Neighborhood Attention Extension
Bringing attention to a neighborhood near you!
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, we recommend referring to our papers, or watching our YouTube video from CVPR 2023.
NATTEN is primarily a C++/CUDA library, which has so far only supported binding with the torch API, and therefore is mostly usable through PyTorch. We plan to eliminate the torch dependency in the future and possibly support other frameworks / engines.
NATTEN's python interface provides Neighborhood Attention (local attention) and Dilated Neighborhood Attention (sparse global attention, a.k.a. dilated local attention) as autograd-compatible PyTorch modules for both 1D, 2D, and 3D data.
It also has experimental support for forward mode automatic differentiation, and nested tensors.
Our CPU implementations are very limited and barely performance-optimized. While we aim to provide the best implementation for different devices, optimizing our CUDA kernels is higher up on the list of priorities. Contributions are always welcomed.
NATTEN generally supports all architectures supported by PyTorch. More specifically, architectures since Kepler (SM35) are supported. However, our most-performant kernels only support architectures since Volta, targeting tensor core math.
Devices with compute capability greater than or equal to 6.0 (Pascal and later) allow running in FP16.
Devices with compute capability greater than or equal to 8.0 (Ampere and later) allow running in BF16.
NATTEN provides more than one set of kernel. Our naive kernels, which were developed during the first phase of the project, provide a very basic implementation of neighborhood attention, and are the last resort for every problem. This means that if your device and software support alternatives to our naive kernels (i.e. GEMM kernels), and your problem size is supported, NATTEN will automatically pick the better kernels for you. (NOTE: this is done based on the class of kernels and not by actual performance via profiling.)
Naive kernels are always usable across different architectures, though not the most performant.
Naive kernels for the 2-dimensional neighborhood attention also come with a tiled implementation for one of the three underlying operations, which is considerably more performant than the original. However, the tiled kernels only support problem sizes with head dim 32, and up to kernel size 13x13. Tiled kernels are also not supported in devices with compute capability smaller than 6.0.
Our GEMM-based kernels depend on and are modeled after CUTLASS's Implicit GEMM kernels for convolution.
Devices with compute capability greater than or equal to 7.0 (Volta, Turing, Ampere, Ada Lovelace, Hopper), can run our GEMM kernels, which are somewhat performance-optimized, thanks to the underlying mainloop from CUTLASS, and target Tensor Core math.
However, do note that their current float16/bfloat16 implementations do not typically result in improved latency, due to a memory alignment issue, which we aim to resolve in future kernels.
Devices with compute capability greater than or equal to 8.0 (Ampere and later) support GEMM kernels with double, full, and half precision (FP64, FP32, FP16, BF16).
Devices with compute capability 7.0 or 7.5 (Volta and Turing) only support GEMM kernels with half precision (FP16). This is because their tensor cores only allow FP16 math.
NOTE: the table presents the average improvement in latency over different problem sizes with full precision (tfloat32).
Simple, just Google your GPU model, and check its compute capability. If you've already set up NATTEN, you could also run:
from natten.functional import get_device_cc
cc = get_device_cc()
cc = get_device_cc(0) # Optionally select a specific GPU
print(f"Your device is SM{cc}.")The new NATTEN library sets up constants that are binded to the python interface, which will allow you to check whether you've compiled with: a. CUDA, b. Float16 (half) support, c. Bfloat16 support, d. New GEMM kernels.
import natten
# Whether NATTEN was built with CUDA kernels,
# and supports running them on this system.
print(natten.has_cuda())
# Whether NATTEN supports running float16 on
# the selected device.
print(natten.has_half())
print(natten.has_half(0)) # Optionally specify a GPU index.
# Whether NATTEN supports running bfloat16 on
# the selected device.
print(natten.has_bfloat())
print(natten.has_bfloat(0)) # Optionally specify a GPU index.
# Whether NATTEN supports running GEMM kernels
# on the selected device.
print(natten.has_gemm())
print(natten.has_gemm(0)) # Optionally specify a GPU index.
# Whether NATTEN supports running GEMM kernels
# in full precision on the selected device.
print(natten.has_fp32_gemm())
print(natten.has_fp32_gemm(0)) # Optionally specify a GPU index.If natten.has_gemm() returns true, by default NATTEN will call the faster GEMM kernels instead of the original naive kernels
for both NA1D and NA2D. 3D Neighborhood attention is not supported at this time, but you can still use the naive kernels.
In addition, we will be adding scripts that allow you to profile and observe latency from the kernels with those options available.
Sliding window self attention mechanisms have been relatively overlooked, in part due to implementation difficulties. For example, in a paper proposing one of the earliest examples of such methods, SASA, it was noted that although such methods are theoretically efficient, they're relatively slow in practice, compared to convolutions, which have been implemented in most well-known deep learning libraries.
That is why we started developing NATTEN, an extension to existing libraries with efficient implementations of sliding window attention mechanisms, which will enable research in this direction including building powerful hierarchical vision transformers.
For more information, we highly recommend reading our preprints NAT and DiNAT, and check out their repository.
- python >= 3.8
- torch >= 2.0
NATTEN supports PyTorch version 2.0 and later, and Python versions 3.7, 3.8, 3.9, 3.10(only torch >= 1.11), and 3.11 (only torch >= 1.13).
Older NATTEN releases supported python >= 3.7 and torch >= 1.8.
NOTE: NATTEN only comes with pre-built Linux wheels, and supports Kepler and above (SM >= 35).
Make sure your GPU is supported by referring to
this webpage.
Future versions will extend support to older GPUs.
Just refer to our website, shi-labs.com/natten, select your PyTorch version and the CUDA version it was compiled with, copy-paste the command and install in seconds!
For example, if you're on torch==2.0.0+cu118, you should install NATTEN using the following wheel:
pip3 install natten -f https://shi-labs.com/natten/wheels/cu118/torch2.0.0/index.htmlMore generally:
pip3 install natten -f https://shi-labs.com/natten/wheels/{cu_version}/torch{torch_version}/index.htmlNOTE: If you do not specify a wheel URL, pip will collect NATTEN and try to compile on locally, which depending on your system might take up to 30 minutes. We strongly recommend using our website if you're a Linux user.
Unfortunately we are not yet able to build Mac wheels (and do not yet have a Metal backend). However, you can compile upon installing and use the CPU kernels:
pip3 install nattenThe current release has not been successfully built Windows devices with CUDA, and therefore does not yet have Windows wheels. If you are a windows user willing to help us figure out building with MSVC, please contact us or open an issue.
Once you've set up your Python environment and installed PyTorch with CUDA, simply clone and build:
git clone https://github.com/SHI-Labs/NATTEN
cd NATTEN
pip install -r requirements.txt
makeNOTE: NATTEN will use the PyTorch API to detect your GPU architecture, and will by default attempt to use 1/4th of the number of processes your system allows to build. You can override them by passing in the following arguments:
# Build with 2 workers/processes
make WORKERS=2
# Build targeting SM89 (Ada Lovelace)
make CUDA_ARCH="8.9"Please also note that building with the latest GEMM kernels can be a bit time consuming, which means at least 10 - 20 minutes given that you use enough workers. It is technically possible to improve build time by generating more source files and using more workers (at the expense of generating a larger binary), but that option will be made available in the future.
You can optionally run unit tests to verify building from source finished successfully:
make test- Neighborhood Attention 1D (CPU, naive)
- Neighborhood Attention 2D (CPU, naive)
- Neighborhood Attention 3D (CPU, naive)
- Neighborhood Attention 1D (CUDA, naive)
- Neighborhood Attention 2D (CUDA, naive)
- Neighborhood Attention 3D (CUDA, naive)
- Neighborhood Attention 1D (CUDA, gemm-based, SM70 and above)
- Neighborhood Attention 2D (CUDA, gemm-based, SM70 and above)
- Dilation support
- Float16 support
- BFloat16 support
- Kepler and Maxwell (30<=SM<60) support
- Windows builds
- Neighborhood Attention 1D (CUDA, fused kernels)
- Neighborhood Attention 2D (CUDA, fused kernels)
Simply import NeighborhoodAttention1D, NeighborhoodAttention2D, or NeighborhoodAttention3D from natten:
from natten import NeighborhoodAttention1D
from natten import NeighborhoodAttention2D
from natten import NeighborhoodAttention3D
na1d = NeighborhoodAttention1D(dim=128, kernel_size=7, dilation=2, num_heads=4)
na2d = NeighborhoodAttention2D(dim=128, kernel_size=7, dilation=2, num_heads=4)
na3d = NeighborhoodAttention3D(dim=128, kernel_size=7, dilation=2, num_heads=4)NA3D also supports different kernel size and dilation values for depth:
na3d = NeighborhoodAttention3D(
dim=128,
kernel_size=7,
kernel_size_d=5,
dilation=2,
dilation_d=3,
num_heads=4)Modules expect inputs of shape [batch_size, *, dim]:
- NA1D:
[batch_size, sequence_length, dim] - NA2D:
[batch_size, height, width, dim] - NA3D:
[batch_size, depth, height, width, dim]
We recommend counting flops through fvcore.
pip install fvcoreOnce you have fvcore installed, you can directly use our dedicated FLOP counter:
from natten.flops import get_flops
flops = get_flops(model, input)Alternatively, if you are using fvcore's FlopCountAnalysis directly, be sure to add our op handles:
from fvcore.nn import FlopCountAnalysis
from natten.flops import add_natten_handle
# ...
flop_ctr = FlopCountAnalysis(model, input)
flop_ctr = add_natten_handle(flop_ctr)
# ...NATTEN is released under the MIT License.
@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)}
}
@article{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}
}We would like to 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 Implicit GEMM NA. We also thank Meta, and the PyTorch project and team.