Building Tensorflow from source with docker: how to change cuda version?


We would like to rebuild tensorflow from source on your compute cluster to take advantage of hardware optimisation. We use the build based on docker since it’s difficult to install build dependencies on this cluster.
The build using cuda default option works but are relatively slow at runtime. We’re wondering if it is due to the fact that the build was done with cuda 11.2 and cuDNN 8 and in the cluster with installed cuda 11.4 ?
The other question is how to change the cuda and cudnn version when building tensorflow from sources ?



1 Like

I think that you can upvote and subscribe to:

E.g. the RBE image used in the CI was updated:

But not in the tensorflow/docker-deve GPU:

/cc @angerson