SageAttention

This repository provides the official implementation of SageAttention.

SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration
Paper: https://arxiv.org/abs/2410.02367
Jintao Zhang, Jia Wei, Haofeng Huang, Pengle Zhang, Jun Zhu, Jianfei Chen

SageAttention2 Technical Report: Accurate 4-Bit Attention for Plug-and-play Inference Acceleration
Paper: https://arxiv.org/abs/2411.10958
Jintao Zhang, Haofeng Huang, Pengle Zhang, Jia Wei, Jun Zhu, Jianfei Chen

SageAttention

SageAttention2

Project Updates

• News 2024-11-19: SageAttention2 will be released soon.
• News 2024-11-12: Support for sageattn_varlen is available now.
• News 2024-11-11: Support for different sequence length between q and k,v, (batch_size, head_num, seq_len, head_dim) or (batch_size, seq_len, head_num, head_dim) input shapes, and group-query attention is available now.

Base environment

python>=3.9
torch>=2.3.0
triton>=2.3.0

We recommend to install: (the kernel will be faster a little)
python>=3.11
torch>=2.4.0
triton-nightly

Installation

Install using pip:

pip install sageattention

Or compiling from source:

cd sageattention 
pip install .

Note: SageAttention is currently optimized for RTX4090 and RTX3090 GPUs. Performance improvements may not be significant on other GPU architectures. We will progressively extend support to other GPUs.

How to use

from sageattention import sageattn
attn_output = sageattn(q, k, v, tensor_layout="HND", is_causal=False, smooth_k=True)

q, k, v are FP16/BF16 type with the shape (batch_size, head_num, seq_len, head_dim) using default tensor_layout="HND". For shape (batch_size, seq_len, head_num, head_dim), set tensor_layout="NHD". is_causal determines the use of a causal mask. smooth_k is a technique we proposed to ensure the accuracy. Disabling smooth_k might slightly increase speed, but could compromise accuracy if the distribution of q, k, v is irregular. In rare cases, setting smooth_k to False may result in better accuracy.

Note: sageattn is an accurate implementation that integrating smoothing K, INT8 per-block quantization for q, k, and a FP16 accumulator for Matmul of $PV$. Support for head_dim values of 64, 96, and 128 is currently available. Extended support for values 48, 72, and 256 will be available soon. Support for different sequence length between q and k,v and group-query attention is available. Support of different sequences length in the same batch is available through sageattn_varlen.

Plug-and-play Example

We can replace scaled_dot_product_attention easily.
We will take Cogvideo as an example:

Just add the following codes and run!

from sageattention import sageattn
import torch.nn.functional as F

F.scaled_dot_product_attention = sageattn

Specifically,

cd example
python sageattn_cogvideo.py

You can get a lossless video in ./example faster than by using python original_cogvideo.py

Note: Not all models use F.scaled_dot_product_attention, so maybe you should replace the original Attention by modifying the Attention Class of the target model (as shown in another example in ./example).

Performance

Speed of Kernels

Note: The TOPS results refer only to the Attention Kernel, excluding the quantization and smoothing K.

End-to-end performance

Citation

If you use this code or find our work valuable, please cite:

@misc{zhang2024sageattention,
      title={SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration}, 
      author={Jintao Zhang and Jia wei and Haofeng Huang and Pengle Zhang and Jun Zhu and Jianfei Chen},
      year={2024},
      eprint={2410.02367},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.02367}, 
}

@misc{zhang2024sageattention2,
      title={SageAttention2 Technical Report: Accurate 4 Bit Attention for Plug-and-play Inference Acceleration}, 
      author={Jintao Zhang and Haofeng Huang and Pengle Zhang and Jia Wei and Jun Zhu and Jianfei Chen},
      year={2024},
      eprint={2411.10958},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.10958}, 
}