Worked on extending the diffusion engine plugin architecture within the vllm-project/vllm-omni repository, focusing on enabling support for out-of-tree hardware backends. Developed a modular plugin extension in Python that decouples backend implementations from the core diffusion logic, allowing for easier integration of new hardware models and facilitating accelerated experimentation. Emphasized backend development and plugin architecture, ensuring the system remains scalable and maintainable as new hardware is introduced. Collaborated with other contributors through signed-off commits and code reviews, and incorporated unit testing to maintain code quality. The work established a foundation for future extensibility and streamlined hardware integration workflows.