When I select:
PyTorch Build: Stable (1.3)
Your OS: Linux
Package: Conda
Language: Python 3.7
CUDA: 9.2
It gives me this command:
conda install pytorch torchvision cudatoolkit=9.2 -c pytorch
My Question is, what does this mean?
(Let me give a starting but not complete answer to my question so you may understand me better. )
Answer (initial): The command means, you will use ‘conda’ package management software to install pytorch (1.3?) and torchvision (0.4?) which is pre-compiled with cuda library version 9.2. (Please correct the first part if it is incorrect or not complete) In addition, what is the ‘-c pytorch’ in the end?
-c pytorch means from the Pytorch anaconda channel, i.e. packages prepared by the PyTorch team (sometimes I think Soumith himself does that).
The pytorch and torchvision selects the latest packages available on that channel (where the stable releases are put, so currently 1.3 / 0.4, as you say).
The cuda library is provided in a separate package cudatoolkit of which you explicitly select version 9.2 on the command line.)
According to your answer, there is a misunderstanding of cuda installation in my prior description. I thought cudatoolkit=9.2 means the installed pytorch is pre-compiled with cudatoolkit=9.2. However, it should be that it will install cudatoolkit 9.2. Base on this PyTorch Forum Discussion, it also means it will not only install a cudatoolkit, but also install a corresponding version of CuDNN.
So, my question is that,
Is this cudatoolkit and cuDNN installation in our conda environment solely or it is system-wide?