KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #codex #typescript #developer-tools #ai-agents #deepseek-harness #claude-code #ai-agent #cli #react #mcp

FlashBias

★ 33 Open GitHub ↗

About Code release for "FlashBias: Fast Computation of Attention with Bias" (NeurIPS 2025), https://arxiv.org/abs/2505.12044

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# FlashBias (NeurIPS 2025)
FlashBias: Fast Computation of Attention with Bias [[Paper]](https://arxiv.org/pdf/2505.12044) [[Slides]](https://wuhaixu2016.github.io/pdf/NeurIPS2025_FlashBias.pdf) [[Poster]](https://wuhaixu2016.github.io/pdf/poster_NeurIPS2025_FlashBias.pdf)

Attention with bias widely exists, such as relative position encoding in [Swin Transformer](https://github.com/microsoft/Swin-Transformer), scientific prior encoding in [AlphaFold (Nature 2024)](https://www.nature.com/articles/s41586-024-07487-w) and [Pangu-Weather (Nature 2023)](https://www.nature.com/articles/s41586-023-06185-3). Surprisingly, despite the common use of attention with bias, its targeted efficiency optimization remains absent.

This paper presents [FlashBias](https://arxiv.org/pdf/2505.12044) based on the low-rank compressed sensing theory, which can provide fast-exact computation for many widely used attention biases, enabling 1.5× speedup for Pairformer in AlphaFold 3, and over 2× speedup for attention with bias in vision and language models without loss of accuracy.

<img src=".\pic\flashbias.png" height = "250" alt="" align=center />
 
Figure 1. Overview of FlashBias.

## Get Started

**Check [./flash_bias](https://github.com/thuml/FlashBias/tree/main/flash_bias) for Triton kernel and comparison among FlashBias, FlashAttention, torch-complie, SDPA and xFormers.**

The following are some representative applications of FlashBias.

1. **Large language model** with ALiBi bias: See [./1_Language_Model](https://github.com/thuml/FlashBias/tree/main/1_Language_Model)
2. **Swin Transformer V2** with relative position bias: See [./2_Vision_Transformer](https://github.com/thuml/FlashBias/tree/main/2_Vision_Transformer)
3. **Transformer PDE Solver** with spatial distance bias: See [./3_Neual_Solver](https://github.com/thuml/FlashBias/tree/main/3_Neural_Solver)
4. **AlphaFold 3** with pair representation bias: [./4_AlphaFold3](https://github.com/thuml/FlashBias/tree/main/4_AlphaFold3)

<img src=".\pic\features.png" height = "200" alt="" align=center />
 
Figure 2. Three types of implementations for FlashBias and corresponding representative applications.

## Usage

```python
## Triton-based FlashBias, See `flash_bias_triton.py` for more details.
from flash_bias_triton import flash_bias_func
output_flash = flash_bias_func(q, k, v, q_bias, k_bias, None, False, 1 / np.sqrt(headdim))

## PyTorch-SDPA-based FlashBias
import torch
# Notably, the dimension of concat[q, q_bias] should be divided evenly by 8; otherwise, you cannot activate flashattention in the backend
output_flash_sdpa = torch.nn.functional.scaled_dot_product_attention(
    query=torch.concat([q * softmax_scale, q_bias], dim=-1),
    key=torch.concat([k, k_bias], dim=-1),
    value=v,
    attn_mask=None,
    dropout_p=0.0,
    scale=1,
    is_causal=causal,
)
```

## Overall Comparison

A significant memory and running time reduction compared to vanilla FlashAttention.

<img src=".\pic\efficiency.png" height = "220" alt="" align=center />
 
Figure 3. Efficiency comparison with vanilla FlashAttention.

## Citation

If you find this repo useful, please cite our paper.

```
@inproceedings{wu2025flashbias,
  title={FlashBias: Fast Computation of Attention with Bias},
  author={Haixu Wu and Minghao Guo and Yuezhou Ma and Yuanxu Sun and Jianmin Wang and Wojciech Matusik and Mingsheng Long},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025}
}
```

## Contact

If you have any questions or want to use the code, please contact wuhaixu98@gmail.com

## Acknowledgement

We appreciate the following GitHub repos a lot for their valuable code base or datasets:

https://github.com/Dao-AILab/flash-attention

https://github.com/meta-pytorch/attention-gym

https://github.com/bytedance/Protenix

https://github.com/microsoft/Swin-Transformer

FlashBias

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README