Yi Ding enhanced the intel/torch-xpu-ops repository by refactoring the scaled dot product attention logic to improve PyTorch compatibility for XPU devices. He introduced a device support stub, streamlining backend selection and removing an unimplemented fallback, which reduced conditional complexity and improved maintainability. Working primarily in C++ and leveraging GPU programming and machine learning expertise, Yi moved key decision logic to the PyTorch layer, aligning the integration more closely with upstream frameworks. This focused engineering effort resulted in a cleaner, more efficient attention mechanism for XPU, laying groundwork for future enhancements and ensuring more reliable performance across supported devices.