Noname010101/sageattention-wheels

SageAttention 2.2.0 wheel (sm_86 / sm_89 / sm_90a, torch 2.10 cu128)

Prebuilt Python wheel of SageAttention 2.2.0.

This is not affiliated with thu-ml. It is a community build of their open-source code. All credit for SageAttention goes to its authors. License: Apache-2.0 (see LICENSE).

Contents

File	Description
sageattention-2.2.0-sm86_89_90-py312-torch2.10-cu128.whl	The wheel (33,062,892 bytes)
sageattention-2.2.0-sm86_89_90-py312-torch2.10-cu128.whl.sha256	SHA-256 checksum of the wheel
LICENSE	Upstream Apache-2.0 license

Compatibility

• Python: 3.12 (cp312)
• PyTorch: 2.10.0+cu128
• CUDA: 12.8 build
• Platform: Linux x86_64
• GPU architectures:
  • sm_86: A40, A6000, RTX 3090
  • sm_89: L40S, RTX 6000 Ada, RTX 4090
  • sm_90a: H100

Other GPUs (for example sm_80 A100 or sm_120 Blackwell) are not included. Use the upstream source build for those.

Install

pip install --no-deps https://huggingface.co/Noname010101/sageattention-wheels/resolve/main/sageattention-2.2.0-sm86_89_90-py312-torch2.10-cu128.whl

--no-deps keeps pip from replacing your installed PyTorch. Install torch 2.10.0+cu128 first.

Verify the download:

sha256sum sageattention-2.2.0-sm86_89_90-py312-torch2.10-cu128.whl
# expected: e9aa95c748a3e85ea3f8d2abbb9609f36613b1630ebd9eca01a280fc73f70f5f

Build notes

• Source: upstream thu-ml/SageAttention at commit eb615cf6cf4d221338033340ee2de1c37fbdba4a.
• Source is unmodified, with one exception: the sm90 extension was built only for sm_90a.
• Build settings:
  • TORCH_CUDA_ARCH_LIST="8.6;8.9;9.0"
  • --no-build-isolation
  • CUDA 12.8
  • LIBRARY_PATH=/usr/local/cuda/lib64/stubs so -lcuda links against the stub library at build time

License

Apache License 2.0. Redistribution of this binary is permitted under the terms of that license, including the requirement to keep the license and notices. See LICENSE.

Source

• Upstream code: https://github.com/thu-ml/SageAttention
• Upstream project credit: thu-ml