Speculative decoding accelerates autoregressive generation by letting a faster draft model propose multiple tokens that a target model verifies in parallel without changing the target distribution. This guide explains the mechanism, trade-offs, evaluation, and controls that matter in practice.