Dear all,

I really don't know what to suggest here, or whether you will regard
it as "someone else's problem" (which would be quite reasonable!).  It
is currently not possible to build TensorFlow from source using the
Debian CUDA packages because of the way these packages have spread the
files across the filesystem (presumably to follow the FHS
requirements).  See
https://github.com/tensorflow/tensorflow/issues/40202 and
https://github.com/tensorflow/tensorflow/pull/42187 - the last two
comments on the pull request in particular.

Perhaps including suitable symlinks in the various CUDA packages so
that everything can be seen from one directory, in the way that it
appears in the NVIDIA packages, would be a way to help with this?

Best wishes,

   Julian