发表机构
The University of Hong Kong; Huawei Technologies Co., Ltd.（香港大学; 华为技术有限公司）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究针对多模态扩散Transformer的W4A4量化局限，提出统一框架\textit{方法}，通过降低海森矩阵与激活噪声代理优化，以秩-4实现优于SVDQuant的量化性能，且成本更低、速度更快。
AI 中文摘要
在扩散Transformer中，低秩分支可通过将每个权重分解为低比特残差和高精度低秩分量，缓解4比特权重-激活（W4A4）后训练量化（PTQ）的损失。然而，现有的低秩PTQ方法要么分别优化低秩补偿和残差量化，通常需要更高的秩；要么依赖二阶权重更新，未显式建模激活量化误差，该误差在4比特量化下尤为明显。为解决这些局限，我们提出\textit{方法}（注：原文未给出具体方法名，保留\textit{方法}占位），这是一个将低秩辅助W4A4 PTQ建模为耦合校准问题的统一框架，并从联合目标中推导基于优化的求解器。消除输出侧低秩因子可得到\textit{降低的海森矩阵}，其会折扣已被低秩分量捕获的残差误差，同时引入激活噪声代理以抑制激活量化误差。在五个扩散骨干网络上，秩-4 \textit{方法}在PSNR和LPIPS指标上始终优于秩-4 SVDQuant；在SANA-1.6B、FLUX.1-schnell和FLUX.1-dev上，其以8倍更小的秩、最高6.25倍更快的量化速度，超越了秩-32 SVDQuant。此外，在Qwen3-8B大语言模型（LLM）上，秩-4 \textit{方法}将MMLU准确率从61.50%提升至68.17%，优于秩-32 SVDQuant。总体而言，\textit{方法}以显著更低的秩和量化成本实现了更优的W4A4性能。
英文摘要
In diffusion transformers, low-rank branches can mitigate 4-bit weight--activation (W4A4) post-training quantization (PTQ) loss by decomposing each weight into a low-bit residual and a high-precision low-rank component. Existing low-rank PTQ approaches, however, either optimize low-rank compensation and residual quantization separately, often requiring higher ranks, or rely on second-order weight updates without explicitly modeling activation quantization error, which becomes particularly pronounced under 4-bit quantization. To address these limitations, we present \method{}, a unified framework modeling low-rank-assisted W4A4 PTQ as a coupled calibration problem and deriving optimization-based solvers from the joint objective. Eliminating the output-side low-rank factor yields a \emph{deflated Hessian} that discounts residual errors already captured by the low-rank component, while an activation-noise surrogate is incorporated to suppress activation quantization error. Across five diffusion backbones, rank-4 \method{} consistently outperforms rank-4 SVDQuant in PSNR and LPIPS. It further surpasses rank-32 SVDQuant on SANA-1.6B, FLUX.1-schnell, and FLUX.1-dev with an $8\times$ smaller rank and up to $6.25\times$ faster quantization. Furthermore, on the Qwen3-8B LLM, rank-4 \method{} improves MMLU accuracy from 61.50\% to 68.17\% over rank-32 SVDQuant. Overall, \method{} achieves better W4A4 performance with substantially lower rank and quantization cost.
