In June 2025, Gu Shqiaio enhanced the ModelTC/LightX2V repository by developing dynamic cache and offloading mechanisms to optimize deep learning model performance. Leveraging Python and PyTorch, Gu refactored cache configuration and enabled CPU offloading for tensors, reducing memory pressure and improving throughput. The work included refining residual connection calculations and addressing issues in feature caching and offloading logic, resulting in more stable and predictable compute paths. By focusing on maintainability and scalability, Gu’s contributions improved the efficiency of the caching subsystem and established a robust foundation for future deployment, demonstrating depth in model optimization and system design.