LongNet: Scaling Transformers to 1,000,000,000 Tokens

TL;DR

LongNet introduces a scalable Transformer variant capable of handling over 1 billion tokens with linear complexity, enabling efficient long-sequence modeling without performance loss on shorter sequences.

Contribution

It proposes dilated attention for Transformers, allowing exponential expansion of the attentive field with linear complexity, and demonstrates its effectiveness on very long sequences.

Findings

Achieves over 1 billion token sequence modeling.

Maintains strong performance on standard language tasks.

Offers a drop-in replacement for standard attention in Transformers.

Abstract

Scaling sequence length has become a critical demand in the era of large language models. However, existing methods struggle with either computational complexity or model expressivity, rendering the maximum sequence length restricted. To address this issue, we introduce LongNet, a Transformer variant that can scale sequence length to more than 1 billion tokens, without sacrificing the performance on shorter sequences. Specifically, we propose dilated attention, which expands the attentive field exponentially as the distance grows. LongNet has significant advantages: 1) it has a linear computation complexity and a logarithm dependency between any two tokens in a sequence; 2) it can be served as a distributed trainer for extremely long sequences; 3) its dilated attention is a drop-in replacement for standard attention, which can be seamlessly integrated with the existing Transformer-based optimization. Experiments results demonstrate that LongNet yields strong performance on both long-sequence modeling and general language tasks. Our work opens up new possibilities for modeling very long sequences, e.g., treating a whole corpus or even the entire Internet as a sequence.