Training-Free Long-Context Scaling of Large Language Models

TL;DR

This paper introduces Dual Chunk Attention, a training-free method enabling large language models like Llama2 70B to handle over 100k token contexts efficiently, matching or surpassing finetuned models on practical tasks.

Contribution

The paper presents Dual Chunk Attention, a novel attention mechanism that allows long-context processing in LLMs without additional training, improving scalability and performance.

Findings

Supports over 100k tokens without training

Achieves comparable or better performance than finetuned models

Reaches 94% of GPT-3.5-16k performance

Abstract

The ability of Large Language Models (LLMs) to process and generate coherent text is markedly weakened when the number of input tokens exceeds their pretraining length. Given the expensive overhead of finetuning large-scale models with longer sequences, we propose Dual Chunk Attention (DCA), which enables Llama2 70B to support context windows of more than 100k tokens without continual training. By decomposing the attention computation for long sequences into chunk-based modules, DCA manages to effectively capture the relative positional information of tokens within the same chunk (Intra-Chunk) and across distinct chunks (Inter-Chunk), as well as integrates seamlessly with Flash Attention. In addition to its impressive extrapolation capability, DCA achieves performance on practical long-context tasks that is comparable to or even better than that of finetuned models. When compared with proprietary models, our training-free 70B model attains 94% of the performance of gpt-3.5-16k, indicating it is a viable open-source alternative. All code and data used in this work are released at \url{https://github.com/HKUNLP/ChunkLlama}.