We highly recommend using the spas_sage2_attn_meansim_topk_cuda and block_sparse_sage2_attn_cuda APIs. They are plug-and-play and customizable:

from spas_sage_attn import spas_sage2_attn_meansim_topk_cuda

attn_output = spas_sage2_attn_meansim_topk_cuda(q, k, v, topk=0.5, is_causal=False)

You can adjust topk to balance between attention accuracy (higher topk is more accurate) and sparsity (lower topk is more sparse).

from spas_sage_attn import block_sparse_sage2_attn_cuda

attn_output = block_sparse_sage2_attn_cuda(q, k, v, mask_id=None):

In this API, we support computing attention in any block sparse mask per attention head. Specifically, the per-head attention mask mask_id has shape (batch_size, num_heads, ⌈seq_len / 128⌉, ⌈seq_len // 64⌉) and consists of 0 and 1. Currently, the block size is 128×64.

The official implementation of SpargeAttn, a universal training-free sparse attention accelerating language, image, and video models.

SpargeAttention: Accurate and Training-free Sparse Attention
Accelerating Any Model Inference

• Please use the spas_sage2_attn_meansim_topk_cuda and block_sparse_sage2_attn_cuda APIs.
• [2025-07]: Release a Triton Kernel example.
• [2025-06]: SpargeAttn based on SageAttention2++ is released.
• [2025-05]: Add a very simple usage without tuning or calibration: o = spas_sage2_attn_meansim_topk_cuda(q, k, v).
• [2025-05]: 🎉SpargeAttn and SageAttention2 are accepted by ICML 2025!
• [2025-03] Support high acceleration on more GPUs, e.g., H100.

• python>=3.9 , torch>=2.3.0

• CUDA:
  • >=12.8 for Blackwell, >=12.4 for fp8 support on Ada, >=12.3 for fp8 support on Hopper, >=12.0 for Ampere

pip install ninja   # for parallel compilation
python setup.py install   # or pip install -e .

• spas_sage2_attn_meansim_topk_cuda: SpargeAttn based on SageAttention2 that we recommend using.

• spas_sage2_attn_meansim_cuda: SpargeAttn based on SageAttention2 that we do not recommend.

• spas_sage_attn_meansim_topk_cuda: SpargeAttn based on SageAttention that we recommend using.

• spas_sage_attn_meansim_cuda: SpargeAttn based on SageAttention that we do not recommend.

Just replace torch.nn.functional.scaled_dot_product_attention API using spas_sage2_attn_meansim_topk_cuda:

from spas_sage_attn import spas_sage2_attn_meansim_topk_cuda

- attn_output = torch.nn.functional.scaled_dot_product_attention(q, k, v, is_causal=False)  # is_causal can be True

+ attn_output = spas_sage2_attn_meansim_topk_cuda(q, k, v, topk=0.5, is_causal=False) # is_causal can be True

from spas_sage_attn import spas_sage2_attn_meansim_topk_cuda

attn_output = spas_sage2_attn_meansim_topk_cuda(q, k, v, topk=0.5, is_causal=False)

You can adjust topk to balance between attention accuracy (higher topk is more accurate) and sparsity (lower topk is more sparse).

from spas_sage_attn import block_sparse_sage2_attn_cuda

attn_output = block_sparse_sage2_attn_cuda(q, k, v, mask_id=None):

In this API, we support computing attention for any block-sparse mask per attention head. Specifically, the per-head attention mask mask_id has shape (batch_size, num_heads, ⌈seq_len / 128⌉, ⌈seq_len // 64⌉) and consists of 0 and 1. Currently, the block size is 128×64.

@inproceedings{zhang2025spargeattn,
  title={Spargeattn: Accurate sparse attention accelerating any model inference},
  author={Zhang, Jintao and Xiang, Chendong and Huang, Haofeng and Wei, Jia and Xi, Haocheng and Zhu, Jun and Chen, Jianfei},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2025}
}

@inproceedings{zhang2025sageattention,
  title={SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration}, 
  author={Zhang, Jintao and Wei, Jia and Zhang, Pengle and Zhu, Jun and Chen, Jianfei},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025}
}

@inproceedings{zhang2024sageattention2,
  title={Sageattention2: Efficient attention with thorough outlier smoothing and per-thread int4 quantization},
  author={Zhang, Jintao and Huang, Haofeng and Zhang, Pengle and Wei, Jia and Zhu, Jun and Chen, Jianfei},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2025}
}