I encountered a bug in training

It seems a problem of higher version PyTorch compatibility about CUDA interface. Since I write CUDA implementation of LIP in PyTorch v1.0, it is highly recommended to run experiments with PyTorch v1.0. Could you please provide the detailed version of your PyTorch?
If you like to stick with a higher version of PyTorch, an alternative way is to use the implementation of LIP consisting up of PyTorch primitives.

My pytorch version is 1.8. If I don't change the version, what should I do now

You could use either

LIP/cuda-lip/LIP/pyinterface.py

Line 41 in ea0631e

def primitive_lip2d(x, logit, kernel=3, stride=2, padding=1):

or

LIP/imagenet/lip_resnet.py

Line 19 in ea0631e

def lip2d(x, logit, kernel=3, stride=2, padding=1):

for your code. It is based on PyTorch primitives and free from the version compatibility problem.

If you have installed CUDA-LIP library already, you can simply call LIP.primitive_lip2d(x, logit, ...) instead of LIP.cuda_lip2d(...) to perform LIP.

The above error occurred when I used lip.cuda_lip2d. Now I use lip.primitive_Lip2d can run successfully. If there is any problem in the follow-up, I will consult again. Thanks for your help.