Worked on stabilizing the omni diffusion backend in the vllm-project/vllm-omni repository, focusing on improving the robustness of local attention mechanisms. Addressed a recurring warning issue by introducing a backend preference attribute within the Attention class, ensuring this preference was consistently set and utilized during local attention execution. This technical approach reduced log noise and improved configuration correctness, enhancing reliability in production-like environments. The work involved targeted debugging and refactoring using Python, PyTorch, and deep learning techniques. No new user-facing features were released during this period, but the groundwork was laid for future enhancements through improved backend stability.