Hi, I was trying to update to pytorch v1.0.1 via conda, specifically—in a new environment—I am running
conda install -c pytorch pytorch==1.0.1=py3.7_cuda9.0.176_cudnn7.4.2_2 torchvision cudatoolkit=9.2
but this failed because that build of pytorch is expecting
Is cuda 9.2 going to be no longer supported?
FWIW, I have been working with:
conda install -c pytorch pytorch==1.0.0 torchvision cudatoolkit=9.2
and things have appeared to work fine for the most part (although I have sporadically received some
RuntimeError: CUDA error: invalid argument errors, which seem to disappear often upon running the same script again). I believe the conda build that was installed with the last command was
The build name suggests that I should use cuda v9.0, but I would like to use cuda v9.2 if possible since the HPC I am running on has CUDA 9.2 drivers. I guess I’m confused about which versions need to be lined up to get correct/optimal performance.
Thanks for the help!