I built vllm from the ground up. Starting from the same kernels, I integrate them into GPT2, build the KV cache manager, the scheduler, and finally the FastAPI server on top. This is purely educational for now, but in the future it will become production ready.