Installing tensorflow for CUDA and ROCM within the same Python installation about tensorflow HOT 4 CLOSED

@samfux84 You're right about the challenges. Currently ROCm and CUDA versions of TensorFlow cannot coexist side-by-side within the same Python package. Could you create separate containers for each TensorFlow version (CUDA and ROCM), then select the desired container based on their specific GPU needs. You can also refer to this doc on multiple GPU architectures.
Thank you!

from tensorflow.

@sushreebarsa : Thank you for your reply. I somehow expected that currently it is not possible to have both packages available within one Python installation, but hoped that there is a workaround available.

A very simple solution would be if both packages (Tensorflow for CUDA and for ROCM) would have different package names. Then both packages could be installed within the same Python version.

I think that our current solution will be to have separate Python installations for CUDA and ROCM. This is not really desirable, as it duplicates the work of installing Python (and we plan to provide multiple versions of Python on our cluster).

Long term, it would be desirable to have a single tensorflow package that checks for the GPU type and then uses the appropriate code (Nvidia GPU/ AMD GPU/ Intel or AMD CPU), so that a single installation is sufficient.

Please feel free to close this ticket.

from tensorflow.

@samfux84 Thank you for the response!
Closing this ticket for now. Thank you!

from tensorflow.

Are you satisfied with the resolution of your issue?
Yes
No

from tensorflow.