Fuzzy PyTorch: Rapid Numerical Variability Evaluation for Deep Learning Models

Problem
This paper addresses the gap in evaluating numerical variability in deep learning models due to floating-point arithmetic, which can significantly affect model performance and reliability. As deep learning is increasingly deployed in critical applications, understanding and managing this variability is essential. The authors present Fuzzy PyTorch, a preprint framework that integrates with existing PyTorch models to facilitate rapid evaluation of numerical variability, aiming to enhance scalability and efficiency while minimizing code modifications.

Method
Fuzzy PyTorch introduces a novel library that incorporates stochastic arithmetic into the PyTorch framework through Probabilistic Rounding with Instruction Set Management. This library features two rounding modes: a stochastic rounding mode and an innovative up-down rounding mode. The integration with Verificarlo, a numerical analysis compiler, allows for efficient management of floating-point operations. The authors conducted comparative evaluations against Verrou, a leading tool in the field, demonstrating that Fuzzy PyTorch can maintain model performance while achieving significant runtime reductions. The framework is designed to handle models ranging from 1 to 341 million parameters, showcasing its scalability across various deep learning architectures.

Results
Fuzzy PyTorch achieves runtime reductions of 5x to 60x compared to Verrou, depending on the specific model and configuration used. The framework maintains comparable model performance, indicating that the introduction of stochastic rounding does not adversely affect the accuracy of the models evaluated. The authors provide empirical evidence demonstrating the framework’s effectiveness across a range of model sizes, confirming its applicability for both small and large-scale deep learning tasks.

Limitations
The authors acknowledge that while Fuzzy PyTorch offers significant improvements in runtime and scalability, it may not address all aspects of numerical variability, particularly in highly specialized or non-standard architectures. Additionally, the reliance on stochastic rounding may introduce variability that could be problematic in certain applications where deterministic outputs are critical. The paper does not discuss the potential overhead introduced by the integration with Verificarlo, nor does it explore the implications of using up-down rounding in detail.

Why it matters
Fuzzy PyTorch represents a significant advancement in the ability to assess and manage numerical variability in deep learning models, which is crucial for ensuring robust performance in real-world applications. By providing a scalable and efficient solution, this framework enables researchers and practitioners to quantify floating-point uncertainty without sacrificing computational efficiency or model accuracy. The implications of this work extend to various domains where deep learning is applied, potentially leading to more reliable and trustworthy AI systems.

Authors: Inés Gonzalez-Pepe, Hiba Akhaddar, Tristan Glatard, Yohan Chatelain
Source: arXiv:2605.25991
URL: https://arxiv.org/abs/2605.25991v1