Longformer

Longformer is a BERT-like model for long documents.

***** New July 23rd, 2020: Speed degradation *****

A significant speed degradation in the hugginface/transformers was recenlty discovered and fixed (check this PR for details). To avoid this problem, either use the old release v2.11.0 but it doesn't support gradient checkpointing, or use the master branch. This problem should be fixed with the next hugginface/transformers release.

***** New June 29th, 2020: Easier to use Gradient checkpointing *****

Gradient checkpointing has been released with huggingface/transformers release v3.0.0. Gradient checkpointing reduces memory by 5x which makes it possible to process longer sequences on smaller GPUs. To use, try something like the following:

from transformers import LongformerModel
model = LongformerModel.from_pretrained('allenai/longformer-base-4096', gradient_checkpointing=True)

***** New June 2nd, 2020: Integrating with Huggingface + Train your own long model + Gradient checkpointing *****

1. Longformer is now integrated in the huggingface/transformers release v2.11.0. Now you can do

from transformers import LongformerModel
model = LongformerModel.from_pretrained("allenai/longformer-base-4096")

The release also includes LongformerForQA and other LongformerForTaskName with automatic setting of global attention.

2. We added a notebook to show how to convert an existing pretrained model into its "long" version.

3. Gradient checkpointing has been merged into HF master (check PR). Gradient checkpointing can reduce memory usage significanlty (5x for longformer-base-4096) allowing longer sequences on smaller gpus.

***** New April 27th, 2020: A PyTorch implementation of the sliding window attention *****

We added a PyTorch implementation of the sliding window attention that doesn't require the custom CUDA kernel. It is limited in functionality but more convenient to use for finetuning on downstream tasks.

Advantage: supports CPU, TPU and fp16, which aren't supported by the custom CUDA kernel

Limitations: uses 2x more memory (but fp16 offsets that), and doesn’t support dilation and autoregressive attention (not needed for finetuning)

therefore, it is suitable for finetuning on downstream tasks but not a good choice for language modeling. The code snippit below and the TriviaQA scripts were updated to use this new implementation.

***** End new information *****

How to use

1. Download pretrained model

• longformer-base-4096
• longformer-large-4096

2. Install environment and code

   conda create --name longformer python=3.7
   conda activate longformer
   conda install cudatoolkit=10.0
   pip install git+https://github.com/allenai/longformer.git

3. Run the model

   import torch
   from longformer.longformer import Longformer, LongformerConfig
   from longformer.sliding_chunks import pad_to_window_size
   from transformers import RobertaTokenizer
   
   config = LongformerConfig.from_pretrained('longformer-base-4096/') 
   # choose the attention mode 'n2', 'tvm' or 'sliding_chunks'
   # 'n2': for regular n2 attantion
   # 'tvm': a custom CUDA kernel implementation of our sliding window attention
   # 'sliding_chunks': a PyTorch implementation of our sliding window attention
   config.attention_mode = 'sliding_chunks'
   
   model = Longformer.from_pretrained('longformer-base-4096/', config=config)
   tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
   tokenizer.model_max_length = model.config.max_position_embeddings
   
   SAMPLE_TEXT = ' '.join(['Hello world! '] * 1000)  # long input document
   
   input_ids = torch.tensor(tokenizer.encode(SAMPLE_TEXT)).unsqueeze(0)  # batch of size 1
   
   # TVM code doesn't work on CPU. Uncomment this if `config.attention_mode = 'tvm'`
   # model = model.cuda(); input_ids = input_ids.cuda()
   
   # Attention mask values -- 0: no attention, 1: local attention, 2: global attention
   attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=input_ids.device) # initialize to local attention
   attention_mask[:, [1, 4, 21,]] =  2  # Set global attention based on the task. For example,
                                        # classification: the <s> token
                                        # QA: question tokens
   
   # padding seqlen to the nearest multiple of 512. Needed for the 'sliding_chunks' attention
   input_ids, attention_mask = pad_to_window_size(
           input_ids, attention_mask, config.attention_window[0], tokenizer.pad_token_id)
   
   output = model(input_ids, attention_mask=attention_mask)[0]

Model pretraining

This notebook demonstrates our procedure for training Longformer starting from the RoBERTa checkpoint. The same procedure can be followed to get a long-version of other existing pretrained models.

TriviaQA

• Training scripts: scripts/triviaqa.py
• Pretrained large model: here (replicates leaderboard results)
• Instructions: scripts/cheatsheet.txt

CUDA kernel

Our custom CUDA kernel is implemented in TVM. For now, the kernel only works on GPUs and Linux. We tested it on Ubuntu, Python 3.7, CUDA10, PyTorch >= 1.2.0. If it doesn't work for your environment, please create a new issue.

Compiling the kernel: We already include the compiled binaries of the CUDA kernel, so most users won't need to compile it, but if you are intersted, check scripts/cheatsheet.txt for instructions.

Known issues

Please check the repo issues for a list of known issues that we are planning to address soon. If your issue is not discussed, please create a new one.

Citing

If you use Longformer in your research, please cite Longformer: The Long-Document Transformer.

@article{Beltagy2020Longformer,
  title={Longformer: The Long-Document Transformer},
  author={Iz Beltagy and Matthew E. Peters and Arman Cohan},
  journal={arXiv:2004.05150},
  year={2020},
}

Longformer is an open-source project developed by the Allen Institute for Artificial Intelligence (AI2). AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.