QuesmaOrg/awesome-ai-tokenomics
A curated reference on AI tokenomics: the pricing, measurement, optimization and governance of the tokens production LLM systems consume, with one-line summaries and the number behind each entry.
Awesome Infra for AI › Educational Resources
datawhalechina/zero-to-sglang
zero-to-sglang is an open educational course, jointly produced by Datawhale and RadixArk (the company behind the SGLang project), that teaches LLM inference engineering by building a working inference server from first principles. Rather than staying at a conceptual level or dropping readers directly into the SGLang source with no onramp, the course is structured in four parts. Part 0 sets up the environment and deploys a first SGLang server. Part I covers inference foundations without requiring a GPU: what an LLM is, how prefill differs from decode, compute-bound versus memory-bound execution, the KV cache as the core data structure of inference, and how to design and read a serving benchmark using metrics like TTFT, TPOT and goodput. Part II has readers build a mini-sglang incrementally, adding a forward pass and generation loop, HTTP serving, continuous batching, paged KV cache and RadixAttention prefix caching one piece at a time, each landing in real code. Part III returns to the production SGLang codebase to explain how its advanced techniques are implemented: attention backends, CUDA graphs, quantization, hierarchical caching, disaggregated prefill-decode serving, and other frontier optimizations. A planned final part covers profiling, trace analysis and how to land a first pull request in SGLang. The course is available in English and Chinese and targets engineers who already know Python, PyTorch and basic linear algebra and want to understand and contribute to modern LLM serving infrastructure rather than just use it.
https://github.com/datawhalechina/zero-to-sglang