Worked on the IBM/vllm repository to deliver a Visual Token Masking Performance Enhancement focused on optimizing visual data processing within deep learning models. The approach introduced caching for visual token IDs and added support for multiple tokens, directly addressing inefficiencies in the existing token masking workflow. By leveraging Python and PyTorch, the solution improved throughput and reduced latency for inference tasks involving large volumes of visual tokens. The work demonstrated a methodical approach to model optimization, with clear, traceable commits and impact-driven messaging. This enhancement enabled better resource utilization and supported higher concurrency for visual-token-heavy machine learning applications.