3 ms·Speculative cascades – A hybrid approach for smarter, faster LLM inference6 points by emschwartz 1y ago