Description

xFormers is a modular and field agnostic library to flexibly generate transformer architectures from interoperable and optimized building blocks. These blocks are not limited to xFormers and can also be cherry picked as the user see fit.

Getting started

The full documentation contains instructions for getting started, deep dives and tutorials about the various APIs. If in doubt, please check out the HOWTO. Only some general considerations are laid out in the README.

For recent changes, you can have a look at the changelog

Installation

To install xFormers, it is recommended to use a dedicated virtual environment, as often with python, through python-virtualenv or conda for instance.

  conda create --name xformer_env
  conda activate xformer_env

Please note that, until Pytorch 1.12 is released, xformers requires Pytorch nightly to be installed. You can fetch it using pip or conda here

There are two ways you can install xFormers locally:

Directly from the pip package

You can fetch the latest release from PyPi. This will not contain the wheels for the sparse attention kernels, for which you will need to build from source.

pip install xformers

Build from source (dev mode)

These commands will fetch the latest version of the code and then install xFormers from source. If you want to build the sparse attention CUDA kernels, please make sure that the next point is covered prior to running these instructions.

git clone [email protected]:facebookresearch/xformers.git
cd xformers
pip install -r requirements.txt
pip install -e .
# or, for OSX
MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ pip install -e .

Installing custom (non-pytorch) parts

Sparse attention kernels

Installing the CUDA-based sparse attention kernels may require extra care, as this mobilizes the CUDA toolchain. As a reminder, these kernels are built when you run pip install -e . and the CUDA buildchain is available (NVCC compiler). Re-building can for instance be done via python3 setup.py clean && python3 setup.py develop, so similarly wipe the build folder and redo a pip install -e.

Some advices related to building these CUDA-specific components, tentatively adressing common pitfalls. Please make sure that:

• NVCC and the current CUDA runtime match. Depending on your setup, you may be able to change the CUDA runtime with module unload cuda module load cuda/xx.x, possibly also nvcc
• the version of GCC that you're using matches the current NVCC capabilities
• the TORCH_CUDA_ARCH_LIST env variable is set to the architures that you want to support. A suggested setup (slow to build but comprehensive) is export TORCH_CUDA_ARCH_LIST="6.0;6.1;6.2;7.0;7.2;8.0;8.6"

Triton

Some parts of xFormers use Triton, and will only expose themselves if Triton is installed, and a compatible GPU is present (nVidia GPU with tensor cores). If Triton was not installed as part of the testing procedure, you can install it directly by running pip install triton. You can optionally test that the installation is successful by running one of the Triton-related benchmarks, for instance python3 xformers/benchmarks/benchmark_triton_softmax.py

Triton will cache the compiled kernels to /tmp/triton by default. If this becomes an issue, this path can be specified through the TRITON_CACHE_DIR environment variable.

Testing the installation

This will run a benchmark of the attention mechanisms exposed by xFormers, and generate a runtime and memory plot. If this concludes without errors, the installation is successful. This step is optional, and you will need some extra dependencies for it to be able to go through : pip install -r requirements-benchmark.txt.

Once this is done, you can run this particular benchmark as follows:

python3 xformers/benchmarks/benchmark_encoder.py --activations relu  --plot -emb 256 -bs 32 -heads 16

Using xFormers

Transformers key concepts

Let's start from a classical overview of the Transformer architecture (illustration from Lin et al,, "A Survey of Transformers")

You'll find the key repository boundaries in this illustration: a Transformer is generally made of a collection of attention mechanisms, embeddings to encode some positional information, feed-forward blocks and a residual path (typically referred to as pre- or post- layer norm). These boundaries do not work for all models, but we found in practice that given some accomodations it could capture most of the state of the art.

Models are thus not implemented in monolithic files, which are typically complicated to handle and modify. Most of the concepts present in the above illustration correspond to an abstraction level, and when variants are present for a given sub-block it should always be possible to select any of them. You can focus on a given encapsulation level and modify it as needed.

Repo map

├── components                  # Parts zoo, any of which can be used directly
│   ├── attention
│   │    └ ...                  # all the supported attentions
│   ├── feedforward             #
│   │    └ ...                  # all the supported feedforwards
│   ├── positional_embedding    #
│   │    └ ...                  # all the supported positional embeddings
│   ├── activations.py          #
│   └── multi_head_dispatch.py  # (optional) multihead wrap
│
├── factory                     # Build model programatically
│   ├── block_factory.py        # (optional) helper to programatically generate layers
│   └── model_factory.py        # (optional) helper to programatically generate models
│
├── benchmarks
│     └ ...                     # A lot of benchmarks that you can use to test some parts
└── triton
      └ ...                     # (optional) all the triton parts, requires triton + CUDA gpu

Attention mechanisms

• Scaled dot product

  • Attention is all you need, Vaswani et al., 2017

• Sparse

  • whenever a sparse enough mask is passed

• BlockSparse

  • courtesy of Triton

• Linformer

  • Linformer, self-attention with linear complexity, Wang et al., 2020

• Nystrom

  • Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention, Xiong et al., 2021

• Local. Notably used in (and many others)

  • Longformer: The Long-Document Transformer, Beltagy et al., 2020
  • BigBird, Transformer for longer sequences, Zaheer et al., 2020

• Favor/Performer

  • Rethinking Attention with Performers, Choromanski et al., 2020

• Orthoformer

  • Keeping Your Eye on the Ball: Trajectory Attention in Video Transformers, Patrick et al., 2021

• Random

  • See BigBird, Longformers,..

• Global

  • See BigBird, Longformers,..

• FourierMix

  • FNet: Mixing Tokens with Fourier Transforms, Lee-Thorp et al.

• CompositionalAttention

  • Compositional Attention: Disentangling search and retrieval, S. Mittal et al.

• 2D Pooling

  • Metaformer is actually what you need for vision, Yu et al.

• Visual Attention

  • Visual Attention Network_, Guo et al

• ... add a new one see Contribution.md

Feed forward mechanisms

• MLP
• Fused
• Mixture of Experts
• Conv2DFeedforward

Positional embedding

• Sine
• Vocabulary
• Rotary
• Simplicial

Residual paths

• Pre
• Post
• DeepNorm

Initializations

This is completely optional, and will only occur when generating full models through xFormers, not when picking parts individually.

There are basically two initialization mechanisms exposed, but the user is free to initialize weights as he/she sees fit after the fact.

• Parts can expose a init_weights() method, which define sane defaults
• xFormers supports specific init schemes which can take precedence over the init_weights()

If the second code path is being used (construct model through the model factory), we check that all the weights have been initialized, and possibly error out if it's not the case (if you set xformers.factory.weight_init.__assert_if_not_initialized = True)

Supported initialization schemes are:

• Small init
• Timm defaults
• ViT defaults
• Moco v3 defaults

One way to specify the init scheme is to set the config.weight_init field to the matching enum value. This could easily be extended, feel free to submit a PR !

Key Features

1. Many attention mechanisms, interchangeables
2. Optimized building blocks, beyond PyTorch primitives
   1. sparse attention
   2. block-sparse attention
   3. fused softmax
   4. fused linear layer
   5. fused layer norm
   6. fused dropout(activation(x+bias))
3. Benchmarking and testing tools
   1. micro benchnmarks
   2. transformer block benchmark
   3. LRA, with SLURM suppot
4. Programatic and sweep friendly layer and model construction
   1. Compatible with hierarchical Transformers, like Swin or Metaformer
5. Hackable
   1. Not using monolithic CUDA kernels, composable building blocks
   2. Using Triton for some optimized parts, explicit, pythonic and user-accessible
   3. Native support for SquaredReLU (on top of ReLU, LeakyReLU, GeLU, ..), extensible activations

FAQ ?

We've tried to collect a relatively exhaustive list of explanations in the HOWTO

License

xFormers has a BSD-style license, as found in the LICENSE file.

Citing xFormers

If you use xFormers in your publication, please cite it by using the following BibTeX entry.

@Misc{xFormers2021,
  author =       {Benjamin Lefaudeux and Francisco Massa and Diana Liskovich and Wenhan Xiong and Vittorio Caggiano and Sean Naren and Min Xu and Jieru Hu and Marta Tintore and Susan Zhang},
  title =        {xFormers: A modular and hackable Transformer modelling library},
  howpublished = {\url{https://github.com/facebookresearch/xformers}},
  year =         {2021}
}

Credits

The following repositories are used in xFormers, either in close to original form or as an inspiration:

• Sputnik
• GE-SpMM
• Triton
• LucidRain Reformer
• RevTorch
• Nystromformer
• FairScale
• Pytorch Image Models