Co-Designing AI Models Using Speculative Decoding for Faster LLM Inference