This research introduces Attention Disaggregation, a technique that significantly improves throughput for large language model serving by optimizing how computational resources are allocated.

• Addresses the resource underutilization problem in current prefill-decoding disaggregation approaches
• Enhances throughput by 1.8-2.9× compared to state-of-the-art serving systems
• Maintains efficient resource utilization across both compute-intensive (prefill) and memory-intensive (decoding) phases
• Implements a practical system design that prevents performance interference while maximizing hardware efficiency

For engineering teams, this approach offers a practical path to serve more LLM requests using existing hardware resources, potentially reducing infrastructure costs and improving system responsiveness.

Injecting Adrenaline into LLM Serving: Boosting Resource Utilization and Throughput via Attention Disaggregation