Developed FP8 quantization support for ERNIE expert weights in the PaddlePaddle/ERNIE repository, enabling parameter storage in FP8 when BF16 is unavailable. This work introduced a new training callback, which was integrated into the pre-training trainer to improve memory efficiency and potentially accelerate distributed training. The implementation focused on deep learning model optimization using Python, with careful alignment to existing training pipelines to maintain compatibility and workflow consistency. By enabling FP8 precision for expert weights, the changes addressed scalability challenges for large models, laying the foundation for faster iteration and reduced memory footprint on supported hardware without disrupting established processes.