Demystifying the Silence of Correctness Bugs in PyTorch Compiler

View PDF HTML (experimental) Abstract:Performance optimization of AI infrastructure is key to the fast adoption of large language models (LLMs). The PyTorch compiler (this http URL), a core optimization tool for deep learning (DL) models (including LLMs), has received due attention. However, this http URL is prone to correctness bugs, which cause incorrect outputs of compiled DL models without triggering exceptions, crashes, or warnings. These bugs pose a serious threat to the reliability of downstream LLM applications. Data from the PyTorch community shows that 19.2% of high-priority issues are incorrect outputs of compiled DL models induced by this http URL bugs, the second-most-common bug category (only behind program crashes at 19.57%). However, no systematic study has been conducted to specifically characterize and thereby detect these bugs. In this paper, we present the first empirical study of the correctness bugs in this http URL, examine their characteristics, and assess the effectiveness of existing fuzzers in detecting them. Based on our findings, we propose a proof-of-concept testing technique named AlignGuard, tailored specifically for detecting correctness bugs in this http URL. AlignGuard incorporates bug characteristics distilled from our empirical study, applying LLM-based test mutation to existing test cases for correctness bug detection. At the time of writing, AlignGuard has successfully detected 23 new correctness bugs in recent this http URL. All these bugs have been confirmed or fixed by the PyTorch development team, and over half (14/23) of them are even marked as high-priority bugs, underscoring the usefulness of our technique. Subjects: Software Engineering (cs.SE); Artificial Intelligence (cs.AI) ACM classes: D.2.5; I.2.5 Cite as: arXiv:2604.08720 [cs.SE] (or arXiv:2604.08720v1 [cs.SE] for this version) https://doi.org/10.48550/arXiv.2604.08720 arXiv-issued DOI via DataCite (pending registration) Submission history From: Meiziniu Li [view email] [v1] Thu, 9 Apr 2026 19:13:15 UTC (358 KB)