2026-06-08
OffQ: Taming Structured Outliers in LLM Quantization by Offsetting
针对低比特量化中激活异常值问题，提出OffQ偏移机制，缓解性能下降
Authors: Haoqi Wang, Lorenz K. Mueller, Jiawei Zhuang, Mathieu Salzmann, Lukas Cavigelli | Date: 2026-06-05
Link: http://arxiv.org/abs/2606.07116v1
Abstract
Low-bit quantization has been widely adopted to accelerate the inference of large language models (LLMs) by significantly reducing computational cost and memory usage. However, activation outliers pose a major challenge to effective quantization, often leading to notable performance degradation. In this paper, we introduce OffQ, a method designed to mitigate activation outliers in low-bit quantization through a novel offsetting mechanism. Specifically, OffQ first identifies a low-dimensional outlier subspace in the activations using a proposed top-1 PCA, and then concentrates high-magnitude activations into 1 channel via rotation. OffQ then absorbs this concentrated outlier channel by converting its magnitude into a shared offset, thereby reducing the standard deviation of the activations. This offsetting strategy enables effective W4A4KV4 quantization of LLMs using deployment-friendly uniform-grid and uniform-precision quantization. Extensive experiments across diverse LLM architectures and benchmarks demonstrate that OffQ outperforms state-of-the-art baselines, consistently improving model accuracy while preserving low-bit efficiency.SABLE: GPU-Based Power Flow Accelerator for Sparsity-Aware Batched Learning
解决GPU稀疏批处理学习中的功率流计算挑战，提出SABLE加速器，提升效率
Authors: Suho Park, Keunju Song, Hongseok Kim | Date: 2026-06-05
Link: http://arxiv.org/abs/2606.07099v1
Abstract
Recent studies have developed GPU-based approaches for solving AC power flow and successfully applied them to standalone power flow problems. However, integrating these approaches into modern differentiable learning frameworks while preserving sparsity remains challenging. To this end, we present SABLE, a GPU-based sparse batched power flow accelerator for differentiable learning via an implicit power flow layer. SABLE leverages a block-diagonal embedding that reformulates batched three-dimensional Jacobians as a fixed-pattern two-dimensional sparse template that is shared across PyTorch, CuPy, and cuDSS. This common template enables zero-copy interoperability and memory-efficient sparse reuse across the software stack. On top of this representation, SABLE accelerates repeated power flow computations through reusable sparse templates, custom GPU kernels, a cuDSS-based sparse-direct LU solver, and mixed-precision techniques. Extensive experiments show that SABLE improves standalone power flow solving throughput by up to 253.4$\times$ over pandapower and 5.7$\times$ over ExaPF. In end-to-end training, evaluated on AC optimal power flow learning models based on DC3 and DeepLDE, SABLE expands the feasible training batch range by up to 64$\times$ and improves training throughput by up to 206.7$\times$ over the corresponding baseline.SigmaScale: LLM Compression with SVD-based Low-Rank Decomposition and Learned Scaling Matrices
针对LLM压缩，提出SigmaScale学习缩放矩阵方法，降低有效秩提升压缩效果
Authors: Ernests Lavrinovics, Marco Letizia, Roy Janco, Shai Segal, Johannes Bjerva, Maurizio Pierini | Date: 2026-06-05
Link: http://arxiv.org/abs/2606.07098v1