Design a dynamic batching mechanism for a high-throughput model serving platform to maximize GPU utilization while maintaining latency SLAs.
