Speculative Streaming: Fast LLM Inference Without Auxiliary Models