dKV-Cache: The Cache for Diffusion Language Models

TL;DR

This paper introduces dKV-Cache, a novel caching mechanism that significantly accelerates diffusion language models during inference, narrowing the gap with autoregressive models without retraining.

Contribution

The paper proposes a delayed and conditioned KV-cache-like mechanism for DLMs, enabling faster inference and improved performance, a novel approach for non-autoregressive models.

Findings

Achieves 2-10x inference speedup across benchmarks.

Provides almost lossless acceleration with improved long-sequence performance.

Enables training-free cache utilization in existing DLMs.

Abstract

Diffusion Language Models (DLMs) have been seen as a promising competitor for autoregressive language models. However, diffusion language models have long been constrained by slow inference. A core challenge is that their non-autoregressive architecture and bidirectional attention preclude the key-value cache that accelerates decoding. We address this bottleneck by proposing a KV-cache-like mechanism, delayed KV-Cache, for the denoising process of DLMs. Our approach is motivated by the observation that different tokens have distinct representation dynamics throughout the diffusion process. Accordingly, we propose a delayed and conditioned caching strategy for key and value states. We design two complementary variants to cache key and value step-by-step: (1) dKV-Cache-Decode, which provides almost lossless acceleration, and even improves performance on long sequences, suggesting that existing DLMs may under-utilise contextual information during inference. (2) dKV-Cache-Greedy, which has aggressive caching with reduced lifespan, achieving higher speed-ups with quadratic time complexity at the cost of some performance degradation. dKV-Cache, in final, achieves from 2-10x speedup in inference, largely narrowing the gap between ARs and DLMs. We evaluate our dKV-Cache on several benchmarks, delivering acceleration across general language understanding, mathematical, and code-generation benchmarks. Experiments demonstrate that cache can also be used in DLMs, even in a training-free manner from current DLMs.