Julia's main selling point over Python is its speed, but most Python deep learning libraries are just wrappers around C/C++/Fortran/CUDA/Cython codebases that are as fast or faster than Julia.
This is the key point. Python is a scripting language with easy integration with libraries written in compiled languages. In most cases there would be no benefit in rewriting those libraries in Julia. Especially considering that doing so render such libraries harder to use by non-Julia users.