Difference between 1.8.1+cu111 and 1.8.1 with cuda preinstalled

Software version: Cuda 11.1, Nvidia driver 470.x.

System 1 (GTX Titan X): I’m running into segmentation fault when training a model with 1.8.1+cu111 but works fine with 1.8.1 with cuda 11.1 preinstalled.
System 2 (Titan RTX): I’m running the same code without issues with 1.8.1+cu111 but gives seg fault with 1.8.1 with cude 11.1 preinstalled.

Can someone point out the difference between two package and any help on solving the above issue where I want to run the same code on two systems.

I don’t quite understand the “1.8.1 with cude 11.1 preinstalled” setup.
If you are installing the binaries (pip or conda) the local CUDA toolkit won’t be used unless you are building custom CUDA extensions, so what would be the difference to “1.8.1+cu111”?

i had met the similar issue.
i equipped double GTX 1080 Ti on the pc.
but it wasn’t work well with libtorch 1.9+cu10.2. also it works well with libtorch 1.9+cpu.
i wanted to handle it, but i didn’t.

also i can run training a model with pytorch 1.9+cu10.2 in pycharm project…