Possible to install PyTorch 1.2 with CUDA 11?

vision
1

Hi, I wonder would that be possible to install an early version of Pytorch, such as 1.2 with CUDA 11?

I want to run an opensource code which was compiled under PyTorch 1.2 and CUDA 10.0. I set up an anaconda environment with the same config. However, I keep getting the error “RuntimeError: cuDNN error: CUDNN_STATUS_EXECUTION_FAILED”, despite Re-installing Pytorch and CUDA of the same versions.

I have a A6000 GPU, and I did not get such errors with Pytorch 1.9 with CUDA 11, but the opensource code did not seem to be compatible with Pytorch 1.9, I’ve got a lot of computability issues with numpy.

I appreciate any kind of help! Thank you!

2

You would have to check the CUDA11 PRs in GitHub, checkout the 1.2 branch, cherry-pick these PRs into this old branch, solve merge conflicts, and rebuild this version locally.
It depend how much time you would like to spend on this workflow, but I would guess that updating the old PyTorch script to be compatible with 1.9.1 might be easier and the better approach.