Speculative Decoding in LLM Inference