@saudet , any idea how we can fix the build? There is a recurrent Cuda package error now on the GPUs platforms.

Looks like the CUDA Docker images from NVIDIA are broken? We might want to switch to some other way of installing CUDA...

Just out of curiosity, I'll try to build on top of Cuda 10.2 docker image to see if that works...

Mmh, worked on Cuda 10.2... https://github.com/tensorflow/java/runs/1101759245?check_suite_focus=true

I don't know if we can have issues at runtime, but according to @frankfliu, it will fallback anyway looking for Cuda 10.1 so this should be impact-less to our users. And according to this thread, the build should be backward compatible. Should I commit then?

No, it's probably not going to work with either CUDA 10.1 or 10.2...