Unlimiformer: Long-Range Transformers with Unlimited Length Input

TL;DR

Unlimiformer enables existing transformer models to process unlimited-length inputs efficiently by using a kNN-based attention approximation, significantly extending their capabilities for long-document tasks without retraining.

Contribution

It introduces a universal method to extend pretrained transformers to handle unlimited input lengths using a kNN index for attention, without additional training or code modifications.

Findings

Successfully processes inputs up to 500k tokens.

Improves performance of BART and Longformer on long-document tasks.

Operates efficiently with sub-linear query time.

Abstract

Since the proposal of transformers, these models have been limited to bounded input lengths, because of their need to attend to every token in the input. In this work, we propose Unlimiformer: a general approach that wraps any existing pretrained encoder-decoder transformer, and offloads the cross-attention computation to a single k-nearest-neighbor (kNN) index, while the returned kNN distances are the attention dot-product scores. This kNN index can be kept on either the GPU or CPU memory and queried in sub-linear time; this way, we can index practically unlimited input sequences, while every attention head in every decoder layer retrieves its top-k keys, instead of attending to every key. We evaluate Unlimiformer on several long-document and book-summarization benchmarks, showing that it can process even 500k token-long inputs from the BookSum dataset, without any input truncation at test time. We demonstrate that Unlimiformer improves pretrained models such as BART and Longformer by extending them to unlimited inputs without additional learned weights and without modifying their code. We make our code and models publicly available at https://github.com/abertsch72/unlimiformer .