torch 1.3 uses CUDA 10.1
But on my system I have CUDA 10.0 which I use for other apps (and for torch 1.2).
Is there a way to install multiple CUDA versions on Windows that won’t trash the system?
I am not very familiar with windows, @peterjc123 might have a better idea.
My advice for linux system would be to install cuda within a conda environment to have nice encapsulation. Can you use conda environment on your setup?
I use primarily Windows for CUDA-accelerated code nowadays. Linux - only occasionally.
I’ve tried to manage all my Python libraries only with pip so far, and I’ll keep doing that as long as possible. But I’ll give Conda some thought.
I don’t use conda either usually.
But it has this nice feature of being able to create a contained environment where you can install libraries without conflicting with the main system. So I use it to test a specific cuda or cudnn version temporarily.
If you are using conda, then it’s not possible to install two cudatoolkit packages in one environment. You’ll have to use virtual environment. Actually, if you use the wheels package, you’ll find that we copied all the CUDA DLLs alongside with the main library.
What’s more, different versions of CUDA won’t conflict with each other because the DLLs have different names, but cuDNN DLLs will.