A Chinese masterclass in the full stack beneath large models

What it does

AIInfra is a comprehensive, Chinese-language course on the engineering stack behind large models. Maintained by ZOMI, it spans seven modules covering AI cluster architecture, high-speed networking, storage, cloud-native deployment, distributed training, inference optimization, and algorithms. Content is delivered as slides, Markdown articles, and Jupyter notebooks, with accompanying video lectures on Bilibili and YouTube.

The interesting bit

Most educational material treats the infrastructure beneath LLMs as a black box; AIInfra opens it wide. The curriculum gives equal airtime to data-center realities—heat dissipation, ten-thousand-card cluster topologies, and high-speed interconnect—as it does to Transformer training and MFU evaluation, making it a rare public resource for the systems work that determines whether a model trains efficiently or at all.

Key highlights

• Full-stack scope: hardware (GPU/NPU), cluster networking, storage, containers, distributed training and inference, plus algorithms and data.
• Practical notebooks: Jupyter examples for building PyTorch DDP from scratch, calculating MoE parameter counts, and evaluating MFU utilization.
• Multimedia format: written content paired with slide decks and video lectures on Bilibili and YouTube.
• Active but incomplete: several modules and subsections remain unfinished, with README status markers distinguishing completed from pending topics.
• Distinct from the maintainer’s companion AISystem repository by focusing specifically on large-model distributed systems rather than general AI systems.

Caveats

• Content is entirely in Chinese, so non-readers will need translation tools.
• Several modules and subsections remain unfinished; the README shows a mix of completed and pending topics.

Verdict

Worth bookmarking if you are a Mandarin-speaking engineer moving from model-tinkering into production-scale training. If you do not read Chinese, this is currently inaccessible.