Worked on the optuna/optuna repository to address stability issues in GPU-to-CPU tensor handling within the Gaussian Process optimization path. Focused on a targeted bug fix that ensures tensors are explicitly transferred to the CPU before conversion to NumPy arrays, preventing runtime errors when tensors reside on the GPU. This change improved reliability for users leveraging GPU-backed workflows in data science and machine learning applications. The solution was implemented in Python, following established review and testing practices, and delivered as an isolated, well-documented patch. The work enhanced code health and traceability while supporting broader adoption of GPU-accelerated optimization workflows.