Tangram: Unlocking Non-Uniform KV Cache for Efficient Multi-turn LLM Serving

2026-06-04

AI Summary

This work addresses the challenges posed by non-uniform KV caching in multi-turn large language model (LLM) serving, where dynamic cache patterns lead to memory fragmentation, scheduling overhead, and degraded computational efficiency. To overcome these issues, the authors propose an LLM serving system that efficiently supports non-uniform KV caches through deterministic memory budget allocation to eliminate prefill stalls, a grouped attention head paging mechanism for high-efficiency physical memory reclamation, and static load balancing with preloading to avoid runtime overhead. This system is the first to enable practical and efficient deployment of non-uniform KV caching without compromising model accuracy, achieving up to 2.6× higher throughput compared to existing baselines.

Abstract

Multi-turn Large Language Model (LLM) serving is critical for consistent user experiences, yet the linear growth of the Key-Value (KV) cache imposes significant pressure on GPU memory and bandwidth. Non-uniform KV compression effectively preserves more information by considering the individual importance of each KV cache. However, such KV cache heterogeneity introduces various systemic challenges - including memory fragmentation, scheduling complexities, and diminished kernel utilization - which collectively lead to significant inefficiencies in existing LLM serving systems. To overcome these challenges, we present Tangram, a novel serving system designed to make Non-uniform KV caches practical. Tangram addresses systemic inefficiencies through three core techniques: (1) Deterministic Budget Allocation assigns a static memory footprint to each head based on its intrinsic pattern, entirely eliminating dynamic scheduling overhead and prefill stalls; (2) Head Group Page clusters attention heads with similar retention demands and manages them with independent, vectorized page tables, thereby maximizing physical memory reclamation; and (3) Ahead-of-Time (AOT) Load Balancing leverages static budget profiles to ensure uniform GPU utilization without runtime overhead. Experimental results show that Tangram improves throughput by up to 2.6x compared to existing baselines, while fully preserving model accuracy. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.

Problem

• KV cache
• non-uniform compression
• LLM serving
• memory fragmentation
• scheduling complexity

Innovation

• Non-uniform KV Cache
• LLM Serving
• Memory Management
• Attention Head Grouping
• Load Balancing