Tong Su focused on stabilizing and improving reliability in deep learning workflows across two repositories. For intel/torch-xpu-ops, Tong addressed flaky test failures by implementing a targeted workaround that skips the Polygamma Float16 test on affected CPUs, enhancing CI stability and accelerating feedback for code reviews. In liguodongiot/transformers, Tong resolved GPU model export failures by introducing device-aware tensor handling in the convert_and_export_with_cache function, ensuring compatibility across diverse GPU configurations. Working primarily in Python with PyTorch, Tong’s contributions centered on debugging, model exporting, and robust testing, demonstrating depth in diagnosing and resolving nuanced hardware and software integration issues.