推测解码缩放法则（SDSL）：简化吞吐量优化

Speculative decoding is a technique that accelerates inference by utilizing multiple language models. Previous research has relied on experimental methods to optimize the throughput of the inference pipeline, which involves training large language models (LLMs) and can be costly. This study proposes a theory that analytically connects the key hyperparameters of pre-trained LLMs to the throughput efficiency of a downstream inference system based on speculative decoding. The theory enables the prediction of throughput-optimal hyperparameters for the components of an inference system prior to their pre-training, thereby simplifying the optimization process.