跳转至

文章背景与核心概要

随着大型语言模型（LLMs）向极限压缩（两比特或更低）迈进，分块后训练量化（PTQ）已成为一种标准方法。然而，现有的“跨块”方法通常采用单向连续扫描的方式处理Transformer块，这往往会导致量化误差不断累积且永远无法被重新修正。

本文引入了一种名为交错式跨块量化（ICBQ）的新型调度策略，该策略对标准的顺序基准进行了改进。通过对“缝合处”（连续分块之间的边界对）进行两次精炼——一次在一个分块结束时，另一次在下一个分块开始时——ICBQ有效地缓解了误差传播。该方法保持了局部双块目标的效率，并与现有的校准流程兼容。理论分析表明，这种方法能够将残差控制在与网络深度无关的有界范围内，实验结果也证明了其在三值、3位和2位量化任务中显著提升了困惑度（Perplexity）表现。

从单向扫描到边界缝合：交错式跨块后训练量化 (From Sweep to Seam: Interleaved Cross-Block Post-Training Quantization)

作者： Achille Jacquemond, Yuma Ichikawa, Akira Sakai
日期： 2026年8月10日
标识符： arXiv:2608.09595 [cs.AI]

摘要 (Summary)

As Large Language Models (LLMs) push toward extreme compression (two bits or fewer), block-wise post-training quantization (PTQ) has become a standard approach. Existing "cross-block" methods typically process Transformer blocks in a single sequential sweep, which often leads to the accumulation of quantization errors that are never revisited.

随着大型语言模型（LLMs）向极限压缩（两比特或更低）迈进，分块后训练量化（PTQ）已成为一种标准方法。现有的“跨块”方法通常采用单向连续扫描的方式处理Transformer块，这往往会导致量化误差不断累积且永远无法被重新修正。

This paper introduces Interleaved Cross-Block Quantization (ICBQ), a novel scheduling strategy that improves upon the standard sequential baseline. By refining the "seam" (the boundary pair between consecutive chunks) twice—once at the end of a chunk and again at the start of the next—ICBQ effectively mitigates error propagation. The method maintains the efficiency of local two-block objectives and is compatible with existing calibration pipelines. Theoretical analysis suggests that this approach keeps residuals bounded independently of network depth, and experimental results demonstrate significant improvements in perplexity for ternary, 3-bit, and 2-bit quantization tasks.

本文引入了一种名为交错式跨块量化（ICBQ）的新型调度策略，该策略对标准的顺序基准进行了改进。通过对“缝合处”（连续分块之间的边界对）进行两次精炼——一次在一个分块结束时，另一次在下一个分块开始时——ICBQ有效地缓解了误差传播。该方法保持了局部双块目标的效率，并与现有的校准流程兼容。理论分析表明，这种方法能够将残差控制在与网络深度无关的有界范围内，实验结果也证明了其在三值、3位和2位量化任务中显著提升了困惑度表现。

核心贡献 (Key Contributions)

  • Interleaved Scheduling: Proposes a "seam-revisit" mechanism that refines boundary blocks twice, preventing the one-way error accumulation inherent in standard sequential sweeps.
  • 交错调度： 提出了一种“缝合处重访”机制，对边界块进行两次精炼，从而防止了标准顺序扫描中固有的单向误差累积。
  • Theoretical Foundation: Derives a depth-wise upper-bound comparison, proving that seam revisits help control the propagation of quantization errors.
  • 理论基础： 推导了深度方向的上界比较，证明了重访缝合处有助于控制量化误差的传播。
  • Versatility: Demonstrates that the ICBQ framework is robust, showing performance gains in ternary quantization and compatibility with established methods like GPTQ (3-bit and 2-bit).
  • 通用性： 证明了ICBQ框架的稳健性，展示了其在三值量化中的性能提升，以及与GPTQ（3位和2位）等现有方法的兼容性。
  • Efficiency: Retains the local two-block objective, ensuring that the method remains computationally feasible and compatible with existing calibration input workflows.
  • 高效性： 保留了局部的双块优化目标，确保该方法在计算上依然可行，并与现有的校准输入工作流保持兼容。

元数据与访问链接 (Metadata & Access)

  • Subjects: Artificial Intelligence (cs.AI)
  • 主题： 人工智能 (cs.AI)

外部资源 (External Resources)