In July 2025, Liyurui developed FP8 quantization support for ERNIE expert weights in the PaddlePaddle/ERNIE repository, targeting improved memory efficiency and scalability for large models. By introducing a new training callback and integrating it into the pre-training trainer, Liyurui enabled the use of FP8-based parameter storage when BF16 is unavailable, optimizing both memory usage and potential training speed. The implementation, written in Python and leveraging deep learning and model optimization techniques, maintained compatibility with existing distributed training workflows. This work addressed the challenge of efficient parameter handling, laying a foundation for faster iteration and reduced memory footprint on supported hardware.