I am thrilled to announce the paper “Fuzzy PyTorch: Rapid Numerical Variability Evaluation for Deep Learning Models“ has been published in Transactions on Machine Learning Research (2026).

Description

This paper introduces Fuzzy PyTorch, a framework for assessing how floating-point arithmetic uncertainties affect deep learning model behavior. It integrates stochastic arithmetic into PyTorch via PRISM (Probabilistic Rounding with Instruction Set Management), a novel library interfacing with the Verificarlo compiler. The approach supports both traditional stochastic rounding and a newly developed up-down rounding mode, achieving runtime reductions of 5× to 60× versus Verrou while scaling to models with up to 341 million parameters.

Code and Data

The code is open source and available on GitHub: big-data-lab-team/fuzzy-pytorch. The repository includes the PRISM library, the modified Verificarlo compiler, a Dockerfile to build Fuzzy PyTorch, and the scripts used to run the experiments.

Authors

  • Inés Gonzalez Pepe
  • Hiba Akhaddar
  • Tristan Glatard
  • Yohan Chatelain