Developed an enhancement for the DarkLight1337/vllm repository focused on improving observability and memory profiling during model graph capture. The work introduced detailed CUDA memory usage logging, enabling developers to gain clearer insights into memory consumption patterns and identify potential optimization opportunities. Leveraging expertise in GPU programming and Python development, the implementation integrated logging and monitoring capabilities directly into the model graph capture process. This feature addressed the need for transparent tracking of GPU memory usage, supporting more efficient debugging and performance tuning. The contribution demonstrated a targeted approach to infrastructure improvement, emphasizing practical solutions for monitoring and optimizing deep learning workflows.