Worked on refactoring multimodal processor registration in the Blaizzy/mlx-vlm repository to improve modularity and deployment flexibility. Focused on decoupling Torch dependencies for the qwen3_5 model by introducing an install_auto_processor_patch mechanism and registering Qwen3VLProcessor, enabling deployment in environments with minimal Torch requirements. The approach removed direct AutoImageProcessor and AutoProcessor registrations for hunyuan_vl and kimi_vl, streamlining processor management and reducing startup overhead. Utilized Python for dependency management, model integration, and processor registration, resulting in a cleaner, more maintainable codebase. The work addressed targeted deployment challenges and enhanced the repository’s support for diverse multimodal model configurations.