Worked on benchmarking and CI/CD improvements for PaddlePaddle/Paddle and PaddleMIX, focusing on stability and workflow modernization. Addressed a compatibility regression in the stable_diffusion_txt2img_laion400m benchmark by pinning the PyArrow version and updating shell scripts, ensuring consistent results across dygraph and dynamicToStatic configurations. Enhanced the model benchmarking CI workflow by upgrading to PaddleX models, which improved the relevance and quality of model evaluations. Used Shell and YAML to streamline benchmark triggering and reporting, and documented new processes for reproducibility. The work emphasized reliable benchmarking, reproducible CI environments, and efficient DevOps practices to support faster, data-driven product decisions.