Support for Cuda 12 and PyTorch 2.1.0

| |

Neighborhood Attention Extension

Bringing attention to a neighborhood near you!

Visualization of neighborhood attention in 2D.

Visualization of dilated neighborhood attention in 2D.

NATTEN is an open-source project dedicated to providing fast implementations for Neighborhood Attention, a sliding window self-attention mechanism.

If you're not familiar with neighborhood attention, please refer to our papers, or watch our YouTube video from CVPR 2023.

To read more about our GEMM-based and fused neighborhood attention kernels, please refer to our new preprint, Faster Neighborhood Attention.

New: Fused Neighborhood Attention now supports backpropagation!

We've released the Fused Neighborhood Attention (FNA) backward kernel and interface, which means you can now train models based on neighborhood attention faster and more efficiently.

FNA can be seen as a generalization of methods such as Flash Attention and FMHA from back-to-back matrix multiplication to back-to-back tensor-tensor contraction, and comes with neighborhood attention masking built in. This accelerates neighborhood attention, a multi-dimensional sliding window attention pattern, by never storing the attention tensor to global memory, which aside from reducing global memory footprint also reduces the memory bandwidth bottleneck.

We highly recommend referring to FNA quick start or the Fused vs unfused NA guide before starting to use FNA, since the interface, memory layout, and feature set can differ from all unfused ops in NATTEN.

Getting started

NATTEN supports PyTorch version 2.0 and later, and Python versions 3.8 and above. Python 3.12 is only supported with torch >= 2.2.0.

Older NATTEN releases supported python >= 3.7 and torch >= 1.8.

Please refer to install instructions to find out whether your operating system and hardware accelerator is compatible with NATTEN.

Feature availability

Problem space	CPU backend	CUDA backend
1D	naive	naive, gemm, fna
2D	naive	naive, gemm, fna
3D	naive	naive, fna

CPU

Problem space	CPU Backend	Causal masking	Varying parameters	Relative positional bias	Autograd support
1D	naive	✅	✅	✅	Forward and reverse mode
2D	naive	✅	✅	✅	Forward and reverse mode
3D	naive	✅	✅	✅	Forward and reverse mode

Notes:

Forward mode autograd does not support relative positional biases and causal masking yet.
Relative positional biases are not yet supported when any axis has causal masking enabled.

CUDA

Problem space	CUDA Backend	Causal masking	Varying parameters	Relative positional bias	Autograd support	Min. Arch
1D	naive	✅	✅	✅	Forward and reverse mode	SM35
2D	naive	✅	✅	✅	Forward and reverse mode	SM35
3D	naive	✅	✅	✅	Forward and reverse mode	SM35
1D	gemm	-	-	✅	Forward and reverse mode	SM70
2D	gemm	-	-	✅	Forward and reverse mode	SM70
1D	fna	✅	✅	✅	Reverse mode	SM50
2D	fna	✅	✅	✅	Reverse mode	SM50
3D	fna	✅	✅	✅	Reverse mode	SM50

Notes:

FP16 kernels are only available on SM50 and above*, and BF16 requires SM80 and above.
- Naive FP16 kernels are only available on SM60 and above.
- FNA FP16 kernels are only available on SM50 and above.
GEMM backend on SM70 and SM75 can only do FP16.
Tiled only implements 1/3 of the ops, is only implemented for 2D problems, and requires head dim = 32.
Forward mode autograd does not support relative positional biases and causal masking yet.
Relative positional biases are not yet supported when any axis has causal masking enabled.
Relative positional biases are not supported in FNA during backward pass.

Features that will likely no longer be worked on or improved:

Relative positional biases
- There's just better alternatives that don't involve explicitly biasing the attention weight matrix, and they will be more performant on top of providing similar or better accuracy levels.
GEMM-based kernels
- Since FNA covers more features than our unfused GEMM-based kernels, and we know it to be a better solution (please refer to Faster Neighborhood Attention for details), we do not plan to extend or improve these kernels.
- This includes support for varying parameters, causal masking, and 3-D problems.

License

NATTEN is released under the MIT License.

Citation

@misc{hassani2024faster,
  title        = {Faster Neighborhood Attention: Reducing the O(n^2) Cost of Self Attention at the Threadblock Level},
  author       = {Ali Hassani and Wen-Mei Hwu and Humphrey Shi},
  year         = 2024,
  url          = {https://arxiv.org/abs/2403.04690},
  eprint       = {2403.04690},
  archiveprefix = {arXiv},
  primaryclass = {cs.CV}
}
@inproceedings{hassani2023neighborhood,
  title        = {Neighborhood Attention Transformer},
  author       = {Ali Hassani and Steven Walton and Jiachen Li and Shen Li and Humphrey Shi},
  year         = 2023,
  booktitle    = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}
}
@misc{hassani2022dilated,
  title        = {Dilated Neighborhood Attention Transformer},
  author       = {Ali Hassani and Humphrey Shi},
  year         = 2022,
  url          = {https://arxiv.org/abs/2209.15001},
  eprint       = {2209.15001},
  archiveprefix = {arXiv},
  primaryclass = {cs.CV}
}

Acknowledgements

We thank NVIDIA, and the CUTLASS project and team for their efforts in creating and open-sourcing CUTLASS. We would also like to thank Haicheng Wu for his valuable feedback and comments which led to the creation of GEMM-based NA. We also thank Meta and the xFormers team for their FMHA kernel, which is what our Fused Neighborhood Attention kernel is based on. We thank the PyTorch project and team.

shi-labs / natten Goto Github PK

natten's Introduction

New: Fused Neighborhood Attention now supports backpropagation!

Getting started

Feature availability

CPU

CUDA

License

Citation

Acknowledgements

natten's People

Contributors

Stargazers

Watchers

Forkers

natten's Issues

Compiled version problem

Build from source problem

encoding: utf-8

module natten._C

from /mnt/c/Users/zzz/AppData/Local/JetBrains/PyCharm2021.3/remote_sources/380714796/-89893896/natten/_C.cpython-38-x86_64-linux-gnu.so

by generator 1.147

no doc

no imports

Whether NATTEN was built with CUDA

Whether NATTEN with CUDA was built with support for float16

Whether NATTEN with CUDA was built with support for bfloat16

Whether NATTEN with CUDA was built with the new GEMM kernels

Recommend Projects

Recommend Topics

Recommend Org