Conda installed cudatoolkit==10.1 but torch.version.cuda returns as 10.2

1

I have CUDA=10.1 installed
And I was trying to install maskrcnn-benchmark of pytorch so I created a conda environment and installed ```
conda install pytorch==1.5.1 torchvision==0.6.1 cudatoolkit=10.1 -c pytorch

And now when I print torch.version.cuda it returns me 10.2 therefore compiling nvidia apex gives me a runtime error!

Please help @ptrblck !
2

Could you check, if you have multiple PyTorch installations on your system and if so delete them and only install the desired one?

What’s your use case that you want to install an older PyTorch version and want to use apex?
Note that PyTorch 1.6 ships with native mixed-precision training via torch.cuda.amp and is the recommended approach.