通过剪枝和动态解码，让扩散语言模型推理快68倍
Streaming-dLLM: Accelerating Diffusion LLMs via Suffix Pruning and Dynamic Decoding
- 用掩码剪枝去除冗余上下文，减少空间计算开销
- 根据置信度动态跳过冗余去噪步骤，提速68.2倍
- 无需训练，适合追求高效生成的部署场景
扩散大语言模型（dLLMs）通过并行解码和双向注意力，在自然语言生成中展现出优于自回归模型的全局连贯性。尽管已有工作通过缓存复用或启发式解码加速推理，但忽略了块级扩散过程中的内在低效：空间上对信息稀疏的后缀区域均匀建模造成冗余，时间上固定去噪调度导致无效迭代。为此，我们提出Streaming-dLLM，一种无需训练的框架，从空间与时间维度同步优化推理效率。空间上，引入衰减引导的后缀建模，通过剪枝冗余掩码标记近似完整上下文；时间上，采用基于置信度的动态策略与提前退出机制，使已收敛的标记跳过不必要的去噪迭代。大量实验表明，Streaming-dLLM在保持生成质量的同时实现最高68.2倍的加速，验证了其在扩散解码中的有效性。代码已开源。
原文摘要 · Abstract (English)
Diffusion Large Language Models (dLLMs) offer a compelling paradigm for natural language generation, leveraging parallel decoding and bidirectional attention to achieve superior global coherence compared to autoregressive models. While recent works have accelerated inference via KV cache reuse or heuristic decoding, they overlook the intrinsic inefficiencies within the block-wise diffusion process. Specifically, they suffer from spatial redundancy by modeling informative-sparse suffix regions uniformly and temporal inefficiency by applying fixed denoising schedules across all the decoding process. To address this, we propose Streaming-dLLM, a training-free framework that streamlines inference across both spatial and temporal dimensions. Spatially, we introduce attenuation guided suffix modeling to approximate the full context by pruning redundant mask tokens. Temporally, we employ a dynamic confidence aware strategy with an early exit mechanism, allowing the model to skip unnecessary iterations for converged tokens. Extensive experiments show that Streaming-dLLM achieves up to 68.2X speedup while maintaining generation quality, highlighting its effectiveness in diffusion decoding. The code is available at https://github.com/xiaoshideta/Streaming-dLLM.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。