Install nvidia-cuda-toolkit 9.0.176-2 on Debian (buster).

At first, need to download all the dependencies from http://ftp.riken.jp/Linux/debian/debian/pool/non-free/n/nvidia-cuda-toolkit/, and install these.

# dpkg -i libaccinj64-9.0_9.0.176-2_amd64.deb libnppig9.0_9.0.176-2_amd64.deb libcublas9.0_9.0.176-2_amd64.deb libnppim9.0_9.0.176-2_amd64.deb libcudart9.0_9.0.176-2_amd64.deb libnppist9.0_9.0.176-2_amd64.deb libcuinj64-9.0_9.0.176-2_amd64.deb libnppisu9.0_9.0.176-2_amd64.deb libcurand9.0_9.0.176-2_amd64.deb libnppitc9.0_9.0.176-2_amd64.deb libcusolver9.0_9.0.176-2_amd64.deb libnpps9.0_9.0.176-2_amd64.deb libcusparse9.0_9.0.176-2_amd64.deb libnvgraph9.0_9.0.176-2_amd64.deb libnppc9.0_9.0.176-2_amd64.deb libnvrtc9.0_9.0.176-2_amd64.deb libnppial9.0_9.0.176-2_amd64.deb libnvtoolsext1_9.0.176-2_amd64.deb libnppicc9.0_9.0.176-2_amd64.deb libnvvm3_9.0.176-2_amd64.deb libnppicom9.0_9.0.176-2_amd64.deb  libnppidei9.0_9.0.176-2_amd64.deb libnppif9.0_9.0.176-2_amd64.deb  

In order to prevent automatic upgrade, mark some packages as hold.

# apt-mark hold libnvtoolsext1 libnvvm3 nvidia-cuda-dev nvidia-cuda-toolkit nvidia-profiler

Indeed, minor dependencies should be downloaded and installed.
# apt install gcc-6 g++-6 clang-4.9

And, finalize the install of nvidia-cuda-toolkit.

# dpkg -i nvidia-profiler_9.0.176-2_amd64.deb
# dpkg -i nvidia-cuda-dev_9.0.176-2_amd64.deb
# dpkg -i nvidia-cuda-toolkit_9.0.176-2_amd64.deb 

Next, download and install remaining dependencies from https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64/

# dpkg -i libcudnn7_7.0.5.15-1+cuda9.0_amd64.deb
# dpkg -i libcudnn7-dev_7.0.5.15-1+cuda9.0_amd64.deb
# dpkg -i libcudnn7-doc_7.0.5.15-1+cuda9.0_amd64.deb

(venv) $ python3.6
>>> pip install tensorflow-gpu

In terms of virtualenv, please see the other post about it.