We just released a tiny (~3kloc) Python library that implements state-of-the-art inference algorithms on GPU and provides performance similar to vLLM. We believe it's a great learning vehicle for inference techniques and the code is quite easy to hack on!
