Worked on expanding hardware compatibility in the PaddlePaddle/PaddleCFD repository by delivering custom device support for CSR sum and mean operations. This involved updating the select.cpp file in C++ to recognize custom devices through the is_custom_device() check, allowing these operations to execute on both standard GPUs and metax-gpu hardware. The approach leveraged device abstraction and GPU computing skills to broaden deployment options for CFD workloads, aligning with the project’s roadmap for heterogeneous hardware acceleration. The work focused on integrating custom-device checks into core CSR operations, enhancing PaddleCFD’s flexibility for diverse computational fluid dynamics environments without introducing new bugs.