Developed and integrated the LLaVA-OneVision-2 multimodal model into the vLLM repository, expanding the framework’s capabilities to support advanced vision-language inference. This work involved designing architecture hooks and implementing custom video processing backends that handle both frame-based and codec-based pipelines, enabling flexible and efficient multimodal workflows. Leveraging expertise in backend development, computer vision, and multimodal AI, the integration was accomplished using Python and PyTorch, with careful tracking and merging of upstream changes. The result broadened model support within vLLM, reduced deployment complexity, and enabled end-to-end inference for new vision-language use cases without introducing new bugs.