Pytorch CUDA Initialization is Extremely Slow in A40 GPU (suddenly)

This problem could be solved by reinstalling CUDA, CUDNN, and NVIDIA Drivers. It was not known what suddenly happened to the normal Conda environment, but it is expected that a conflict between internal drivers may have been the cause.

First, when I deleted the driver with apt-get purge, there was a remaining file, so I had to search and remove it one by one with dpkg -l | grep. And to match the compatibility between CUDA versions, nvidia-driver 510, cuda 11.6, and cudnn 8.4.0.27 versions were installed. When installing dev from the NVIDIA official website, the latest version is installed, so I installed it through runfile. At this time, nvidia-driver should not be installed together, and nvidia-driver for CUDA 11.6 version had to be reinstalled separately. When installing 470 and 520 versions with different nvidia-driver versions, the existing problem could not be solved.