# Inside caomaolufei/AIInfraGuide: an AI infrastructure curriculum built with Astro and Pagefind > AIInfraGuide is a Chinese-language Astro site that organises AI infrastructure study into four modules, a performance profiling section and a bank of interview questions, and it ships as a documentation project with no licence file and no releases. **caomaolufei/AIInfraGuide** — AI Infra 全栈从0入门学习资料：https://caomaolufei.github.io/AIInfraGuide/ - Repository: https://github.com/caomaolufei/AIInfraGuide - Stars: 2,534 · Forks: 211 - Language: Astro - License: not declared - Published: 2026-09-30 · Updated: 2026-09-30 · Language: en - Canonical page: https://hysenlabs.com/projects/caomaolufei-aiinfraguide ## A syllabus for infrastructure work, ordered by dependency The organising idea of AIInfraGuide is that AI infrastructure is learned in one direction only. The sequence starts with prerequisites and ends with inference, and nothing later in the list is reachable without what came before it. Module one holds the foundations: programming language fundamentals, mathematics, the Transformer architecture, the PyTorch framework, GPU hardware, and collective communication. Module two is CUDA programming and operator optimisation, split into eight chapters that move from a development environment through the programming and memory models into the classic operators, Reduce, GEMM, Softmax and Attention, then into AI compilers and the profiling toolchain. Module three is distributed training, and module four is inference optimisation, where the listed topics run from the basics of large language model serving through engine internals, mainstream frameworks such as vLLM, quantisation, speculative decoding, prefill and decode disaggregation, and benchmarking. A separate performance analysis section collects Nsight Systems, Nsight Compute and the Roofline model. Every article is described as following the same order, an explanation in plain language of what something is and why it exists, followed by the technical detail. For an engineer who already works in this field, that ordering is the interesting part, because it exposes which dependencies the author considered load bearing. ## The interview bank is the part with a countable claim Two supporting boards sit alongside the four modules, and one of them makes a specific numerical claim that is easy to check and easy to compare. The interview bank is described as holding more than 180 real interview questions drawn from more than 60 companies, sorted into tiers. Documentation projects rarely state their size in those terms, so this is the one number in the repository that a reader can hold the project to. The other supporting board is the learning route, a single article that presents the overall shape of the stack together with a knowledge map and recommended resources. The tables also distinguish between planned chapters and published ones. A row such as the first chapter of module two appears as a bare heading with its contents described in text, while the chapter beneath it has a working link. That distinction is the honest signal of where the project stands: outlines exist for the full arc, links exist only for what has been written. The entries that do link out are unusually specific for a survey. Module two names FlashAttention versions one, two and three, online Softmax, operator fusion, warp shuffle reductions, shared memory tiling against cuBLAS, PagedAttention written in CUDA, Triton, torch.compile, and TVM and XLA. Module one includes a chapter on collective communication running from primitives through the ring all-reduce algorithm to NCCL. Naming a specific kernel version or a specific comparison target is a commitment the author would not make in a generic index, and it tells you the intended reader is someone who will run the code rather than someone browsing for orientation. ## Astro and Pagefind are load bearing here The project is a documentation site, and the toolchain is chosen for a documentation site rather than for content. The manifest pins Astro 4, adds Tailwind through the official integration, and brings in the typography plugin along with two self-hosted font packages, Inter for text and JetBrains Mono for code. Search is handled by Pagefind, which indexes the built output rather than the source, and the search interface comes from the matching default UI package. The build script is where the interesting decision sits: ``` "build": "astro check && astro build && pagefind --site dist && cp -r dist/pagefind public/pagefind" ``` Astro's type check runs first, then the static build, then Pagefind indexes the generated `dist` directory, and finally the generated index is copied back into `public`. That last copy is the part worth understanding. Placing the index inside the public source directory is what makes search available while running the dev server locally, where the output has never been built, not only in the deployed site. It also means the index exists as a committed artefact in the source tree, which is a trade a documentation project makes deliberately to keep local preview and production behaving the same way. The remaining dependencies are a unified set of remark and rehype plugins that map onto what a technical curriculum needs: math support through remark-math and rehype-katex for the equations a Transformer chapter cannot avoid, heading anchors and external link handling for cross-references, and slug generation for stable fragment identifiers. ## The MIT badge and the missing licence file The repository carries a licence badge reading MIT in the README, and the licence field in the project metadata is unresolved. Those two facts sit next to each other without agreement, and the file listing settles the question: at the top level there is a `.github` directory, a `.gitignore`, the README, the Astro and Tailwind configuration files, `docs`, `public`, `scripts`, `src`, and the two package manifests. There is no licence file among them. A badge is a claim rendered as an image. A licence file is the grant itself. For a project whose entire value is its text, that difference is larger than it would be for a tool, because a tool's licence matters at install time while a document's licence matters at redistribution time. Someone who wants to fork the CUDA chapters, adapt the interview questions for their own team's preparation, or host a mirror is left guessing at the terms, and the answer is not a matter of interpretation because there is nothing to interpret. This does not make the project unusable, and it is not a reason to distrust the content. It is a reason to read before copying. The same caution applies to the deployment: the site is published under the author's account as a GitHub Pages URL, so the hosting is tied to that account rather than to an organisation that could outlive it. ## A documentation repository with no releases and a version that never moves The project has no GitHub releases, and the version field in the manifest reads 1.0.0 with the package left unscoped. For a documentation site that is unremarkable. For anything else it would be a red flag, and the contrast is instructive because the repository looks enough like a software project, with a lockfile, a TypeScript configuration, a lint-free but typed build and a scripts directory, to invite the assumption that it is installable. It is not. The package name is unscoped, the manifest declares only module type and version at the top, and there is no publish configuration. The build produces a static directory that is served, not a library that is resolved. Anyone reaching for this project as a dependency will find nothing to install, and the absence of releases is the honest signal rather than an oversight. The maintenance picture is straightforward to state from dates. The last push landed on 2026-09-17, the repository is not archived, and the README describes the content as continuously being updated. The pull requests welcome badge suggests contributions are welcome, and the presence of a `.github` directory indicates where that workflow lives, though the repository does not describe a contribution guide or a review process in what is published. For a corpus that grows one chapter at a time, that is a reasonable amount of process, and a reader who cares about accuracy over breadth will find the open pull request path more useful than any version number. ## What a reader should do with the outline that is not written yet The most useful thing in the repository is the shape of the incomplete part, so it is worth separating what exists from what is promised. The tables list the full arc of an infrastructure curriculum, and the project is candid that this is a work in progress. Several rows are headings with a description but no link, and one of the two supporting boards is described as a planned collection rather than a finished one. For a reader, that has a practical consequence. The order of the modules is trustworthy as a dependency graph, since a curriculum that places collective communication after GPU hardware and before distributed training is making a claim about prerequisites rather than about taste. The coverage of any individual chapter is not, because a described chapter and a written chapter are different objects and the tables mark them differently. A sensible use is to treat the tables as a syllabus to be checked against a local plan, then spend time only in the chapters that have links and that cover ground the reader cannot fill from elsewhere. The strongest entries by their own descriptions are the ones anchored to a concrete implementation target: a vector addition kernel walked end to end, a shared memory tiled GEMM compared against cuBLAS, a CUDA implementation of PagedAttention, a ring all-reduce derivation leading into NCCL, and a Transformer forward pass written in PyTorch. Those are the chapters where a survey either teaches something a reference manual will not, or it does not. ## Conclusion Read AIInfraGuide if you want a Chinese-language map of the AI infrastructure stack and can accept that its value is a syllabus rather than a maintained library. Do not adopt it as a dependency, and do not redistribute its text, because the repository carries an MIT badge in the README while shipping no licence file, which leaves the grant ambiguous. Check the chapter list against what your team actually owns before committing study time, since the published tables are described as still being filled in and some entries are chapter headings with no linked article yet. ## FAQ ### What does caomaolufei/AIInfraGuide cover? It covers AI infrastructure in four ordered modules: prerequisites, CUDA programming and operator optimisation, distributed training, and inference optimisation, plus a performance analysis section and an interview bank. Prerequisites include programming, mathematics, the Transformer, PyTorch, GPU hardware and collective communication. ### Is AIInfraGuide written in English or Chinese? The articles are written in Chinese and hosted at a GitHub Pages URL under the author's account. Each article is described as explaining a topic in plain language first and technical detail second. ### Does AIInfraGuide ship with a licence file? No. The README displays an MIT badge, but no licence file appears among the top level repository entries, so the terms attached to the text are not stated in the repository itself. ### How does AIInfraGuide implement search on the site? Search runs on Pagefind, which indexes the built output directory after the Astro build. The build script then copies the generated index back into the public directory so search also works when previewing locally. ### Is AIInfraGuide usable as an npm package? No. It is a static documentation site built with Astro and served from a GitHub Pages URL. The package version reads 1.0.0, the repository has no GitHub releases, and nothing in the manifest is configured for publishing. ## Sources - [caomaolufei/AIInfraGuide on GitHub](https://github.com/caomaolufei/AIInfraGuide) - [Issues](https://github.com/caomaolufei/AIInfraGuide/issues) - [README](https://github.com/caomaolufei/AIInfraGuide/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/caomaolufei-aiinfraguide