博客

AI Infra 领域的推荐博客

vLLM

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

推测解码

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

GPU

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

Harness Engineering

Harness Engineering 指的是一种开发方式：工程师不直接写大量代码，而是设计环境、规则和测试反馈系统，让 AI Agent 自动生成并改进代码。

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

自动化 AI 研究

• autoresearch — 让 AI Agent 在单 GPU 上自动修改 nanochat 训练代码，以固定 5 分钟预算运行实验、评估结果并保留有效改动，形成最小化的自动研究闭环。
• First Steps Toward Automated AI Research — Recursive 对自动研究系统的早期实践总结，涵盖长周期并行探索、结果验证，以及在 NanoChat、NanoGPT Speedrun 和 GPU Kernel 优化任务上的实验。

量化

• PyTorch 官方资料
• PyTorch 的量化实战项目

CUTLASS

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

Triton

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

FlashAttention

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

CUDA

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