Developed a configurable GPU memory allocation feature for the NVIDIA-NeMo/Eval repository, enabling users to specify the fraction of GPU memory dedicated to vLLM deployments. This work involved updating YAML-based configuration schemas and enhancing Markdown documentation to support the new parameter, making resource budgeting more precise for inference workloads. By parameterizing GPU memory utilization, the solution improved deployment flexibility and scalability, allowing for more predictable performance and efficient resource sharing in GPU-bound environments. The approach demonstrated practical application of configuration management and DevOps skills, focusing on maintainability and reproducibility while addressing the need for cost-efficient, scalable inference infrastructure.