Developed a scalable vision-language inference optimization for the jeejeelee/vllm repository, focusing on efficient processing of multimodal inputs. Built a CUDA graph-based budgeted inference workflow for the Vision Encoder, enabling dynamic batching and improved throughput for images with varying token budgets. The approach utilized a CUDA graph manager to capture and replay execution graphs, reducing per-inference overhead and aligning with production scalability requirements. Leveraged Python and PyTorch to integrate deep learning and multimodal processing techniques, ensuring compatibility with existing machine learning pipelines. The work addressed performance bottlenecks in vision encoder inference, delivering a robust solution for high-throughput, budgeted inference scenarios.