WaterSIC: information-theoretically (near) optimal linear layer quantization

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Lifar, Egor, Savkin, Semyon, Ordentlich, Or, Polyanskiy, Yury
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866908868058021888
author Lifar, Egor
Savkin, Semyon
Ordentlich, Or
Polyanskiy, Yury
author_facet Lifar, Egor
Savkin, Semyon
Ordentlich, Or
Polyanskiy, Yury
contents This paper considers the problem of converting a given dense linear layer to low precision. The tradeoff between compressed length and output discrepancy is analyzed information theoretically (IT). It is shown that a popular GPTQ algorithm may have an arbitrarily large gap to the IT limit. To alleviate this problem, a novel algorithm, termed ''WaterSIC'', is proposed and is shown to be within a rate gap of 0.255 bits to the IT limit, uniformly over all possible covariance matrices of input activations. The key innovation of WaterSIC's is to allocate different quantization rates to different columns (in-features) of the weight matrix, mimicking the classical IT solution known as ''waterfilling''. Applying WaterSIC to the Llama and Qwen family of LLMs establishes new state-of-the-art performance for all quantization rates from 1 to 4 bits.
format Preprint
id arxiv_https___arxiv_org_abs_2603_04956
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle WaterSIC: information-theoretically (near) optimal linear layer quantization
Lifar, Egor
Savkin, Semyon
Ordentlich, Or
Polyanskiy, Yury
Machine Learning
Information Theory
This paper considers the problem of converting a given dense linear layer to low precision. The tradeoff between compressed length and output discrepancy is analyzed information theoretically (IT). It is shown that a popular GPTQ algorithm may have an arbitrarily large gap to the IT limit. To alleviate this problem, a novel algorithm, termed ''WaterSIC'', is proposed and is shown to be within a rate gap of 0.255 bits to the IT limit, uniformly over all possible covariance matrices of input activations. The key innovation of WaterSIC's is to allocate different quantization rates to different columns (in-features) of the weight matrix, mimicking the classical IT solution known as ''waterfilling''. Applying WaterSIC to the Llama and Qwen family of LLMs establishes new state-of-the-art performance for all quantization rates from 1 to 4 bits.
title WaterSIC: information-theoretically (near) optimal linear layer quantization
topic Machine Learning
Information Theory
url https://arxiv.org/abs/2603.04956