Speculative Speculative Decoding

TL;DR

Speculative speculative decoding (SSD) is a novel method that parallelizes token prediction and verification in autoregressive models, significantly speeding up inference by pre-emptively generating and verifying tokens.

Contribution

We introduce SSD, a new parallel decoding algorithm that overcomes sequential dependence in speculative decoding, leading to faster inference with minimal overhead.

Findings

SSD achieves on average 30% faster inference than existing speculative decoding methods.

SSD can be up to 5 times faster than traditional autoregressive decoding.

Our implementation demonstrates practical speed improvements on open source inference engines.

Abstract

Autoregressive decoding is bottlenecked by its sequential nature. Speculative decoding has become a standard way to accelerate inference by using a fast draft model to predict upcoming tokens from a slower target model, and then verifying them in parallel with a single target model forward pass. However, speculative decoding itself relies on a sequential dependence between speculation and verification. We introduce speculative speculative decoding (SSD) to parallelize these operations. While a verification is ongoing, the draft model predicts likely verification outcomes and prepares speculations pre-emptively for them. If the actual verification outcome is then in the predicted set, a speculation can be returned immediately, eliminating drafting overhead entirely. We identify three key challenges presented by speculative speculative decoding, and suggest principled methods to solve each. The result is Saguaro, an optimized SSD algorithm. Our implementation is on average 30% faster than optimized speculative decoding baselines and up to 5x faster than autoregressive decoding with open source inference engines.