Comment by meatmanek

This is super cool. Do you know if any of the inference backends (llama.cpp, vllm, etc) support this technique?