harryming/sageattention-blackwell-wheels

SageAttention 2 and 3 wheels for Blackwell / CUDA 13

Prebuilt Linux x86_64 wheels used by the AI Love Island MiniMax H3 ComfyUI production pod. They are compiled from the official thu-ml/SageAttention source.

Compatibility

• Linux x86_64
• CPython 3.12 (cp312)
• PyTorch 2.10.0 + CUDA 13.0
• NVIDIA Blackwell, tested on RTX PRO 6000 Blackwell Server Edition
• SageAttention 2.2.0: Blackwell sm_120
• SageAttention 3 1.0.0: Blackwell FP4 sm_120a

These wheels are environment-specific. Do not install them on a different Python, Torch, CUDA, OS, or GPU architecture without a separate compatibility test.

Install

pip install --no-deps https://huggingface.co/harryming/sageattention-blackwell-wheels/resolve/main/wheels/torch2.10.0-cu130-cp312-blackwell/sageattention-2.2.0-cp312-cp312-linux_x86_64.whl

pip install --no-deps https://huggingface.co/harryming/sageattention-blackwell-wheels/resolve/main/wheels/torch2.10.0-cu130-cp312-blackwell/sageattn3-1.0.0-cp312-cp312-linux_x86_64.whl

In ComfyUI KJNodes:

• Sage 2/2++: PathchSageAttentionKJ / auto
• Sage 3: PathchSageAttentionKJ / sageattn3
• Keep allow_compile=false for the validated baseline.

SageAttention 3 may be faster on Blackwell but is not guaranteed to be lossless for every video model. Compare temporal stability, faces, hands, text, lip sync, and audio before replacing a Sage 2 production baseline. sageattn3_per_block_mean is not the approved default in this package.

Verification

sha256sum -c SHA256SUMS

License and source

The upstream project is Apache-2.0 licensed. Source, papers, and build instructions are maintained by the official project: https://github.com/thu-ml/SageAttention.