本地笔记目录
先按论文、项目、基础区分材料类型，再按主题查找。论文清单中未写笔记的条目见论文归档。历史 issue 的来源与迁移情况见 GitHub 笔记归档，个人外部技术问答见 AI Infra 面试基础归档。
论文
Attention
- DeepSeek Sparse Attention (DSA) 算法流程
- FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling：阅读材料
- IndexCache：跨层复用稀疏索引，以及它为什么能被训练
- MiniMax Sparse Attention：论文算法与代码实现精确对应
- MSA 中的 KL 对齐与局部知识蒸馏
- MSA 为什么采用 KV-outer Sparse Attention Forward
- MSA Indexer 与 DSA Lightning Indexer 的差别
- MiniMax Sparse Attention (MSA)
- Fast Transformer Decoding: One Write-Head is All You Need
- A Survey of Efficient Attention Methods: Hardware-efficient, Sparse, Compact, and Linear Attention
- Attention is all you need：历史笔记
- FLASHATTENTION: Fast and Memory-Efficient Exact Attention with IO-Awareness：历史笔记
- Kimi Linear：状态更新与混合注意力
- DeepSeek-V3.2 与 DSA：年度总结摘记
推理与 KV Cache
- Dual Chunk Attention
- HiSparse：稀疏注意力的分层 KV Cache 管理
- A Survey of LLM Inference Systems
- Microsecond-scale Preemption for Concurrent GPU-accelerated DNN Inferences：未完成摘记
分布式训练与 MoE
- MegaMoE：把 MoE 前向路径压进一个通信计算融合核
- PithTrain: A Compact and Agent-Native MoE Training System
- PockEngine: Sparse and Efficient Fine-tuning in a Pocket：历史笔记
- Seer：同步 RL 的 rollout 切分、调度与分组投机
- SonicMoE：梯度重排、异步 IO 与 tile 感知路由
GPU 编译与性能
- HIDA: A Hierarchical Dataflow Compiler for High-Level Synthesis
- MLIR: Scaling Compiler Infrastructure for Domain Specifific Computation
- The Deep Learning Compiler: A Comprehensive Survey
- TIRAMISU: A Polyhedral Compiler for Expressing Fast and Portable Code
- Categorical Foundations for CuTe Layouts
- Event Tensor：用“事件张量”编译动态 Megakernel
- CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution
- AVO: Agentic Variation Operators for Autonomous Evolutionary Search
- Cocktailer: Analyzing and Optimizing Dynamic Control Flow in Deep Learning：历史笔记
- Rammer: Enabling Holistic Deep Learning Compiler Optimizations with rTasks：历史笔记
- Roller: Fast and Efficient Tensor Compilation for Deep Learning：历史笔记
- Welder: Scheduling Deep Learning Memory Access via Tile-graph：历史笔记
- AStitch: Enabling a New Multi-dimensional Optimization Space for Memory-Intensive ML Training and Inference on Modern SIMT Architectures：历史笔记
- Graphene: An IR for Optimized Tensor Computations on GPUs：历史笔记
- BOLT: BRIDGING THE GAP BETWEEN AUTO-TUNERS AND HARDWARE-NATIVE PERFORMANCE：历史笔记
- Chimera: An Analytical Optimizing Framework for Effective Compute-intensive Operators Fusion：历史笔记
- AMOS: Enabling Automatic Mapping for Tensor Computations On Spatial Accelerators with Hardware Abstraction：历史笔记
- Mirage Persistent Kernel：任务图与持久化运行时
- Tawa：用异步引用自动生成 warp specialization
- Look Ma, No Bubbles!：Llama-1B Megakernel 阅读摘记
模型架构与推理能力
- SkVM: Revisiting Language VM for Skills across Heterogenous LLMs and Harnesses
- DeepSeek-V4: 迈向高效百万级上下文智能
- Distilling the Knowledge in a Neural Network
计算基础
- The Deminkernel Datapath OS Architecture for Microsecond-scale Datacenter Systems
- RedLeaf: Isolation and Communication in a Safe Operating System
- HyperBench: A Benchmark Suite for Virtualization Capabilities
- Security and Performance in the Delegated User-level Virtualization
- Honeycomb: Secure and Efficient GPU Executions via Static Validation
- Nephele: Extending Virtualization Environments for Cloning Unikernel-based VMs
- Unikraft: Fast, Specialized Unikernels the Easy Way：历史来源
- A Loop Transformation Theory and an Algorithm to Maximize Parallelism：未完成摘记
- Unikraft：模块化 library OS 与专用内核
项目
推理与 KV Cache
分布式训练与 MoE
GPU 编译与性能
计算基础
- Arceos irq 分析
- smoltcp
- boot_linux
- QEMU
- Virtio
- xv6-riscv 中的 KPTI 机制
- xv6-riscv 设备
- 动态内存分配
- 启动
- 环境配置
- GDB调试
- 中断
- 锁
- 进程与线程
- 调度
- 虚拟内存
- xv6 文件系统手册和代码阅读笔记
- xv6 rust design
- ArceOS：从 feature 配置到 unikernel 启动
基础
Attention
- FlashAttention 笔记
- Flash-Decoding for long-context inference
- Linear Attention
- Multi-head Latent Attention
- Sparse Attention
- Attention Sink、Learnable Sink 与 dsink：从输出含义到反向梯度
- [译] 用 60 幅图从零构建 FlashAttention-4
推理与 KV Cache
分布式训练与 MoE
GPU 编译与性能
- CUDA 内存与同步
- GPU 内存池：空间管理与异步生命周期
- GPU 执行、编译与排错
- RMSNorm、Softmax 与在线归约
- GEMM Pipeline 技术
- Tensor Core 笔记
- 代码生成设计提案
- Cute
- Memory Coalescing
- swizzle
- 向量化内存访问
- wmma
- Warp-Level Primitives
- Scatter 算子
- [译] tcgen05 入门：从零实现 B200 GEMM
模型架构与推理能力
- Transformer
- LLM Training, Inference & Algorithms — Recommended Reading Path
- 神经网络
- Adam 论文前置学习路线
- 2025 年度 LLM 研究总结：知乎原文归档
计算基础
- Instruction Level Parallelism(ILP) and Its Exploitation
- RISC V Compressed
- RISC -V N 扩展
- RISC -V H 扩展
- RISC -V Interrupt
- RISC-V IOMMU
- SBI
- TLB MMU
- 寄存器分配
- C++ 对象、生命周期与并发
- 迭代器
- Rust 智能指针
- Rust trait
- 传输层
- FAT文件系统知识概览
- 文件系统
- io_uring 阅读笔记
- 内存管理
- 模块化内核
- 多核
- 虚拟化和云
- Virtualizing I/O
- Virtualizing Memory
- QEMU/KVM
- 虚拟化硬件加速器调研
- Unikernel、Unikraft 与 ArceOS：知乎原文归档
教材与技术参考
- AI Infra 面试基础：Notion 总结归档
- 超标量处理器概览
- Cache
- Divide and Conquer
- Introduction
- CUDA
- 现代多核处理器
- 并行编程
- 并发控制
- Deep Learning System -- Introduction
- 神经网络反向传播：理论与数学推导
- Lab-2A Leader Election
- Lab-1 MapReduce
- Copy on Write Fork
- Lazy Page Allocation
- Networking
- User Level Threads
- 张量程序抽象
- [译] 深度学习所需的矩阵微积分
- Docker
- riscv-gnu-toolchain
- Verilator 学习笔记
- Visual Studio Code