caomaolufei/AIInfraGuide

What it solves

AIInfraGuide addresses the lack of systematic, up-to-date Chinese-language resources for learning AI infrastructure. It helps engineers navigate the rapidly evolving landscape of large model training and inference, bridging the gap between hardware knowledge and software implementation.

How it works

The project provides a structured knowledge base organized into four core learning modules: pre-requisites (like GPU architecture and Transformer basics), CUDA programming and operator optimization, distributed training strategies, and inference optimization (centered on vLLM). It follows a "plain language first, technical details second" approach to ensure concepts are accessible yet rigorous. Additionally, it includes a comprehensive interview preparation guide with over 180 real-world questions from 60+ companies.

Who it’s for

This guide is designed for software engineers, students, and researchers who want to build a professional-grade understanding of the AI technology stack, ranging from hardware-level optimization to large-scale distributed training and model deployment.

Highlights

• Systematic learning paths covering the full AI infrastructure stack.
• Deep dives into CUDA programming, including classic operator implementations like GEMM and FlashAttention.
• Comprehensive coverage of distributed training techniques (DDP, FSDP, 3D parallelism).
• Extensive inference optimization resources, featuring vLLM source code analysis and quantization techniques.
• A curated interview database categorized by company tier to assist with career development.