AdapTive-LeArning Speculator System (ATLAS): Faster LLM inference