How Efficient Are Diffusion Language Models? A Critical Examination of Efficiency Evaluation Practices

TL;DR

This work investigates why diffusion language models, despite their parallelizable decoding, often underperform autoregressive models in efficiency. By combining systematic empirical benchmarks with a theory grounded in hardware performance and model-side work per token, the authors show that AR models typically achieve higher throughput, while DLMs are more compute-intensive per token and increasingly bottlenecked by sequence length. They analyze acceleration strategies (e.g., dual cache and parallel decoding) and demonstrate that their benefits are largest at single-batch (1) scenarios but diminish as batch size grows, highlighting a mismatch between current acceleration techniques and real-world deployment scales. The study underscores the need for standardized evaluation protocols and improved acceleration methods to advance the practical use of DLMs, particularly in settings with small batch sizes or specialized tasks where diffusion’s parallelism could be advantageous.

Abstract

Diffusion language models (DLMs) have emerged as a promising alternative to the long-dominant autoregressive (AR) paradigm, offering a parallelable decoding process that could yield greater efficiency. Yet, in practice, current open-source DLMs often underperform their AR counterparts in speed, limiting their real-world utility. This work presents a systematic study of DLM efficiency, identifying key issues in prior evaluation methods. Through empirical benchmarking and a theoretical analysis, we demonstrate that AR models generally achieve higher throughput, while DLMs consistently lag. We also investigate acceleration strategies, finding that techniques like dual cache and parallel decoding mainly offer gains at small batch sizes, with their benefits diminishing upon scaling. Our findings underscore the necessity of robust evaluation methods and improved acceleration strategies to advance research on DLMs.

Figures (6)

• Figure 1: (a) and (b) show the throughput comparison across different generation lengths under the 0-shot and 5-shot settings, respectively, while (c) presents the throughput comparison between AR and block diffusion models under longer generation lengths.
• Figure 2: The throughput comparison across different batch sizes under the 0-shot (a) and 5-shot (b) settings, respectively.
• Figure 3: (a) and (b) show the arithmetic intensity comparison across different generation lengths under the 0-shot and 5-shot settings, respectively, while (c) presents the arithmetic intensity comparison between AR and block diffusion models under longer generation lengths.
• Figure 4: The arithmetic intensity comparison across different batch sizes under the 0-shot (a) and 5-shot (b) settings, respectively.
• Figure 5: (a) and (b) show the throughput comparison across different generation lengths under the 0-shot and 5-shot settings, respectively.