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

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Bian, Zhuohang, Wu, Feiyang, Li, Zhuoran, Ma, Teng, Zhuo, Youwei
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918512348364800
author Bian, Zhuohang
Wu, Feiyang
Li, Zhuoran
Ma, Teng
Zhuo, Youwei
author_facet Bian, Zhuohang
Wu, Feiyang
Li, Zhuoran
Ma, Teng
Zhuo, Youwei
contents Large Language Models (LLMs) are increasingly deployed in complex multi-agent applications that rely on external function calls. This workload creates severe performance challenges for the KV Cache: spatial contention leads to the eviction of critical agents' caches and temporal underutilization leaves the cache of agents stalled on long-running function calls idling in GPU memory. We present TokenCake, a KV-Cache-centric serving framework that bridges this gap by co-optimizing scheduling and memory management through an agent-aware design. TokenCake's Temporal Scheduler employs an event-driven, opportunistic policy to proactively offload idle KV Caches during function calls and uses predictive uploading to hide data transfer latency. TokenCake's Spatial Scheduler uses dynamic memory partitioning, guided by a hybrid priority metric combining graph structure and runtime state, to reserve GPU memory for critical-path agents. Our evaluation on representative multi-agent benchmarks shows that TokenCake reduces end-to-end latency by over 47.06% and improves effective GPU memory utilization by up to 16.9% compared to vLLM.
format Preprint
id arxiv_https___arxiv_org_abs_2510_18586
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications
Bian, Zhuohang
Wu, Feiyang
Li, Zhuoran
Ma, Teng
Zhuo, Youwei
Distributed, Parallel, and Cluster Computing
C.4; D.4.2; I.2.11
Large Language Models (LLMs) are increasingly deployed in complex multi-agent applications that rely on external function calls. This workload creates severe performance challenges for the KV Cache: spatial contention leads to the eviction of critical agents' caches and temporal underutilization leaves the cache of agents stalled on long-running function calls idling in GPU memory. We present TokenCake, a KV-Cache-centric serving framework that bridges this gap by co-optimizing scheduling and memory management through an agent-aware design. TokenCake's Temporal Scheduler employs an event-driven, opportunistic policy to proactively offload idle KV Caches during function calls and uses predictive uploading to hide data transfer latency. TokenCake's Spatial Scheduler uses dynamic memory partitioning, guided by a hybrid priority metric combining graph structure and runtime state, to reserve GPU memory for critical-path agents. Our evaluation on representative multi-agent benchmarks shows that TokenCake reduces end-to-end latency by over 47.06% and improves effective GPU memory utilization by up to 16.9% compared to vLLM.
title TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications
topic Distributed, Parallel, and Cluster Computing
C.4; D.4.2; I.2.11
url https://arxiv.org/abs/2510.18586