Diffusion Models are Secretly Exchangeable: Parallelizing DDPMs via Autospeculation

TL;DR

This paper reveals that diffusion models can be parallelized by exploiting an exchangeability property, leading to a new decoding method that significantly speeds up inference without extra models.

Contribution

It proves an exchangeability property for DDPMs and introduces Autospeculative Decoding, enabling near-linear parallel speedup during inference.

Findings

Autospeculative Decoding achieves old speedup over sequential DDPMs.

Theoretical analysis shows old runtime reduction with ASD.

Practical implementation accelerates inference across various domains.

Abstract

Denoising Diffusion Probabilistic Models (DDPMs) have emerged as powerful tools for generative modeling. However, their sequential computation requirements lead to significant inference-time bottlenecks. In this work, we utilize the connection between DDPMs and Stochastic Localization to prove that, under an appropriate reparametrization, the increments of DDPM satisfy an exchangeability property. This general insight enables near-black-box adaptation of various performance optimization techniques from autoregressive models to the diffusion setting. To demonstrate this, we introduce \emph{Autospeculative Decoding} (ASD), an extension of the widely used speculative decoding algorithm to DDPMs that does not require any auxiliary draft models. Our theoretical analysis shows that ASD achieves a $\tilde{O} (K^{\frac{1}{3}})$ parallel runtime speedup over the $K$ step sequential DDPM. We also demonstrate that a practical implementation of autospeculative decoding accelerates DDPM inference significantly in various domains.