n* 轻量化微调: [LoRA](#)，[QLoRA](#)\n\n### Attention 优化\n- [FlashAttention](#)\n- [FlashAttention-2](#)\n- [RoPE (Rotary Position Embeddings)](#)\n- [ALiBi](#)\n- [Multi-Query Attention (MQA)](#)\n- [Grouped-Query Attention (GQA)](#)\n\n### 推测解码\n- [Speculative Decoding](#)\n- [Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads](#)\n- [Fast Inference from Transformers via Speculative Decoding](#)\n- [Break the Sequential Dependency of LLM Inference Using Lookahead Decoding](#)\n- [Accelerating Large Language Model Decoding with Speculative Sampling](#)\n\n### KV Cache & 推理存储\n- [PagedAttention (vLLM)](#)\n- [Efficient Memory Management for Large Language Model Serving with PagedAttention](#)\n- [KV Cache Compression & Optimization](#)\n\n### PD 分离\n- [Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving](#)\n- [Splitwise: Efficient Generative LLM Inference Using Phase Splitting](#)\n- [DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM Inference](#)\n- [DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving](#)\n- [MemServe: Context Caching for Disaggregated LLM Serving with Elastic Memory Pool](#)\n- [TetriInfer: Inference without Interference: Disaggregate LLM Inference for Mixed Downstream Workloads](#)\n\n### LLM 应用与提示工程\n- [Retrieval-Augmented Generation (RAG)](#)\n- [METIS: Fast Quality-Aware RAG Systems with Configuration Adaptation](#)\n- [CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion](#)\n- [Parrot: Efficient Serving of LLM-based Applications with Semantic Variable](#)\n- [Towards End-to-End Optimization of LLM-based Applications with Ayo](#)\n- [Chain-of-Thought Prompting](#)\n- [Tree of Thoughts](#)\n- [ReAct](#)\n\n### MoE 混合专家\n- [Mixture of Experts (Switch Transformer)](#)\n- [DeepSeekMoE](#)\n\n### 调度与批处理\n- [DeepSpeed-FastGen: High-throughput Text Generation for LLMs](#)\n- [SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills](#)\n- [Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve](#)\n\n### 训练优化与缩放\n- [Test-Time Scaling](#)\n- [Muon Optimizer](#)\n\n## 3、Blog\n### FlashAttention\n- [ELI5: FlashAttention](#)\n- [FlashAttention from First Principles](#)\n- [Flash Attention 2.0 with Tri Dao (author)!](#)\n- [FlashAttention学习过程【详】解](#)\n- [FlashAttention — Visually and Exhaustively Explained](#)\n- [Designing Hardware-Aware Algorithms: FlashAttention](#)\n- [FlashAttention: Fast and Memory-Efficient Exact Attention With IO-Awareness](#)\n\n### Harness Engineering\n设计环境、规则、测试反馈系统，让 AI Agent 自动生成并改进代码\n- [Minions: Stripe’s one-shot, end-to-end coding agents—Part 2](#)\n- [Effective harnesses for long-running agents](#)\n- [Minions: Stripe’s one-shot, end-to-end coding agents](#)\n- [Harness engineering: leveraging Codex in an agent-first world](#)\n- [Vibe Coding AReaL：零手打代码开发分布式 RL 训练框架](#)\n\n### Triton\n- [Deep Dive into Triton Internals (Part 3)](#)\n- [Deep Dive into Triton Internals (Part 1)](#)\n- [Deep Dive into Triton Internals (Part 2)](#)\n\n### vLLM\n- [vLLM源码解析](#)\n- [Inside vLLM: Anatomy of a High-Throughput LLM Inference System](#)\n\n### GPU\n- [A history of NVidia Stream Multiprocessor](#)\n- [Building a Tiny GPU to Understand AI Hardware Engineering](#)\n\n### CUTLASS\n- [Learn CUTLASS the hard way - part 2!](#)\n- [Learn CUTLASS the hard way! (Video)](#)\n\n### 量化\n- [PyTorch 的量化实战项目](#)\n- [PyTorch 官方量化资料](#)\n\n### 推测解码\n- [How Speculative Decoding Boosts vLLM Performance by up to 2.8x](#)\n\n### CUDA\n- [LeetCUDA](https:\u002F\u002Fgithub.com\u002Fxlite-dev\u002FLeetCUDA)\n- [How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog](#)\n\n## 4、Book\n- [Build a Large Language Model (From Scratch)](#)\n- [AI Systems Performance Engineering](#)：GPU CUDA Kernel 调优、PyTorch 算法优化、多节点训练推理系统调优...\n","AI-wiki 是一个面向大语言模型（LLM）全栈技术体系的开源知识整理与学习笔记仓库，系统覆盖数据、算法、模型、硬件及系统架构五大维度。核心内容包括课程资源（LLM\u002FAgent）、技术博客（FlashAttention、量化、推测解码等）、经典书籍与前沿论文分类索引（涵盖基础模型、微调方法、Attention优化、KV Cache管理、Prefill-Decode分离、RAG等）。项目采用结构化组织方式，强调工程落地与系统级理解，适合AI工程师、系统研究员及进阶学习者开展LLM推理优化、服务部署与架构设计相关工作。",2,"2026-07-08 04:30:16","CREATED_QUERY"]