Stochastic KV Routing: Enabling Adaptive Depth-Wise Cache Sharing

TL;DR

This paper introduces a stochastic training method that enables flexible depth-wise cache sharing in transformer models, reducing memory usage without sacrificing performance.

Contribution

It proposes a simple training approach, random cross-layer attention, to make models robust to depth-wise cache sharing, facilitating efficient KV cache reduction.

Findings

Depth-wise cache sharing can significantly reduce memory footprint.

Random cross-layer attention maintains or improves model performance.

The method is effective during pre-training and fine-tuning across various models.

Abstract

Serving transformer language models with high throughput requires caching Key-Values (KVs) to avoid redundant computation during autoregressive generation. The memory footprint of KV caching is significant and heavily impacts serving costs. This work proposes to lessen these memory requirements. While recent work has largely addressed KV cache reduction via compression and eviction along the temporal axis, we argue that the \emph{depth} dimension offers an orthogonal and robust avenue for optimization. Although prior research suggests that a full cache for every layer is redundant, implementing cross-layer cache sharing remains a practical challenge; existing methods typically suffer from reduced throughput or increased time-to-first-token. In this paper, we demonstrate that dropping a layer's cache offers efficient optimization without information loss. We propose a simple training approach: random cross-layer attention. During training, layers randomly choose to attend either to their own KV states or those of a preceding layer. This stochastic process adapts the model to be robust to various depth-wise cache sharing strategies, ensuring flexibility for unknown hardware constraints at deployment time. Our evaluations show that applying this scheme during pre-training or fine-tuning enables depth-wise cache sharing for various model families. Furthermore, for larger models in data-constrained settings, this approach is suggestive of a regularization-like effect, frequently preserving or improving performance while significantly reducing the cache's memory footprint.