Build from source

Build a TensorFlow pip package from source and install it on Ubuntu Linux and macOS. While the instructions might work for other systems, it is only tested and supported for Ubuntu and macOS.

Setup for Linux and macOS

Install the following build tools to configure your development environment.

Install Python and the TensorFlow package dependencies

sudo apt install python3-dev python3-pip

brew install python

Install the TensorFlow pip package dependencies (if using a virtual environment, omit the --user argument):

pip install -U --user pip

Install Bazel

To build TensorFlow, you will need to install Bazel. Bazelisk is an easy way to install Bazel and automatically downloads the correct Bazel version for TensorFlow. For ease of use, add Bazelisk as the bazel executable in your PATH.

If Bazelisk is not available, you can manually install Bazel. Make sure to install the correct Bazel version from TensorFlow's .bazelversion file.

Install Clang (recommended, Linux only)

Clang is a C/C++/Objective-C compiler that is compiled in C++ based on LLVM. It is the default compiler to build TensorFlow starting with TensorFlow 2.13. The current supported version is LLVM/Clang 17.

LLVM Debian/Ubuntu nightly packages provide an automatic installation script and packages for manual installation on Linux. Make sure you run the following command if you manually add llvm apt repository to your package sources:

sudo apt-get update && sudo apt-get install -y llvm-17 clang-17

Now that /usr/lib/llvm-17/bin/clang is the actual path to clang in this case.

Alternatively, you can download and unpack the pre-built Clang + LLVM 17.

Below is an example of steps you can take to set up the downloaded Clang + LLVM 17 binaries on Debian/Ubuntu operating systems:

1. Change to the desired destination directory: cd <desired directory>

2. Load and extract an archive file...(suitable to your architecture):

   wget https://github.com/llvm/llvm-project/releases/download/llvmorg-17.0.2/clang+llvm-17.0.2-x86_64-linux-gnu-ubuntu-22.04.tar.xz
   
   tar -xvf clang+llvm-17.0.2-x86_64-linux-gnu-ubuntu-22.04.tar.xz
   
   

3. Copy the extracted contents (directories and files) to /usr (you may need sudo permissions, and the correct directory may vary by distribution). This effectively installs Clang and LLVM, and adds it to the path. You should not have to replace anything, unless you have a previous installation, in which case you should replace the files:

   cp -r clang+llvm-17.0.2-x86_64-linux-gnu-ubuntu-22.04/* /usr
   

4. Check the obtained Clang + LLVM 17 binaries version:

   clang --version
   

5. Now that /usr/bin/clang is the actual path to your new clang. You can run the ./configure script or manually set environment variables CC and BAZEL_COMPILER to this path.

Install GPU support (optional, Linux only)

There is no GPU support for macOS.

Read the GPU support guide to install the drivers and additional software required to run TensorFlow on a GPU.

Download the TensorFlow source code

Use Git to clone the TensorFlow repository:

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow

The repo defaults to the master development branch. You can also check out a release branch to build:

git checkout branch_name  # r2.2, r2.3, etc.

Configure the build

TensorFlow builds are configured by the .bazelrc file in the repository's root directory. The ./configure or ./configure.py scripts can be used to adjust common settings.

Please run the ./configure script from the repository's root directory. This script will prompt you for the location of TensorFlow dependencies and asks for additional build configuration options (compiler flags, for example). Refer to the Sample session section for details.

./configure

There is also a python version of this script, ./configure.py. If using a virtual environment, python configure.py prioritizes paths within the environment, whereas ./configure prioritizes paths outside the environment. In both cases you can change the default.

Sample session

The following shows a sample run of ./configure script (your session may differ):

./configure
You have bazel 6.1.0 installed.
Please specify the location of python. [Default is /Library/Frameworks/Python.framework/Versions/3.9/bin/python3]: 


Found possible Python library paths:
  /Library/Frameworks/Python.framework/Versions/3.9/lib/python3.9/site-packages
Please input the desired Python library path to use.  Default is [/Library/Frameworks/Python.framework/Versions/3.9/lib/python3.9/site-packages]

Do you wish to build TensorFlow with ROCm support? [y/N]:
No ROCm support will be enabled for TensorFlow.

Do you wish to build TensorFlow with CUDA support? [y/N]:
No CUDA support will be enabled for TensorFlow.

Do you want to use Clang to build TensorFlow? [Y/n]:
Clang will be used to compile TensorFlow.

Please specify the path to clang executable. [Default is /usr/lib/llvm-16/bin/clang]:

You have Clang 16.0.4 installed.

Please specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is -Wno-sign-compare]:


Would you like to interactively configure ./WORKSPACE for Android builds? [y/N]: n
Not configuring the WORKSPACE for Android builds.

Do you wish to build TensorFlow with iOS support? [y/N]: n
No iOS support will be enabled for TensorFlow.

Preconfigured Bazel build configs. You can use any of the below by adding "--config=<>" to your build command. See .bazelrc for more details.
    --config=mkl            # Build with MKL support.
    --config=mkl_aarch64    # Build with oneDNN and Compute Library for the Arm Architecture (ACL).
    --config=monolithic     # Config for mostly static monolithic build.
    --config=numa           # Build with NUMA support.
    --config=dynamic_kernels    # (Experimental) Build kernels into separate shared objects.
    --config=v1             # Build with TensorFlow 1 API instead of TF 2 API.
Preconfigured Bazel build configs to DISABLE default on features:
    --config=nogcp          # Disable GCP support.
    --config=nonccl         # Disable NVIDIA NCCL support.

Configuration options

GPU support

from v.2.18.0

For GPU support, set cuda=Y during configuration and specify the versions of CUDA and cuDNN if required. Bazel will download CUDA and CUDNN packages automatically or point to CUDA/CUDNN/NCCL redistributions on local file system if required.

before v.2.18.0

For GPU support, set cuda=Y during configuration and specify the versions of CUDA and cuDNN. If your system has multiple versions of CUDA or cuDNN installed, explicitly set the version instead of relying on the default. ./configure creates symbolic links to your system's CUDA libraries—so if you update your CUDA library paths, this configuration step must be run again before building.

Optimizations

For compilation optimization flags, the default (-march=native) optimizes the generated code for your machine's CPU type. However, if building TensorFlow for a different CPU type, consider a more specific optimization flag. Check the GCC manual for examples.

Preconfigured configurations

There are some preconfigured build configs available that can be added to the bazel build command, for example:

• --config=dbg —Build with debug info. See CONTRIBUTING.md for details.
• --config=mkl —Support for the Intel® MKL-DNN.
• --config=monolithic —Configuration for a mostly static, monolithic build.

Build and install the pip package

Bazel build options

Refer to the Bazel command-line reference for build options.

Building TensorFlow from source can use a lot of RAM. If your system is memory-constrained, limit Bazel's RAM usage with: --local_ram_resources=2048.

The official TensorFlow packages are built with a Clang toolchain that complies with the manylinux2014 package standard.

Build the package

To build pip package, you need to specify --repo_env=WHEEL_NAME flag. depending on the provided name, package will be created, e.g:

To build tensorflow CPU package:

bazel build //tensorflow/tools/pip_package:wheel --repo_env=WHEEL_NAME=tensorflow_cpu

To build tensorflow GPU package:

bazel build //tensorflow/tools/pip_package:wheel --repo_env=WHEEL_NAME=tensorflow --config=cuda --config=cuda_wheel

To build tensorflow TPU package:

bazel build //tensorflow/tools/pip_package:wheel --repo_env=WHEEL_NAME=tensorflow_tpu --config=tpu

To build nightly package, set tf_nightly instead of tensorflow, e.g. to build CPU nightly package:

bazel build //tensorflow/tools/pip_package:wheel --repo_env=WHEEL_NAME=tf_nightly_cpu

As a result, generated wheel will be located in

bazel-bin/tensorflow/tools/pip_package/wheel_house/

Install the package

The filename of the generated .whl file depends on the TensorFlow version and your platform. Use pip install to install the package, for example:

pip install bazel-bin/tensorflow/tools/pip_package/wheel_house/tensorflow-version-tags.whl

Docker Linux builds

TensorFlow's Docker development images are an easy way to set up an environment to build Linux packages from source. These images already contain the source code and dependencies required to build TensorFlow. Go to the TensorFlow Docker guide for installation instructions and the list of available image tags.

CPU-only

The following example uses the :devel image to build a CPU-only package from the latest TensorFlow source code. Check the Docker guide for available TensorFlow -devel tags.

Download the latest development image and start a Docker container that you'll use to build the pip package:

docker pull tensorflow/tensorflow:devel
docker run -it -w /tensorflow_src -v $PWD:/mnt -e HOST_PERMS="$(id -u):$(id -g)" \
    tensorflow/tensorflow:devel bash

git pull  # within the container, download the latest source code

The above docker run command starts a shell in the /tensorflow_src directory—the root of the source tree. It mounts the host's current directory in the container's /mnt directory, and passes the host user's information to the container through an environmental variable (used to set permissions—Docker can make this tricky).

Alternatively, to build a host copy of TensorFlow within a container, mount the host source tree at the container's /tensorflow directory:

docker run -it -w /tensorflow -v /path/to/tensorflow:/tensorflow -v $PWD:/mnt \
    -e HOST_PERMS="\\((id -u):\\)(id -g)" tensorflow/tensorflow:devel bash

With the source tree set up, build the TensorFlow package within the container's virtual environment:

1. Optional: Configure the build—this prompts the user to answer build configuration questions.
2. Build the pip package.
3. Adjust the ownership permissions of the file for outside the container.

./configure  # if necessary

bazel build //tensorflow/tools/pip_package:wheel --repo_env=WHEEL_NAME=tensorflow_cpu --config=opt
`
chown $HOST_PERMS bazel-bin/tensorflow/tools/pip_package/wheel_house/tensorflow-version-tags.whl

Install and verify the package within the container:

pip uninstall tensorflow  # remove current version

pip install bazel-bin/tensorflow/tools/pip_package/wheel_house/tensorflow-version-tags.whl
cd /tmp  # don't import from source directory
python -c "import tensorflow as tf; print(tf.__version__)"

On your host machine, the TensorFlow pip package is in the current directory (with host user permissions): ./tensorflow-version-tags.whl

GPU support

Docker is the easiest way to build GPU support for TensorFlow since the host machine only requires the NVIDIA® driver (the NVIDIA® CUDA® Toolkit doesn't have to be installed). Refer to the GPU support guide and the TensorFlow Docker guide to set up nvidia-docker (Linux only).

The following example downloads the TensorFlow :devel-gpu image and uses nvidia-docker to run the GPU-enabled container. This development image is configured to build a pip package with GPU support:

docker pull tensorflow/tensorflow:devel-gpu
docker run --gpus all -it -w /tensorflow -v $PWD:/mnt -e HOST_PERMS="$(id -u):$(id -g)" \
    tensorflow/tensorflow:devel-gpu bash
git pull  # within the container, download the latest source code

Then, within the container's virtual environment, build the TensorFlow package with GPU support:

./configure  # if necessary

bazel build //tensorflow/tools/pip_package:wheel --repo_env=WHEEL_NAME=tensorflow --config=cuda --config=cuda_wheel --config=opt

chown $HOST_PERMS bazel-bin/tensorflow/tools/pip_package/wheel_house/tensorflow-version-tags.whl

Install and verify the package within the container and check for a GPU:

pip uninstall tensorflow  # remove current version

pip install bazel-bin/tensorflow/tools/pip_package/wheel_house/tensorflow-version-tags.whl
cd /tmp  # don't import from source directory
python -c "import tensorflow as tf; print(\"Num GPUs Available: \", len(tf.config.list_physical_devices('GPU')))"

Tested build configurations

Linux

CPU

GPU

macOS

CPU

GPU