cs.AR提交于 2026-10-09已译

DynaTE:基于动态令牌执行的扩散大语言模型加速架构

DynaTE: Accelerating Diffusion LLMs via Dynamic Token Execution

Minghan Jiang · Jiayi Wang · Shuaiting Li · Haibin Shen · Kejie Huang

中文摘要

基于扩散机制的大语言模型(dLLMs)通过支持双向并行优化,缓解了自回归(AR)生成中的串行解码瓶颈,近年来成为 AR LLM 的一种有前景的替代方案。然而,其并行迭代优化与针对串行解码优化的 AR 加速器不匹配,且离散令牌生成也与面向连续去噪的 DiT 加速器不同。已有 dLLM 加速器探索了面向负载的特定优化,以降低词表处理开销和去噪迭代间的冗余计算,但这些方法仍保留全部令牌的并行执行,未考虑不同令牌在优化效用和执行需求上的差异。提出 DynaTE,一种软硬件协同设计架构,可在 dLLM 解码过程中使加速器执行动态适配不断变化的令牌状态。DynaTE 首先通过跳过低效用令牌计算实现自适应令牌执行,同时采用维度可重构的 PE 阵列在变化的活跃令牌模式下保持高利用率。其次,DynaTE 通过 FLDD 挖掘动态令牌依赖,在当前迭代内优化少量局部相关令牌,从而减少整体去噪迭代次数,并配合 Merge–Split–Merge 数据流隐藏由此产生的串行开销。第三,流式词表引擎将来自 LM head 的多条令牌流交织调度,以适应由选择性令牌计算和不均匀词表选择需求引起的输出变化。在两种代表性 dLLM 上的评估表明,DynaTE 相对于当前最优 dLLM 加速器实现了 2.05–2.78× 加速和 2.99–3.93× 能效提升,相对 Jetson AGX Orin 实现 2.55× 加速和 6.07× 能效提升。

关键要点

01问题:已有 dLLM 加速器对所有令牌一律并行执行,未考虑令牌间优化效用与执行需求的差异,带来词表处理与去噪迭代的冗余开销。
02方法:提出软硬件协同架构 DynaTE,通过跳过低效用令牌实现自适应执行,并以维度可重构 PE 阵列维持高利用率。
03方法:利用 FLDD 在当前迭代内优化少量局部相关令牌以减少去噪迭代次数,配合 Merge–Split–Merge 数据流隐藏串行开销。
04方法:流式词表引擎交织调度 LM head 多条令牌流,缓解选择性令牌计算与不均匀词表选择带来的输出波动。
05结果:在两种代表 dLLM 上相对 SOTA dLLM 加速器取得 2.05–2.78× 加速与 2.99–3.93× 能效提升,相对 Jetson AGX Orin 取得 2.55× 加速与 6.07× 能效提升。

解读

尚无解读。

原始英文摘要

arXiv:2610.11284v1 Announce Type: cross Abstract: Diffusion-based LLMs (dLLMs) have recently emerged as a promising alternative to autoregressive (AR) LLMs by enabling bidirectional parallel refinement, alleviating the sequential decoding bottleneck of AR generation. However, their parallel iterative refinement mismatches AR accelerators optimized for sequential decoding and their discrete token generation differs from DiT accelerators designed for continuous denoising. Recent dLLM accelerators have explored workload-specific optimizations to reduce vocabulary processing overhead and redundant computation across denoising iterations. However, these approaches retain all tokens in parallel execution, despite varying token refinement utility and execution requirements. This paper presents DynaTE, a hardware--software co-design architecture that dynamically adapts accelerator execution to evolving token states during dLLM decoding. DynaTE first enables adaptive token execution by skipping low-utility token computation, while a dimension-reconfigurable PE array maintains high utilization under varying active-token patterns. Second, DynaTE exploits dynamic token dependencies through FLDD to refine a small number of locally dependent tokens within the current iteration, reducing the overall number of denoising iterations, while a Merge--Split--Merge dataflow hides the resulting serial overhead. Third, a streaming vocabulary engine interleaves multiple token streams from the LM head to accommodate irregular output variations caused by selective token computation and uneven vocabulary-selection demands. Evaluated on two representative dLLMs, DynaTE achieves 2.05--2.78$\times$ speedup and 2.99--3.93$\times$ higher energy efficiency over state-of-the-art dLLM accelerators, while delivering 2.55$\times$ speedup and 6.07$\times$ higher energy efficiency over Jetson AGX Orin.