Follow these steps:

1. git clone && checkout r1.14
2. ./configure
   1. we could follow the default settings for the configure
   2. But on the server, we might want xla, cuda, tensorRT, MPI configured.
3. pip install numpy keras_preprocessing to installed some requirement.
4. bazel build -c dbg --copt="-DNDEBUG" --config cuda --strip=never //tensorflow/tools/pip_package:build_pip_package to build the wheel package
   1. -DNDEBUG: workaround absl string view problem
   2. --copt="-Og": will crash a internal parser crash, do not use.
5. ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg to generate the package
6. pip install tensorflow.whl to install.

more build instructions improvement It seems easier to just use nvidia’s build

clion

1. install the bazel plugin
2. ./configure the tensorflow environment
3. import tensorflow folder as bazel workspace
4. create new workspace file as

directories:
  .

derive_targets_from_directories: false

targets:
  //tensorflow/tools/pip_package:build_pip_package

additional_languages:
  python