Full-Stack AI Resource Directory

• Course
  • LLM
  • Agent
• Blog
  • FlashAttention
  • Harness Engineering
  • Quantization
  • Speculative Decoding Blog
  • CUDA
• Book
• Paper
  • Base Model
  • Fine-tuning
  • Attention Optimization
  • Speculative Decoding
  • KV Cache & Inference Storage
  • Prefill-Decode Disaggregation
  • LLM Application & Prompt Engineering
  • MoE Mixture of Experts
  • Scheduling & Batching
  • Training Optimization & Scaling

一、Course

LLM

• CS224n：系统讲解词向量、Transformer、大模型训练微调、LLM 应用落地全流程
• CS336: Language Modeling from Scratch (Stanford / Spring 2026)
  • 从 tokenizer、Transformer 到分布式训练、数据处理和 RLHF 对齐，5 个assignment 覆盖 LLM 全链路
• CSCI 1390, Spring 2025: Systems for Machine Learning：高效训练和推理、理解如何构建硬件 ML 算法、理解 ML 算法性能、GPU 编程和 CUDA、Transformer 架构、高效检索…

Agent

• 从零开始理解 Agent：基于极简开源项目 nanoAgent，拆解 OpenClaw / Claude Code 等 AI Agent 核心概念
• Learn Claude Code：从零到一构建类 Claude Code AI Agent，12 节渐进课程，含 Agent Loop、工具调用、子代理、上下文压缩、任务系统、多代理协作、Worktree 隔离等机制

2、Paper

底座

PaLM，OPT，BLOOM，LLaMA

微调

• 对齐微调: InstructGPT (RLHF)，Constitutional AI，Self-Instruct，Direct Preference Optimization (DPO)，ORPO，GRPO

• 轻量化微调: LoRA，QLoRA

Attention 优化

• FlashAttention
• FlashAttention-2
• RoPE (Rotary Position Embeddings)
• ALiBi
• Multi-Query Attention (MQA)
• Grouped-Query Attention (GQA)

推测解码

• Speculative Decoding
• Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
• Fast Inference from Transformers via Speculative Decoding
• Break the Sequential Dependency of LLM Inference Using Lookahead Decoding
• Accelerating Large Language Model Decoding with Speculative Sampling

KV Cache & 推理存储

• PagedAttention (vLLM)
• Efficient Memory Management for Large Language Model Serving with PagedAttention
• KV Cache Compression & Optimization

PD 分离

• Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving
• Splitwise: Efficient Generative LLM Inference Using Phase Splitting
• DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM Inference
• DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving
• MemServe: Context Caching for Disaggregated LLM Serving with Elastic Memory Pool
• TetriInfer: Inference without Interference: Disaggregate LLM Inference for Mixed Downstream Workloads

LLM 应用与提示工程

• Retrieval-Augmented Generation (RAG)
• METIS: Fast Quality-Aware RAG Systems with Configuration Adaptation
• CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion
• Parrot: Efficient Serving of LLM-based Applications with Semantic Variable
• Towards End-to-End Optimization of LLM-based Applications with Ayo
• Chain-of-Thought Prompting
• Tree of Thoughts
• ReAct

MoE 混合专家

• Mixture of Experts (Switch Transformer)
• DeepSeekMoE

调度与批处理

• DeepSpeed-FastGen: High-throughput Text Generation for LLMs
• SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills
• Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve

训练优化与缩放

• Test-Time Scaling
• Muon Optimizer

3、Blog

FlashAttention

• ELI5: FlashAttention
• FlashAttention from First Principles
• Flash Attention 2.0 with Tri Dao (author)!
• FlashAttention学习过程【详】解
• FlashAttention — Visually and Exhaustively Explained
• Designing Hardware-Aware Algorithms: FlashAttention
• FlashAttention: Fast and Memory-Efficient Exact Attention With IO-Awareness

Harness Engineering

设计环境、规则、测试反馈系统，让 AI Agent 自动生成并改进代码

• Minions: Stripe’s one-shot, end-to-end coding agents—Part 2
• Effective harnesses for long-running agents
• Minions: Stripe’s one-shot, end-to-end coding agents
• Harness engineering: leveraging Codex in an agent-first world
• Vibe Coding AReaL：零手打代码开发分布式 RL 训练框架

Triton

• Deep Dive into Triton Internals (Part 3)
• Deep Dive into Triton Internals (Part 1)
• Deep Dive into Triton Internals (Part 2)

vLLM

• vLLM源码解析
• Inside vLLM: Anatomy of a High-Throughput LLM Inference System

GPU

• A history of NVidia Stream Multiprocessor
• Building a Tiny GPU to Understand AI Hardware Engineering

CUTLASS

• Learn CUTLASS the hard way – part 2!
• Learn CUTLASS the hard way! (Video)

量化

• PyTorch 的量化实战项目
• PyTorch 官方量化资料

推测解码

• How Speculative Decoding Boosts vLLM Performance by up to 2.8x

CUDA

• LeetCUDA
• How to Optimize a CUDA Matmul Kernel for cuBLAS-like Performance: a Worklog

4、Book

• Build a Large Language Model (From Scratch)
• 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…