Sink-Aware Pruning for Diffusion Language Models
This work addresses the efficiency problem for users of Diffusion Language Models, though it is incremental as it adapts pruning techniques from autoregressive models to DLMs.
The paper tackles the high inference cost of Diffusion Language Models (DLMs) by showing that attention sink tokens, which are preserved in pruning methods for autoregressive models, are unstable in DLMs. The proposed Sink-Aware Pruning method automatically identifies and prunes these unstable sinks, achieving a better quality-efficiency trade-off and outperforming prior pruning baselines under matched compute without retraining.
Diffusion Language Models (DLMs) incur high inference cost due to iterative denoising, motivating efficient pruning. Existing pruning heuristics largely inherited from autoregressive (AR) LLMs, typically preserve attention sink tokens because AR sinks serve as stable global anchors. We show that this assumption does not hold for DLMs: the attention-sink position exhibits substantially higher variance over the full generation trajectory (measured by how the dominant sink locations shift across timesteps), indicating that sinks are often transient and less structurally essential than in AR models. Based on this observation, we propose ${\bf \texttt{Sink-Aware Pruning}}$, which automatically identifies and prunes unstable sinks in DLMs (prior studies usually keep sinks for AR LLMs). Without retraining, our method achieves a better quality-efficiency trade-off and outperforms strong prior pruning baselines under matched compute. Our code is available at https://github.com/VILA-Lab/Sink-Aware-Pruning.