XQuant: Breaking the Memory Wall for LLM Inference with KV Cache Rematerialization

TL;DR

XQuant introduces a novel method for LLM inference that significantly reduces memory usage by quantizing and rematerializing key-value caches, leveraging cross-layer similarities to achieve near-FP16 accuracy with substantial memory savings.

Contribution

The paper presents XQuant, a new approach that rematerializes KV caches with low-bit quantization, enabling up to 12.5× memory savings with minimal accuracy loss.

Findings

Achieves up to 7.7× memory savings with <0.1 perplexity degradation.

XQuant-CL attains up to 12.5× memory savings with 0.1 perplexity degradation.

Outperforms state-of-the-art KV cache quantization methods.

Abstract

Although LLM inference has emerged as a critical workload for many downstream applications, efficiently inferring LLMs is challenging due to the substantial memory footprint and bandwidth requirements. In parallel, compute capabilities have steadily outpaced both memory capacity and bandwidth over the last few decades, a trend that remains evident in modern GPU hardware and exacerbates the challenge of LLM inference. As such, new algorithms are emerging that trade increased computation for reduced memory operations. To that end, we present XQuant, which takes advantage of this trend, enabling an order-of-magnitude reduction in memory consumption through low-bit quantization with substantial accuracy benefits relative to state-of-the-art KV cache quantization methods. We accomplish this by quantizing and caching the layer input activations X, instead of using standard KV caching, and then rematerializing the Keys and Values on-the-fly during inference. This results in an immediate 2$\times$ memory savings compared to KV caching. By applying XQuant, we achieve up to $\sim 7.7\times$ memory savings with $<0.1$ perplexity degradation compared to the FP16 baseline. Furthermore, our approach leverages the fact that X values are similar across layers. Building on this observation, we introduce XQuant-CL, which exploits the cross-layer similarity in the X embeddings for extreme compression. Across different models, XQuant-CL attains up to 10$\times$ memory savings relative to the FP16 baseline with only 0.01 perplexity degradation, and 12.5$\times$ memory savings with only $0.1$ perplexity degradation. XQuant exploits the rapidly increasing compute capabilities of hardware platforms to eliminate the memory bottleneck, while surpassing state-of-the-art KV cache quantization methods and achieving near-FP16 accuracy across a wide range of models.