Unlimiformer: Long-Range Transformers with Unlimited Length Input (NeurIPS 2023)

This is the official implementation of the paper:

Amanda Bertsch, Uri Alon, Graham Neubig, and Matthew R. Gormley:
Unlimiformer: Long-Range Transformers with Unlimited Length Input (to appear in NeurIPS 2023)

Unlimiformer is a method for augmenting pretrained encoder-decoder models with retrieval-based attention, without changing the mathematical definition of attention. This allows the use of unlimited length inputs with any pretrained encoder-decoder!
See also our Tweet.

Unlimiformer can be used to improve the performance of an already-trained model. For best results, the model can be trained with Unlimiformer training.

If you have any questions on this work, please open a GitHub issue or email the authors at abertsch@cs.cmu.edu, ualon@cs.cmu.edu

October 2023 - Unlimiformer will appear at NeurIPS 2023!

August 2023 - Unlimiformer now supports Llama-2 (and all its derivatives)!

To prompt Llama-2 with extremely long inputs, for example, the content of an entire book, use:

python src/run_generation.py --model_type llama --model_name_or_path meta-llama/Llama-2-13b-chat-hf \
    --prefix "<s>[INST] <<SYS>>\n You are a helpful assistant. Answer with detailed responses according to the entire instruction or question. \n<</SYS>>\n\n Summarize the following book: " \
    --prompt example_inputs/harry_potter_full.txt \
    --suffix " [/INST]" --test_unlimiformer --fp16 --length 200 --layer_begin 16 \
    --index_devices 1 --datastore_device 1

The final prompt will be a concatenation of the content of the flags: --prefix, --prompt, --suffix.
The flag --prompt may contain either a path to a text file (e.g., example_inputs/harry_potter_full.txt) or the concrete prompt string.
The flag --test_unlimiformer is required to enable Unlimiformer.
The flag --length determines the desired output length.
The flag --layer_begin determines the layer from which Unlimiformer will start to be applied. For example, if we set --layer_begin 20, the first 20 layers of the model will perform the standard attention over the last context_window_size tokens of the prompt as usual, and the 21st layer and above will attend to the entire long input. From our initial experiments, the value of --layer_begin should be more than half of the total number of layers in the model, and tuning it dramatically changes the quality of the output.
The flags: --datastore_device N and --index_devices N1 N2 N3 ... specify on which GPUs to store Unlimiformer's datastore and index (the base model will be stored on GPU #0).
Add the flag --stream_output to make the generated tokens appear one by one as they are generated.

Getting Started

General Instructions

Copy the files from src into your source code folder.

You'll need to set values for the Unlimiformer-specific arguments outlined in usage.py - you can add these arguments wherever you usually process hyperparameters. To use the model, you must set test_unlimiformer=True. For datastore usage, the model must be in evaluation model (e.g. call model.eval() before inference).

inference-example.py outlines a minimal example for running a sequence through an Unlimiformer model, using the default arguments.

run.py is an example of a full training setup that integrates Unlimiformer, adopted from SLED. See full command lines below.

Reproducing the Experiments from the Paper - Command Lines

To run a standard finetuning + evaluation of BART-base on the GovReport dataset (as examples), use:

python src/run.py \
    src/configs/training/base_training_args.json \
    src/configs/data/gov_report.json \
    --output_dir output_train_bart_base_local/ \
    --learning_rate 1e-5 \
    --model_name_or_path facebook/bart-base \
    --max_source_length 1024 \
    --eval_max_source_length 1024 --do_eval=True \
    --eval_steps 1000 --save_steps 1000 \
    --per_device_eval_batch_size 1 --per_device_train_batch_size 2 \
    --extra_metrics bertscore

To use Unlimiformer at training time (called "Retrieval training" in the paper), use: --unlimiformer_training --max_source_length 16384
- In this case, you might want to use Unlimiformer also at test/validation time, and use also: --test_unlimiformer --eval_max_source_length 999999
Alternatively, to use the computationally cheaper "Random-encoded" at training time, use --random_unlimiformer_training --max_source_length 16384
To alternate between "retrieval training" and "random-encoded training", use both flags: --unlimiformer_training --random_unlimiformer_training --max_source_length 16384

For additional flags and options, see usage.py

Recommended settings

To evaluate with Unlimiformer

At evaluation time, we recommend the default value for each setting.

To train with Unlimiformer

For an inexpensive method, we recommend training as usual and using Unlimiformer during early stopping. To do so, set knn=True and leave all other values at default.

For best performance, there are 3 expensive settings for training. The best one varies by dataset.

Set random_unlimiformer_training=True: this is the random-encoded training setting from the paper
Set unlimiformer_training=True: this is the retrieval training setting from the paper
Set random_unlimiformer_training=True AND unlimiformer_training=True: this is the alternating training setting from the paper

See Table 5 in the paper for a more detailed breakdown of relative training costs.

Tips for very large inputs

For training

you may need to truncate your inputs at training time, e.g. to 8k or 16k tokens. You can use the full inputs at evaluation time
you can also try splitting your inputs into 16k-token-chunks and training on each one as its own example

For evaluation (including early stopping)

if you're consistently running out of CUDA memory, set use_datastore=True to use a Faiss datastore to store hidden states.
if you're still having issues, set gpu_datastore=False or gpu_index=False, but note that this will degrade performance

Trained models

The following models from the paper are available on Hugging Face. Please note that you must add the Unlimiformer-specific files to your repository, and load these models with test_unlimiformer=True. If you download these models from Hugging Face, they may not use Unlimiformer by default!

Table 3: low-cost training methods

Dataset	Method	Hugging Face link
GovReport	Baseline: BART-base	abertsch/bart-base-govreport
GovReport	BART-base + Unlimiformer early stopping	abertsch/unlimiformer-bart-govreport-earlyk
SummScreen	Baseline: BART-base	abertsch/bart-base-summscreen
SummScreen	BART-base + Unlimiformer early stopping	abertsch/unlimiformer-bart-summscreen-earlyk

Table 4: Long-range training methods

Dataset	Method	Hugging Face link
GovReport	BART + Unlimiformer (alternating training)	abertsch/unlimiformer-bart-govreport-alternating
SummScreen	BART + Unlimiformer (retrieval training)	abertsch/unlimiformer-bart-summscreen-retrieval

Table 5: BookSum

Dataset	Method	Hugging Face link
BookSum	Baseline: BART-base	abertsch/bart-base-booksum
BookSum	BART-base + Unlimiformer early stopping	abertsch/unlimiformer-bart-booksum-earlyk
Booksum	BART-base + Unlimiformer (random-encoding training)	abertsch/unlimiformer-bart-booksum-random-encoding
Booksum	BART-base + Unlimiformer (alternating training)	abertsch/unlimiformer-bart-booksum-alternating

Results

Citation

If you use our method or models, please cite our paper:

@article{bertsch2023unlimiformer,
  title={Unlimiformer: Long-Range Transformers with Unlimited Length Input},
  author={Bertsch, Amanda and Alon, Uri and Neubig, Graham and Gormley, Matthew R},
  journal={arXiv preprint arXiv:2305.01625},
  year={2023}
}

abertsch72 / unlimiformer