Fast Transformers

Transformers are very successful models that achieve state of the art performance in many natural language tasks. However, it is very difficult to scale them to long sequences due to the quadratic scaling of self-attention.

This library was developed for our research on fast attention for transformers. You can find a list of our papers in the docs as well as related papers and papers that we have implemented.

Quick-start

The following code builds a transformer with softmax attention and one with linear attention and compares the time required by each to encode a sequence with 1000 elements.

import torch
from fast_transformers.builders import TransformerEncoderBuilder

# Create the builder for our transformers
builder = TransformerEncoderBuilder.from_kwargs(
    n_layers=8,
    n_heads=8,
    query_dimensions=64,
    value_dimensions=64,
    feed_forward_dimensions=1024
)

# Build a transformer with softmax attention
builder.attention_type = "full"
softmax_model = builder.get()

# Build a transformer with linear attention
builder.attention_type = "linear"
linear_model = builder.get()

# Construct the dummy input
X = torch.rand(10, 1000, 8*64)

# Prepare everythin for CUDA
X = X.cuda()
softmax_model.cuda()
softmax_model.eval()
linear_model.cuda()
linear_model.eval()

# Warmup the GPU
with torch.no_grad():
    softmax_model(X)
    linear_model(X)
torch.cuda.synchronize()

# Measure the execution time
softmax_start = torch.cuda.Event(enable_timing=True)
softmax_end = torch.cuda.Event(enable_timing=True)
linear_start = torch.cuda.Event(enable_timing=True)
linear_end = torch.cuda.Event(enable_timing=True)

with torch.no_grad():
    softmax_start.record()
    y = softmax_model(X)
    softmax_end.record()
    torch.cuda.synchronize()
    print("Softmax: ", softmax_start.elapsed_time(softmax_end), "ms")
    # Softmax: 144 ms (on a GTX1080Ti)

with torch.no_grad():
    linear_start.record()
    y = linear_model(X)
    linear_end.record()
    torch.cuda.synchronize()
    print("Linear: ", linear_start.elapsed_time(linear_end), "ms")
    # Linear: 68 ms (on a GTX1080Ti)

Dependencies & Installation

The fast transformers library has the following dependencies:

• PyTorch
• C++ toolchain
• CUDA toolchain (if you want to compile for GPUs)

For most machines installation should be as simple as:

pip install --user pytorch-fast-transformers

Documentation

There exists a dedicated documentation site but you are also encouraged to read the source code.

Research

Ours

To read about the theory behind some attention implementations in this library we encourage you to follow our research.

• Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention (2006.16236)
• Fast Transformers with Clustered Attention (2007.04825)

If you found our research helpful or influential please consider citing

@inproceedings{katharopoulos_et_al_2020,
    author = {Katharopoulos, A. and Vyas, A. and Pappas, N. and Fleuret, F.},
    title = {Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention},
    booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
    year = {2020}
}

@article{vyas_et_al_2020,
    author={Vyas, A. and Katharopoulos, A. and Fleuret, F.},
    title={Fast Transformers with Clustered Attention},
    journal={arXiv preprint arXiv:2007.04825},
    year={2020}
}

By others

• Efficient Attention: Attention with Linear Complexities (1812.01243)
• Linformer: Self-Attention with Linear Complexity (2006.04768)
• Reformer: The Efficient Transformer (2001.04451)

Support, License and Copyright

This software is distributed with the MIT license which pretty much means that you can use it however you want and for whatever reason you want. All the information regarding support, copyright and the license can be found in the LICENSE file in the repository.