“The pre-fill and decode stages of LLM inference have distinct computational profiles, creating an opportunity for specialized hardware optimized for each stage.”