Developed and integrated the flash_attention_backward operation for the FlagOpen/FlagGems repository, targeting improved efficiency in backpropagation for attention mechanisms within deep learning models. Leveraged CUDA and PyTorch to implement the new operator, focusing on both performance optimization and correctness. Enhanced the project’s testing infrastructure by adding dedicated benchmarks and refactoring test scripts, including renaming and reorganizing files for greater clarity and maintainability. Aligned all changes with SiliconFlow standards, ensuring consistency and reliability. This work enabled faster iteration cycles and reduced regression risk, directly supporting business goals through improved model training speed and more robust, maintainable code for future development.