Equivalence Checking of ML GPU Kernels

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Dubey, Kshitij, Driscoll, Benjamin, Wei, Anjiang, Kayal, Neeraj, Sharma, Rahul, Aiken, Alex
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911273258582016
author Dubey, Kshitij
Driscoll, Benjamin
Wei, Anjiang
Kayal, Neeraj
Sharma, Rahul
Aiken, Alex
author_facet Dubey, Kshitij
Driscoll, Benjamin
Wei, Anjiang
Kayal, Neeraj
Sharma, Rahul
Aiken, Alex
contents With the rapid progress of deep learning and large language models (LLMs), companies now spend enormous sums executing GPU kernels. These kernels have, therefore, become prime targets for aggressive optimization. Recent efforts increasingly leverage LLMs to generate GPU kernels, but make no formal guarantees about the generated kernels. We present the first equivalence checker for GPU kernels and use it to formally verify the correctness of machine learning (ML) kernels optimized by hand, by LLMs, and by compilers. We show that our equivalence checker is sound and, for a well-defined class of GPU kernels which includes the programs of interest, complete. Our implementation, VOLTA, can verify ML computations such as convolutions, matrix multiplications, and various attention mechanisms.
format Preprint
id arxiv_https___arxiv_org_abs_2511_12638
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Equivalence Checking of ML GPU Kernels
Dubey, Kshitij
Driscoll, Benjamin
Wei, Anjiang
Kayal, Neeraj
Sharma, Rahul
Aiken, Alex
Programming Languages
With the rapid progress of deep learning and large language models (LLMs), companies now spend enormous sums executing GPU kernels. These kernels have, therefore, become prime targets for aggressive optimization. Recent efforts increasingly leverage LLMs to generate GPU kernels, but make no formal guarantees about the generated kernels. We present the first equivalence checker for GPU kernels and use it to formally verify the correctness of machine learning (ML) kernels optimized by hand, by LLMs, and by compilers. We show that our equivalence checker is sound and, for a well-defined class of GPU kernels which includes the programs of interest, complete. Our implementation, VOLTA, can verify ML computations such as convolutions, matrix multiplications, and various attention mechanisms.
title Equivalence Checking of ML GPU Kernels
topic Programming Languages
url https://arxiv.org/abs/2511.12638