text.md

Build TensorFlow Text from source

Note ⚠️

Please use Xcode 14.3 and Apple clang version 14.0.3 (clang-1403.0.22.14.1) or later versions is required.

Prerequisites

It is assumed here that you have the necessary Unix-Like knowledge, brew and conda have been installed in your terminal, and the installation and use methods of brew and conda will not be repeated here; most importantly, this tutorial is completely based on Apple silicon build, so make sure your Mac is Apple silicon.

Step by Step

1. Create a new Env and install the dependencies provided by Apple.

   conda create -n tensorflow-macos python=3.12 # Python 3.9, 3.10 and 3.11 are also supported.
   conda activate tensorflow-macos

2. Install the tensorflow.

   pip install tensorflow==2.16.1

3. Install bazel 6.5.0.

   wget https://github.com/bazelbuild/bazel/releases/download/6.5.0/bazel-6.5.0-darwin-arm64 -O bazel
   chmod +x bazel
   sudo mv bazel /usr/local/bin/
   bazel --version # Make sure the version is 6.5.0.

4. Download and extract text 2.16.1.

   wget https://github.com/tensorflow/text/archive/refs/tags/v2.16.1.zip
   unzip ./v2.16.1.zip
   cd text-2.16.1

5. Modify some parameters of the source code to ensure correct build.

   • oss_scripts/configure.sh to modify line 49:

     pip install tensorflow==2.16.1

   • oss_scripts/pip_package/setup.py to modify line 75:

     install_requires=[
         (
             'tensorflow>=2.16.1, <2.17',
         ),
     ],

6. Run the script.

   ./oss_scripts/run_build.sh

7. Please do not forget to install the whl file.

   pip install ./*.whl

Tips&Refer

1. Text needs to correspond to a minor version of tensorflow (e.g. tensorflow-macos==2.7.0 and tensorflow-text==2.7.3)
2. Please ensure that your network is stable during the compilation process, and the compilation needs to use the network.
3. I add a PR for Apple Silicon support.