一句话：提出 OATest，从文档测试中提取优化模式并嵌入多样的计算图上下文，合成优化感知测试以暴露 DL 编译器缺陷。
深度学习编译器依靠计算图层面的优化来提升模型部署效率，而这些优化往往针对特定的计算图结构，保证其可靠性至关重要。现有 DL 编译器模糊测试技术并未针对每个优化所匹配的图结构来生成测试，导致优化代码覆盖不足、缺陷难以暴露。本文提出 OATest：它先从文档化测试中抽取优化模式，再将这些模式嵌入多样化的计算图上下文中，从而合成优化感知的测试用例。为解决合成合法且优化感知计算图这一关键技术难题，OATest 引入两种合成策略：复用已有节点的兼容输入/输出，以及创建带兼容输入/输出的新节点，从而在模式与上下文之间建立有效连接。作者在 TVM 与 ONNXRuntime 两个主流 DL 编译器上以崩溃和不一致两类缺陷进行评估，结果表明 OATest 在发现缺陷数量和优化代码覆盖率上均显著优于当时最先进的 DL 编译器模糊测试技术，并发现 56 个此前未知的缺陷，其中 42 个已被开发者确认、24 个已修复。
Deep Learning (DL) compilers have been widely utilized to optimize DL models for efficient deployment across various hardware. Due to their vital role in the DL ecosystem, ensuring their reliability is critical. Such model optimizations are often designed to match specific computational graph structures. However, existing DL compiler fuzzing techniques do not generate tests for each optimization aware of its matched graph structures. In this paper, we propose OATest, a novel technique that synthesizes optimization-aware tests by extracting patterns from documented tests and integrating them into diverse computational graph contexts. To address the key technical challenge of synthesizing valid and optimization-aware computational graphs, OATest introduces two synthesis strategies: (1) reusing compatible inputs/outputs from existing nodes and (2) creating new nodes with compatible inputs/outputs, establishing effective connections between patterns and contexts. OATest is evaluated on two popular DL compilers, TVM and ONNXRuntime, regarding the bugs revealed by crashes and inconsistencies. The experimental results show that OATest significantly outperforms the state-of-the-art DL compiler fuzzing techniques by detecting more bugs and covering more optimization code in TVM and ONNXRuntime. Particularly, OATest uncovers 56 previously unknown bugs, 42/24 of which have been confirmed/fixed by developers.
@inproceedings{abs-2511-18918,
title = {Optimization-Aware Test Generation for Deep Learning Compilers},
author = {Qingchao Shen and Zan Wang and Haoyang Ma and Yongqiang Tian and Lili Huang and Zibo Xiao and Junjie Chen and Shing-Chi Cheung},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3744916.3773216}
}