You Only Cache Once: Decoder-Decoder Architectures for Language Models

TL;DR

YOCO introduces a decoder-decoder architecture for large language models that caches key-value pairs only once, reducing memory and speeding up inference while maintaining global attention capabilities.

Contribution

The paper presents a novel decoder-decoder architecture, YOCO, which efficiently caches key-value pairs once, enabling large-scale language models to operate with reduced memory and faster inference.

Findings

YOCO reduces GPU memory usage significantly.

YOCO accelerates prefill latency and increases throughput.

YOCO achieves comparable performance to traditional Transformers across various scales.

Abstract

We introduce a decoder-decoder architecture, YOCO, for large language models, which only caches key-value pairs once. It consists of two components, i.e., a cross-decoder stacked upon a self-decoder. The self-decoder efficiently encodes global key-value (KV) caches that are reused by the cross-decoder via cross-attention. The overall model behaves like a decoder-only Transformer, although YOCO only caches once. The design substantially reduces GPU memory demands, yet retains global attention capability. Additionally, the computation flow enables prefilling to early exit without changing the final output, thereby significantly speeding up the prefill stage. Experimental results demonstrate that YOCO achieves favorable performance compared to Transformer in various settings of scaling up model size and number of training tokens. We also extend YOCO to 1M context length with near-perfect needle retrieval accuracy. The profiling results show that YOCO improves inference memory, prefill latency, and throughput by orders of magnitude across context lengths and model sizes. Code is available at https://aka.ms/YOCO.