native-sparse-attention-pytorch

Native Sparse Attention

Implementation of the sparse attention pattern proposed by the Deepseek team in their Native Sparse Attention paper

This will be my last open sourced project under Meta

Appreciation

Phil Tillet for democratizing CUDA kernel hacking with Triton
Flex Attention for allowing for rapid prototyping
@Mr-Grin for the code review and pointing out an inaccuracy with the implementation
Eric Pasewark for submitting a simple transformer based compression network
@Mr-Grin for a pull request that fixes compression block hyperparameters
@StrongSpoon for a memory access guard

Install

$ pip install native-sparse-attention-pytorch

Usage

import torch
from native_sparse_attention_pytorch import SparseAttention

attn = SparseAttention(
    dim = 512,
    dim_head = 64,
    heads = 8,
    sliding_window_size = 2,
    compress_block_size = 4,
    compress_block_sliding_stride = 2,
    selection_block_size = 4,
    num_selected_blocks = 2
)

tokens = torch.randn(2, 31, 512)

attended = attn(tokens)

assert tokens.shape == attended.shape

Example

Enwik8 language modeling

$ pip install .[examples]

Then

$ python train.py

To record some of your experiments, just invoke wandb login first before modifying the training script

Citations

@inproceedings{Yuan2025NativeSA,
    title   = {Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention},
    author  = {Jingyang Yuan and Huazuo Gao and Damai Dai and Junyu Luo and Liang Zhao and Zhengyan Zhang and Zhenda Xie and Y. X. Wei and Lean Wang and Zhiping Xiao and Yuqing Wang and Chong Ruan and Ming Zhang and Wenfeng Liang and Wangding Zeng},
    year    = {2025},
    url     = {https://api.semanticscholar.org/CorpusID:276408911}
}

@inproceedings{Keles2022OnTC,
    title   = {On The Computational Complexity of Self-Attention},
    author  = {Feyza Duman Keles and Pruthuvi Maheshakya Wijewardena and Chinmay Hegde},
    booktitle = {International Conference on Algorithmic Learning Theory},
    year    = {2022},
    url     = {https://api.semanticscholar.org/CorpusID:252198880}
}

Metadata

Release files for native-sparse-attention-pytorch 0.2.3

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for native-sparse-attention-pytorch 0.2.3
File	Size	Uploaded
native_sparse_attention_pytorch-0.2.3.tar.gz	36.8 MB	Aug 15, 2025	Details

Built distribution (wheel)

Table of built distributions (wheels) for native-sparse-attention-pytorch 0.2.3
File	Interpreter	ABI	Platform	Reset
native_sparse_attention_pytorch-0.2.3-py3-none-any.whl 27.2 kB Aug 15, 2025	Python 3	none	any	Details

Total release size: 36.9 MB

Release files / native_sparse_attention_pytorch-0.2.3.tar.gz

Download URL	native_sparse_attention_pytorch-0.2.3.tar.gz
Size	36.8 MB
Tags	Source
SHA-256 checksum How to use checksums	`7b9481bb52ee565d5098868c7345285e88342248915ff8e3e62f7fe945be2da0`
BLAKE2b-256 checksum How to use checksums	`faee1c0f8875883103492e6d22a80507c2e5a768ccfcb97c28c9cc9d9ebc7d2c`
Upload date	Aug 15, 2025
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.1.0 CPython/3.9.23`

Release files / native_sparse_attention_pytorch-0.2.3-py3-none-any.whl

Download URL	native_sparse_attention_pytorch-0.2.3-py3-none-any.whl
Size	27.2 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`3a9ac0ebb423594aef3b2a2557f22ce55588395b6b41f53b64c198e2f2fee341`
BLAKE2b-256 checksum How to use checksums	`f0a11f6ed96f00cd306cf3e364b7f593a4d47fc5f5d2feef4d50e5b38c445dd3`
Upload date	Aug 15, 2025
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.1.0 CPython/3.9.23`