XNews

SGLang open-source course from Datawhale and RadixArk

SGLang open-source course from Datawhale and RadixArk

Datawhale and RadixArk (SGLang) launched an open-source course that teaches LLM inference concepts, builds a mini-sglang from scratch, and then explores real SGLang source code. The course includes a chapter on coding ethics and aims to turn learners into SGLang contributors.

Key insights

Course builds mini-sglang from scratch to teach inference engine design

The course has four parts: Part 0 covers coding ethics and deploying SGLang; Part I teaches LLM inference concepts (prefill, decode, KV cache, roofline model) with zero GPU requirement; Part II builds a mini-sglang in ten chapters covering correctness, efficiency, serving, scheduling, memory management, and scaling; Part III explores real SGLang source code for attention backends, quantization, and PD separation; Part IV covers profiling and PR workflow.

Written by SGLang core team members with strong engineering practices

Advanced parts are written by SGLang official members (YichiZ03, fcranzhou). The course includes a unified writing template, CI scripts for structure checking, and bilingual (Chinese/English) support. The coding ethics chapter explicitly opposes AI slop contributions and emphasizes profiling before optimization.

Source posts