There is an error in the blog,
I checked a bunch at the beginning, but in fact, upgrading cudnn to the latest version solved the problem.
Prompt that cuda and cudnn versions are not compatible
When the problem appeared, the machine was already CUDA v10.0.130, cudnn forgot to check, anyway, nvcc -V did not find it out.
According to this method,
Go here to download cudnn, remember to open VPN, otherwise the speed will be slower
https://developer.nvidia.com/rdp/cudnn-archive