Dynamic Speculation Lookahead Accelerates Speculative Decoding of Large Language Models

Speculative decoding is commonly used for reducing the inference latency of large language models. Its effectiveness depends highly on the speculation lookahead (SL)--the number of tokens generated by the draft model at each iteration. In this work we show that the common practice of using the same SL for all iterations (static SL) is suboptimal. We introduce DISCO (DynamIc SpeCulation lookahead Optimization), a novel method for dynamically selecting the SL. Our experiments Figure 1: An illustration of a single speculative decoding with four datasets show that DISCO reaches an iteration with Speculation Lookahead (SL) = 5. average speedup of 10% compared to the best Given a prompt t