This project is a lightweight vLLM (large language model inference engine) implemented in Python. The core code is only over 1000 lines. It has a clear structure and is easy to read. The inference speed is comparable to the original vLLM and integrates inference optimization techniques such as prefix caching, tensor parallelism, and Torch compilation.