积跬步，至千里

静下心，去做事

Contents

• Pytorch Notes
• AI Infra
• Quantization
• Attention
• MLA
• Sparse Attention
• CUDA Notes
• Low Bit training
• Technique
• Profiler
• Docker
• Network
• Git
• pixi
• Slurm

AI Infra

• [2025.08.21] 入局AI Infra：程序员必须了解的AI系统设计与挑战知识
• [2025.08.26] 硬核：分布式推理优化思路分享
• [2025.08.26] [chat] 硬核：分布式推理优化思路分享
• [2025.08.26] [chat] Transformer模型参数训练过程
• [2025.08.26] [paper] PagedAttentioin vLLM

Quantization

• [2025.09.11] [weixin] 量化综述：5.5 outlier的来源
• [2025.09.12] [paper] LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale

Attention

MLA

• [2025.08.02] 再读MLA，还有多少细节是你不知道的
• [2025.08.04] deepseek技术解读(1)-彻底理解MLA（Multi-Head Latent Attention）
• [2025.08.04] 旋转位置编码RoPE
• [2025.08.04] 大模型推理加速：看图学KV Cache
• [2025.08.07] [code] Developing sageMLA on 5090 with per_warp quantization

Sparse Attention

• [2025.08.05] 全网首发！ DeepSeekNSA:Native Sparse Attention 最通俗讲解
• [2025.08.05] [paper] Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
• [2025.08.05] [paper] MOBA: Mixture of Block Attention for Long-context LLMs
• [2025.08.05] 笔记：MoBA 与 Native Sparse Attention
• [2025.08.06] [paper] SeerAttention: Learning Intrinsic Sparse Attention in Your LLMs
• [2025.08.07] [video] MoBA：基于MoE的sparse block attention
• [2025.08.12] [paper] H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models
• [2025.08.12] [paper] InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory

Python Notes

Notes of python and pytorch

• [2025.07.20] Python 3.12泛型+静态检查的终极防坑指南

CUDA Notes

Notes of cuda

• [2025.07.20] CUDA shared memory避免bank conflict的swizzling机制解析
• [2025.07.24] Swizzling for bank conflict的交流理解
• [2025.07.27] ldmatrix指令的详细解释
• [2025.08.11] [paper] Demystifying NCCL: An In-depth Analysis of GPU Communication Protocols and Algorithms
• [2025.08.25] [chat] 为什么Pinned memory（CUDAMallocHost, CUDAHostAlloc）在Host to Device copy中比较快

Low Bit training

• [2025.07.21] [paper] COAT: Compressing Optimizer states and Activation for Memory-Efficient FP8 Training
• [2025.07.22] [paper] Memory Efficient Optimizers with 4-bit States
• [2025.07.26] [chat] 混合精度训练流程以及master weight
• [2025.08.21] [paper] iFairy: the First 2-bit Complex LLM with All Parameters in {±1, ±i}
• [2025.08.21] [chat] iFairy paper reading
• [2025.08.21] [chat] LittleBit paper reading
• [2025.09.15] [chat-paper] Training Transformers with 4-bit Integers
• [2025.09.16] [paper] Quartet: Native FP4 Training Can Be Optimal for Large Language Models
• [2025.09.16] [chat-paper] QuEST: Stable Training of LLMs with 1-Bit Weights and Activations

Technique

Profiler

• [2025.07.21] [zhihu] 性能分析 torch.profiler

Docker

• [2025.07.21] [chat] Docker 相关命令

Network

• [2025.07.22] Macbook Wi-Fi不能上网
• [2025.07.22] Orin 配置k8s

Git

• [2025.07.22] [chat] Git check PR before merge
• [2025.08.22] [chat] Git .gitattributes 文件作用

pixi

• export PIXI_CACHE_DIR=/path/to/your/pixi-cache

• [2025.08.22] pixi 主页
• [2025.08.22] [chat] pixi 介绍

Slurm

• [2025.08.07] Slurm 任务调度