BinaryAttention: One-Bit QK-Attention for Vision and Diffusion Transformers

TL;DR

BinaryAttention introduces a 1-bit quantization method for attention in vision and diffusion transformers, significantly reducing computational cost while maintaining or improving accuracy through learnable biases and training techniques.

Contribution

The paper presents a novel 1-bit attention mechanism that preserves essential similarity relationships and accelerates transformer computations for vision tasks.

Findings

BinaryAttention is over 2x faster than FlashAttention2 on A100 GPUs.

It matches or exceeds full-precision attention accuracy in vision and diffusion transformer benchmarks.

The method effectively mitigates information loss through learnable biases and quantization-aware training.

Abstract

Transformers have achieved widespread and remarkable success, while the computational complexity of their attention modules remains a major bottleneck for vision tasks. Existing methods mainly employ 8-bit or 4-bit quantization to balance efficiency and accuracy. In this paper, with theoretical justification, we indicate that binarization of attention preserves the essential similarity relationships, and propose BinaryAttention, an effective method for fast and accurate 1-bit qk-attention. Specifically, we retain only the sign of queries and keys in computing the attention, and replace the floating dot products with bit-wise operations, significantly reducing the computational cost. We mitigate the inherent information loss under 1-bit quantization by incorporating a learnable bias, and enable end-to-end acceleration. To maintain the accuracy of attention, we adopt quantization-aware training and self-distillation techniques, mitigating quantization errors while ensuring sign-aligned similarity. BinaryAttention is more than 2x faster than FlashAttention2 on A100 GPUs. Extensive experiments on vision transformer and diffusion transformer benchmarks demonstrate that BinaryAttention matches or even exceeds full-precision attention, validating its effectiveness. Our work provides a highly efficient and effective alternative to full-precision attention, pushing the frontier of low-bit vision and diffusion transformers. The codes and models can be found at https://github.com/EdwardChasel/BinaryAttention.