Worked on the PaddlePaddle/FastDeploy repository to enable dynamic optimization of static computational graphs by introducing piecewise CUDA Graph execution. Leveraging Python and CUDA, the developer refactored the CudaGraphPiecewiseBackend to support splitting static graphs into manageable segments, allowing for more flexible graph capture and replay during model execution. This approach established new classes and methods for managing CUDA graph states, improving code maintainability and setting the stage for future runtime optimizations. The work focused on backend development and graph optimization, laying a technical foundation for enhanced inference performance and easier extension of dynamic optimization pathways in future releases.