Medusa: Simple Framework for Accelerating LLM Generation with Multiple Decoding Heads - View it on GitHub