Windows wheel sageattention-2.2.0(+post3)+cu130torch2.10.0-cp312 crashes at import (_fused DLL load failed)
The Windows wheel sageattention-2.2.0(+post3)+cu130torch2.10.0-cp312-win_amd64.whl installs successfully, but crashes immediately at import time due to a failing native CUDA extension.
This appears to be a wheel build / ABI compatibility issue, similar to earlier FlashAttention Windows wheel problems.
Error message:
ImportError: DLL load failed while importing _fused: The specified procedure could not be found.
Failing import chain
from sageattention import sageattn
→ sageattention.core
→ sageattention.quant
→ import sageattention._fused
Environment
OS: Windows 11
Python: 3.12.11
PyTorch: 2.10.0 + CUDA 13.0
GPU: NVIDIA RTX 5090 (Blackwell / SM100)
Wheel: sageattention-2.2.0(+post3)+cu130torch2.10.0-cp312-win_amd64.whl
it's issue with not matching with torch 2.11
please add wheel of (cu130torch2.11.0)
sageattention-2.2.0.post3+cu130torch2.11.0-cp312-cp312-win_amd64
sageattention-2.2.0.post3+cu130torch2.11.0-cp313-cp313-win_amd64