From GPU silicon to LLM inference: a Chinese AI Infra field guide

What it does

AIInfraGuide is a curated, static-site curriculum covering the full AI infrastructure stack for Chinese-speaking engineers. It walks from GPU hardware architecture and CUDA kernel optimization through distributed training paradigms (DDP, FSDP, 3D parallelism) to LLM inference engines and quantization. The project also bundles an interview-prep module with 180+ real questions collected from 40+ companies.

The interesting bit

The field moves faster than textbooks can be printed, so the guide treats itself as a living document—content is explicitly marked as continuously updating. Its pedagogical hook is a “plain language first, jargon second” rule, which is a polite admission that most AI Infra documentation assumes you already work at NVIDIA.

Key highlights

• Full-stack scope: GPU hardware → CUDA programming → collective communication → distributed training → inference optimization (vLLM, PagedAttention, speculative decoding)
• Interview module with 180+ questions from 40+ companies, organized by company tier
• Pedagogical rule: explain concepts in plain language before introducing rigorous terminology
• Built with Astro and served via GitHub Pages
• MIT licensed and open to contributions

Caveats

• Explicitly a work in progress (“内容持续更新中”); not all planned modules are fully published yet
• Entirely Chinese-language; no English translation is currently available
• This is a content repository and static site, not a framework or executable tool

Verdict

Worth bookmarking if you are a Mandarin-speaking engineer trying to break into AI Infra or patch holes in your knowledge systematically. Skip it if you are looking for a drop-in library or prefer your learning materials in English.