openmlsys: a machine learning systems textbook you can read online or build with mdbook
Machine Learning Systems: Design and Implementation (V2 is launching soon).
At a glance
- What is it?
- The OpenMLSys book covers nine layers from tensor abstraction to GPU cluster scheduling, in Chinese and English from one Rust powered mdbook build, and its install steps send you to a differently named repository to get it.
- Who is it for?
- Read this book if you are the person who builds the training and serving stack rather than the one who tunes a model, because the nine chapters are organised by system layer and skip the model theory you already have. Do not lift chapters into paid training without checking the licence, which is CC BY-NC-SA 4.0 and names non-commercial use in the title.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 20 days ago.
- What is it written in?
- Mainly TeX, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- openmlsys/openmlsys4,878 ★ · +23
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| openmlsys/openmlsys | 4.9k | 4.9k |
23 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Nine chapters, one per layer of the stack
The book is organised by system layer, from the programming interface down to the cluster that runs it. The second edition is laid out as nine chapters:
| 章节 | 内容 | |---|---| | 第1章 导论 | 机器学习系统架构和技术栈概述 | | 第2章 编程接口与计算图 | 张量抽象、自动微分、图表示与执行 | | 第3章 AI加速器与编程 | GPU 架构与 CUDA/Triton/CUTLASS 编程模型 | | 第4章 AI编译器与运行时系统 | IR 设计、图优化、算子生成与运行时执行 | | 第5章 数据处理系统 | 数据加载、数据管道和分布式数据处理 | | 第6章 训练系统 | 单节点与分布式训练、并行策略与训练优化 | | 第7章 模型服务 | 推理优化、在线服务与模型管理 | | 第8章 强化学习系统 | 强化学习管道、环境交互与 RL 系统设计 | | 第9章 大规模GPU集群管理 | GPU 调度、资源管理与大规模训练基础设施 |
Three audiences are named explicitly: students who have the theory and want the systems, researchers who need custom operators or distributed execution for large models, and engineers who own machine learning infrastructure and do performance tuning. Chapters 3 and 4 are the part almost nobody writes, which is where this book's value sits.
Each chapter link points into v2/zh_chapters/, so the source is one directory per chapter rather than a monolithic document.
The install steps clone openmlsys-zh, not the repository you are reading
The build instructions and the repository name do not match. The install section says:
git clone https://github.com/openmlsys/openmlsys-zh.git
cd openmlsys-zhThat is a different repository name from the one hosting this README. The badges at the top make the same trip: the actions workflow link, the site link, the licence link and the stargazers link all point into openmlsys-zh.
So the repository hands you a differently named one for building. That is defensible, since one book in two languages has to be built from one source of truth, but it means the clone command is not the command you would guess from the URL you arrived on. Check which repository holds the branch you want before you start editing files.
The root does carry both build scripts, build_mdbook_v1.sh and build_mdbook_v2.sh, so the two lines above are about where the maintained build lives rather than about what is possible here.
The v2 book needs Rust and mdbook, not the Sphinx chain
Building the second edition means installing a Rust toolchain and mdbook:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | shcargo install mdbookThen one script produces both languages:
sh build_mdbook_v2.shThe output paths are the interesting part. The English result lands in .mdbook-v2/book and the Chinese result in .mdbook-v2-zh/book, so a single build command gives you two books rather than a bilingual merge in one file. The stated prerequisites for all of this are curl, git and Python 3, which undersells what the Rust step actually pulls in.
The older build is still in the tree as build_mdbook_v1.sh, and the two versions coexist because v1/ and v2/ are separate directories. If you are following the book rather than editing it, the v2 script is the one that matches the nine chapter table.
requirements.txt exists to keep an old toolchain from breaking
The dependency file is three lines long and two of them exist for compatibility reasons rather than function:
bs4
setuptools<81
sphinxcontrib-bibtex>=2.5.0The comments say why. setuptools is pinned below 81 so that pkg_resources, which notedown uses, is still available. sphinxcontrib-bibtex is pinned at 2.5.0 or higher because d2lbook asks for a version below 2.0.0, and that older version depends on collections.MutableSet, which Python removed in 3.10. The override is installed after d2lbook so it wins the conflict.
That tells you this file belongs to the Sphinx era of the project, the one that needs d2lbook and notedown, rather than to the mdbook build described above. Nothing in the requirements file mentions Rust or mdbook, because that toolchain does not go through pip.
If you are on Python 3.10 or newer and only want to build v2, this file is not your starting point.
CC BY-NC-SA 4.0 puts non-commercial use in the licence name
The project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International licence. The NonCommercial element is the part to read twice if you intend to reuse the text: it forbids commercial use of the book, and the ShareAlike element requires derivatives to keep the same licence.
If you want to put a chapter into a paid course, you need permission that this licence does not grant. Nothing in the README discusses an exception, and the chapter files are the unit people actually reuse.
The README also asks readers to cite the work, giving a plain text form crediting OpenMLSys Team with the title and the year 2022 plus the site URL, and a BibTeX entry carrying the same fields with a note that it is an open source textbook.
Both forms are dated 2022, which is when the project started, so a citation of the current second edition still points at the first.
The newest milestone is a bilingual rebuild from March 2026
The update log has four rows and the last one matters most:
| 日期 | 事件 | |---|---| | 2022-01 | 项目初始化,开始中文内容编写 | | 2022-05 | 完成扩展篇各章节(联邦学习、强化学习、可解释 AI) | | 2023-05 | 适配 MindSpore 2.0 | | 2026-03 | 中英文双语构建架构重构;启动英文版 |
So the Chinese book was started in January 2022, extension chapters on federated learning, reinforcement learning and explainable AI landed in May 2022, the text was adapted to MindSpore 2.0 in May 2023, and March 2026 brought a rebuild of the bilingual build architecture plus the start of the English edition.
The last push to the repository was on 2026-10-26. Read that as a maintenance signal on its own terms: the build moved to a bilingual architecture and the English edition began, with nothing after it in the log.
Meanwhile the repository description still says V2 is launching soon, and the homepage recorded for this repository points at the v1 Chinese site.
Contributing means a style guide and a bilingual glossary
The contribution path is documented in three files under CONTRIBUTING/, and all three are named before a pull request is expected. CONTRIBUTING/CONTRIBUTING_zh.md holds the process, CONTRIBUTING/style_zh.md the writing style rules, and CONTRIBUTING/terminology.md a Chinese and English terminology mapping.
That glossary is the part that reveals how the bilingual edition works. A single term list is what keeps 算子 and operator from drifting apart across the two versions of a chapter, and it is the file to read before translating anything.
The community channel is a WeChat group reached by scanning a QR code in the README, which makes sense for the audience this book names and rules out most asynchronous international discussion.
The remaining root entries support the build and the checks: info/, tests/, tools/, requirements.txt, the two build scripts, and the v1/ and v2/ content directories.
Editorial conclusion
Read this book if you are the person who builds the training and serving stack rather than the one who tunes a model, because the nine chapters are organised by system layer and skip the model theory you already have. Do not lift chapters into paid training without checking the licence, which is CC BY-NC-SA 4.0 and names non-commercial use in the title. Two things to sort out before you start: that you clone openmlsys-zh as the instructions say rather than the repository you found the link to, and that you have the curl, git and Python 3 prerequisites plus the Rust toolchain and mdbook the v2 build needs.
Frequently asked questions
What is the openmlsys book?
An open source textbook, 机器学习系统:设计和实现, on the design and implementation of modern machine learning systems. The second edition runs to nine chapters, from programming interfaces and computation graphs through AI accelerators, compilers, data processing, training, serving, reinforcement learning and large-scale GPU cluster management.
How do I build the openmlsys book locally?
The stated prerequisites are curl, git and Python 3, then install the Rust toolchain and mdbook with cargo install mdbook, and run sh build_mdbook_v2.sh. The English output goes to .mdbook-v2/book and the Chinese output to .mdbook-v2-zh/book, and the clone command in the README points at the openmlsys-zh repository.
What licence is the openmlsys book under?
The Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International licence, CC BY-NC-SA 4.0. The non-commercial term means chapters cannot be reused in paid courses without separate permission, and the share-alike term applies to derivatives.
How do I contribute to the openmlsys book?
Read CONTRIBUTING/CONTRIBUTING_zh.md for the process, the writing style rules in CONTRIBUTING/style_zh.md, and CONTRIBUTING/terminology.md for the Chinese and English terminology mapping. The README links a WeChat group as the community channel.
Is the second edition of the openmlsys book finished?
Not according to the repository. The description says V2 is launching soon, the update log records March 2026 as the bilingual build refactor and the start of the English edition, and the homepage points at the v1 Chinese site.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/openmlsys-openmlsys)Star history chart for your README
[](https://hysenlabs.com/projects/openmlsys-openmlsys)Was this useful?