arXiv subjects

Lior Dikstein

Publications and source records attributed to Lior Dikstein.

5 recordsLinked to original sources

arXiv2026

LATMiX: Learnable Affine Transformations for Microscaling Quantization of LLMs

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.

cs.LG↗

arXiv2024

Data Generation for Hardware-Friendly Post-Training Quantization

Zero-shot quantization (ZSQ) using synthetic data is a key approach for post-training quantization (PTQ) under privacy and security constraints. However, existing data generation methods often struggle to effectively generate data suitable for hardware-friendly quantization, where all model layers are quantized. We analyze existing data generation methods based on batch normalization (BN) matching and identify several gaps between synthetic and real data: 1) Current generation algorithms do not optimize the entire synthetic dataset simultaneously; 2) Data augmentations applied during training are often overlooked; and 3) A distribution shift occurs in the final model layers due to the absence of BN in those layers. These gaps negatively impact ZSQ performance, particularly in hardware-friendly quantization scenarios. In this work, we propose Data Generation for Hardware-friendly quantization (DGH), a novel method that addresses these gaps. DGH jointly optimizes all generated images, regardless of the image set size or GPU memory constraints. To address data augmentation mismatches, DGH includes a preprocessing stage that mimics the augmentation process and enhances image quality by incorporating natural image priors. Finally, we propose a new distribution-stretching loss that aligns the support of the feature map distribution between real and synthetic data. This loss is applied to the model's output and can be adapted to various tasks. DGH demonstrates significant improvements in quantization performance across multiple tasks, achieving up to a 30% increase in accuracy for hardware-friendly ZSQ in both classification and object detection, often performing on par with real data.

cs.LG↗

arXiv2021

HPTQ: Hardware-Friendly Post Training Quantization

Neural network quantization enables the deployment of models on edge devices. An essential requirement for their hardware efficiency is that the quantizers are hardware-friendly: uniform, symmetric, and with power-of-two thresholds. To the best of our knowledge, current post-training quantization methods do not support all of these constraints simultaneously. In this work, we introduce a hardware-friendly post training quantization (HPTQ) framework, which addresses this problem by synergistically combining several known quantization methods. We perform a large-scale study on four tasks: classification, object detection, semantic segmentation and pose estimation over a wide variety of network architectures. Our extensive experiments show that competitive results can be obtained under hardware-friendly constraints.

cs.CV↗

arXiv2014

On State Dependent Broadcast Channels with Cooperation

In this paper, we investigate problems of communication over physically degraded, state-dependent broadcast channels (BCs) with cooperating decoders. Two different setups are considered and their capacity regions are characterized. First, we study a setting in which one decoder can use a finite capacity link to send the other decoder information regarding the messages or the channel states. In this scenario we analyze two cases: one where noncausal state information is available to the encoder and the strong decoder and the other where state information is available only to the encoder in a causal manner. Second, we examine a setting in which the cooperation between the decoders is limited to taking place before the outputs of the channel are given. In this case, one decoder, which is informed of the state sequence noncausally, can cooperate only to send the other decoder rate-limited information about the state sequence. The proofs of the capacity regions introduce a new method of coding for channels with cooperation between different users, where we exploit the link between the decoders for multiple-binning. Finally, we discuss the optimality of using rate splitting techniques when coding for cooperative BCs. In particular, we show that rate splitting is not necessarily optimal when coding for cooperative BCs by solving an example in which our method of coding outperforms rate splitting.

cs.IT↗

arXiv2012

MAC with Action-Dependent State Information at One Encoder

Problems dealing with the ability to take an action that affects the states of state-dependent communication channels are of timely interest and importance. Therefore, we extend the study of action-dependent channels, which until now focused on point-to-point models, to multiple-access channels (MAC). In this paper, we consider a two-user, state-dependent MAC, in which one of the encoders, called the informed encoder, is allowed to take an action that affects the formation of the channel states. Two independent messages are to be sent through the channel: a common message known to both encoders and a private message known only to the informed encoder. In addition, the informed encoder has access to the sequence of channel states in a non-causal manner. Our framework generalizes previously evaluated settings of state dependent point-to-point channels with actions and MACs with common messages. We derive a single letter characterization of the capacity region for this setting. Using this general result, we obtain and compute the capacity region for the Gaussian action-dependent MAC. The unique methods used in solving the Gaussian case are then applied to obtain the capacity of the Gaussian action-dependent point-to-point channel; a problem was left open until this work. Finally, we establish some dualities between action-dependent channel coding and source coding problems. Specifically, we obtain a duality between the considered MAC setting and the rate distortion model known as "Successive Refinement with Actions". This is done by developing a set of simple duality principles that enable us to successfully evaluate the outcome of one problem given the other.

cs.IT↗