Developed and delivered a new operator, update_attn_mask_offsets, for the PaddlePaddle/FastDeploy repository, focusing on efficient attention mask handling for XPU hardware. The implementation optimized the management of attention mask offsets during sequence processing, particularly benefiting models that use padding and variable-length sequences. By supporting both CPU and XPU execution, the operator ensured consistent performance across different hardware environments. The work involved C++ and Python, leveraging CUDA programming and XPU development skills. Code quality was maintained through unit testing and adherence to pre-commit formatting standards, resulting in a robust feature that improves inference efficiency for sequence-based deep learning models.