mini-SGLang — optimized minimalist inference engine

mini-SGLang — optimized minimalist inference engine The codebase is capable of full inference for Qwen 3 (Dense) and Llama 3 at the performance level of a large SGLang, which has two orders of magnitude more code. The project is intended both for training in the operation of modern inference engines and as a minimalist codebase for research. In about 5,000 lines of Python code, the main optimizations of SGLang and quite a lot of functionality fit. The engine supports both online (via OpenAI API) and offline inference, inference on multiple GPUs, and context caching. However, many had to be sacrificed — support for most models, MoE support, AMD support, etc., was removed. But I like the idea of having a minimalist version of the project with the same architecture for experimentation and onboarding new contributors; I would like to see more of this.