NVIDIA/flownet2-pytorch

CUDA error: the provided PTX was compiled with an unsupported toolchain

zinuok opened this issue · 0 comments

I built the custom layers via "python3 setup.py install --user" in each directory (of resample2d, channelnorm, correlation).
However, after importing the following custom layer,

from ...(some_path)... correlation_package.correlation import Correlation

I cannot transfer the tensor to the GPU:

a = np.array(1)
torch.from_numpy(a).to('cuda')

producing the following error.

RuntimeError: CUDA error: the provided PTX was compiled with an unsupported toolchain.

(but, torch.cuda.is_available() returns True)

my settings are as follows:

  • GPU: RTX 3080 Ti
  • driver: 510.108.03
  • CUDA: 11.7 (I also tried with 11.1 but the same error occurred!)
  • PyTorch: 2.0.0

Is there any related-issue about this?

Thank you