Ofir Gordon

Ofir Gordon, Elad Cohen, Hai Victor Habi et al.

Quantization is a key method for deploying deep neural networks on edge devices with limited memory and computation resources. Recent improvements in Post-Training Quantization (PTQ) methods were achieved by an additional local optimization process for learning the weight quantization rounding policy. However, a gap exists when employing network-wise optimization with small representative datasets. In this paper, we propose a new method for enhanced PTQ (EPTQ) that employs a network-wise quantization optimization process, which benefits from considering cross-layer dependencies during optimization. EPTQ enables network-wise optimization with a small representative dataset using a novel sample-layer attention score based on a label-free Hessian matrix upper bound. The label-free approach makes our method suitable for the PTQ scheme. We give a theoretical analysis for the said bound and use it to construct a knowledge distillation loss that guides the optimization to focus on the more sensitive layers and samples. In addition, we leverage the Hessian upper bound to improve the weight quantization parameters selection by focusing on the more sensitive elements in the weight tensors. Empirically, by employing EPTQ we achieve state-of-the-art results on various models, tasks, and datasets, including ImageNet classification, COCO object detection, and Pascal-VOC for semantic segmentation.

Ofir Gordon, Lior Dikstein, Arnon Netzer et al.

Post-training quantization (PTQ) is a widely used approach for reducing the memory and compute costs of large language models (LLMs). Recent studies have shown that applying invertible transformations to activations can significantly improve quantization robustness by reducing activation outliers; however, existing approaches are largely restricted to rotation or Hadamard-based transformations. Moreover, most studies focused primarily on traditional quantization schemes, whereas modern hardware increasingly supports the microscaling (MX) data format. Attempts to combine both showed severe performance degradation, leading prior work to introduce assumptions on the transformations. In this work, we take a complementary perspective. First, we provide a theoretical analysis of transformations under MX quantization by deriving a bound on the quantization error. Our analysis emphasizes the importance of accounting for both the activation distribution and the underlying quantization structure. Building on this analysis, we propose LATMiX, a method that generalizes outlier reduction to learnable invertible affine transformations optimized using standard deep learning tools. Experiments show consistent improvements in average accuracy for MX low-bit quantization over strong baselines on a wide range of zero-shot benchmarks, across multiple model sizes.