Worked on the vllm-spyre repository to address a critical issue in the continuous batching model runner, focusing on padding and memory management for decode operations. Using Python and backend development skills, implemented a minimum batch size of two by padding input tokens and dynamically allocating memory blocks when batch sizes fell below the threshold. Developed an automatic cleanup mechanism to free padding resources, ensuring efficient memory usage and preventing leaks. This optimization reduced peak memory footprint and improved throughput predictability, contributing to greater production stability for continuous decoding workloads. Demonstrated expertise in batch processing, model optimization, and low-level resource management.