Speculative decoding accelerates large language model generation by allowing multiple tokens to be drafted swiftly by a lightweight model before being verified by a larger, more powerful one. This ...