ANPRED: Anchor-based Precision Error Detection and Localization Framework for Deep Learning Libraries

Deep learning libraries are the backbone of constructing deep learning systems, yet their operator implementations are vulnerable to numerical precision errors caused by floating-point limitations, implementation choices, and performance-oriented optimizations. Such errors may be introduced at intermediate execution stages, then propagate through long operator chains, and eventually lead to inconsistent or degraded model behaviors across different libraries. A series of works are devoted to detect and localize such precision issues; nevertheless, most existing approaches focus on examining the final outputs of the model, offering limited observability into where precision errors introduced and how they evolve during execution. To bridge this gap, in this paper, we presents ANPRED, an anchor-based precision error detection and localization framework for deep learning libraries. ANPRED conducts semantically equivalent model implementations with stage-boundary anchors to expose intermediate numerical values and applies multi-dimensional differential analysis to detect, localize, and characterize precision errors. We evaluate ANPRED on six models across three mainstream deep learning libraries and compare it with threshold-based output-level baselines. The results show that ANPRED preserves the precision-error detection capability of output-level baselines while revealing substantially more anchor-level discrepancies that are invisible at the final output. Moreover, ANPRED localizes precision errors to specific execution stages and shows that such errors are stage-concentrated, dimension-sensitive, and non-monotonic in propagation. These findings demonstrate that anchor-level analysis provides a more informative basis for understanding and diagnosing library-related precision errors. The relevant code has been uploaded and is available in this repository. Due to space and size limitation, some typical data is uploaded, but complete data is supported, and the author can be contacted to obtain it if needed.