Developed and delivered Dynamic Batch Optimization (DBO) for ROCm within the jeejeelee/vllm repository, focusing on enhancing large-model inference performance. The work centered on adaptive resource allocation between communication and computation units, optimizing throughput and reducing latency for ROCm-enabled hardware. Using Python and leveraging GPU programming and performance optimization skills, the implementation improved batch processing efficiency and laid the foundation for more effective ROCm deployments. The feature was fully integrated and prepared for production testing, with no major bug fixes required during the period. This contribution addressed resource management challenges and supported broader adoption of ROCm for high-performance inference tasks.