Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding

TL;DR

Fast-dLLM introduces a training-free method to accelerate diffusion-based large language models by enabling KV cache reuse and confidence-aware parallel decoding, significantly improving throughput while maintaining quality.

Contribution

It proposes a novel block-wise approximate KV Cache and confidence-aware decoding to enhance inference speed of diffusion LLMs without retraining.

Findings

Achieves up to 27.6× throughput improvement.

Maintains high generation quality with minimal accuracy loss.

Bridges the performance gap between diffusion and autoregressive models.

Abstract

Diffusion-based large language models (Diffusion LLMs) have shown promise for non-autoregressive text generation with parallel decoding capabilities. However, the practical inference speed of open-sourced Diffusion LLMs often lags behind autoregressive models due to the lack of Key-Value (KV) Cache and quality degradation when decoding multiple tokens simultaneously. To bridge this gap, we introduce a novel block-wise approximate KV Cache mechanism tailored for bidirectional diffusion models, enabling cache reuse with negligible performance drop. Additionally, we identify the root cause of generation quality degradation in parallel decoding as the disruption of token dependencies under the conditional independence assumption. To address this, we propose a confidence-aware parallel decoding strategy that selectively decodes tokens exceeding a confidence threshold, mitigating dependency violations and maintaining generation quality. Experimental results on LLaDA and Dream models across multiple LLM benchmarks demonstrate up to \textbf{27.6$\times$ throughput} improvement with minimal accuracy loss, closing the performance gap with autoregressive models and paving the way for practical deployment of Diffusion LLMs.