PAM: Processing Across Memory Hierarchy for Efficient KV-centric LLM Serving System

TL;DR

PAM is a hierarchical, memory-aware system that optimizes key-value operations in large language model serving by balancing bandwidth and capacity through novel memory and computation coordination.

Contribution

The paper introduces PAM, a hierarchical, heterogeneous memory system with new algorithms for efficient KV processing in LLM serving, addressing bandwidth and capacity bottlenecks.

Findings

PAM significantly improves LLM serving efficiency.

PAM balances memory bandwidth and capacity effectively.

PAM demonstrates scalability in large-scale AI deployments.

Abstract

The widespread adoption of Large Language Models (LLMs) has exponentially increased the demand for efficient serving systems. With growing requests and context lengths, key-value (KV)-related operations, including attention computation and KV cache storage, have emerged as critical bottlenecks. They require massive memory bandwidth and capacity. Unfortunately, existing LLM serving systems, optimized for compute-bound workloads, fail to handle these memory-intensive operations effectively. Even with Processing-In-Memory (PIM) technology, current single-level memory designs cannot simultaneously satisfy the bandwidth and capacity requirements. To address these challenges, we propose Processing Across Memory (PAM), a KV-centric LLM serving system that coordinates heterogeneous PIM-enabled memory devices within a hierarchical architecture. PAM introduces a novel computing paradigm to balance high memory bandwidth with scalable capacity. First, PAM exploits the inherent context locality in KV access patterns to intelligently distribute KV tokens across the memory hierarchy. Second, to further exploit context locality, it introduces the PAMattention algorithm, enabling fine-grained parallel attention computation across heterogeneous PIM devices. Finally, PAM incorporates an intra-device KV mapping, inter-device KV migration interface, and an inter-device online KV scheduling algorithm to dynamically balance computational workloads. By addressing both bandwidth and capacity demands simultaneously, PAM significantly enhances the efficiency and scalability of LLM serving systems, paving the way for cost-effective, high-performance solutions in the era of large-scale AI.