SGLang: An Open-Source Inference Engine Transforming LLM Deployment through CPU Scheduling, Cache-Aware Load Balancing, and Rapid Structured Output Generation