Contributed to the PaddlePaddle and PaddleCustomDevice repositories by enhancing documentation and improving GPU kernel stability over a two-month period. Focused on refining example code for tensor operations using C++ and Python, ensuring clarity, correctness, and adherence to code style and xdoctest standards. Addressed documentation gaps by updating complex number representations and output formatting, making examples more accessible and testable for developers. Improved the reliability of CUDA-based matrix power kernels, targeting stability and integration within GPU-accelerated workflows. Emphasized code quality, maintainability, and onboarding experience, resulting in clearer documentation and more robust GPU programming support for the Paddle ecosystem.