Developed advanced vision-language model support and backend enhancements across ggml-org/llama.cpp and jeejeelee/vllm repositories. Delivered HunyuanVL integration by implementing M-RoPE position encoding, GGUF conversion, and a dedicated projector type, addressing model conversion and testing challenges to improve deployment reliability. Later, migrated HunyuanVL processing to align with Hugging Face Transformers 5.13, replacing deprecated APIs and introducing robust image placeholder token wrapping for accurate semantic handling. Focused on maintainability and compatibility, the work leveraged Python, PyTorch, and deep learning techniques to streamline model deployment and future-proof the VLLM pipeline, reducing upgrade risks and supporting evolving machine learning workflows.