Tokencake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications

2025-10-21

AI Summary

To address two critical bottlenecks in LLM-driven multi-agent systems—KV cache spatial contention (where high-priority agents’ caches are frequently evicted) and low temporal utilization (caused by prolonged tool-call latencies leading to cache idleness)—this paper proposes a KV-cache-centric service framework. Our method introduces: (1) an agent-behavior-aware dynamic memory partitioning mechanism that enforces cache priority for mission-critical agents; and (2) a proactive cache offloading and prefetching strategy guided by predicted function-call waiting times, enabling spatiotemporal coordination of cache resource allocation. Evaluated on representative multi-agent benchmarks, our framework reduces end-to-end latency by 47.06% and improves effective GPU memory utilization by 16.9% over vLLM.

Technology Category

• Multiagent Systems: Agent Communication
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Agent Architectures

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large Language Models (LLMs) are increasingly deployed in complex multi-agent applications that use external function calls. This workload creates severe performance challenges for the KV Cache: space contention leads to the eviction of critical agents' caches and time underutilization leaves the cache of agents stalled on long-running tool calls idling in GPU memory. We present Tokencake, a KV-Cache-centric serving framework that co-optimizes scheduling and memory management with an agent-aware design. Tokencake's Space Scheduler uses dynamic memory partitioning to shield critical agents from contention, while its Time Scheduler employs a proactive offload and predictive upload mechanism to repurpose GPU memory during function call stalls. Our evaluation on representative multi-agent benchmarks shows that Tokencake can reduce end-to-end latency by over 47.06%, improve effective GPU memory utilization by up to 16.9% compared to vLLM.

Problem

• Addresses KV cache space contention in multi-agent LLM applications
• Solves GPU memory underutilization during agent tool call stalls
• Optimizes scheduling and memory management for LLM-based agents

Innovation

• Dynamic memory partitioning shields critical agents from contention
• Proactive offload and upload repurposes GPU memory during stalls
• Co-optimizes scheduling and memory management for multi-agent applications