Training-Free Long-Context Scaling of Large Language Models
This addresses the challenge of expensive long-context scaling for AI researchers and practitioners, offering a training-free solution that is incremental but practical.
The paper tackles the problem of large language models struggling with long input sequences by proposing Dual Chunk Attention, which enables a 70B model to handle over 100k tokens without training, achieving performance comparable to finetuned models and 94% of a proprietary model's capability.
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}.