Installation

Installing ModelServer

Installing using Docker

The easiest and most straight-forward way of using TensorFlow Serving is with Docker images. We highly recommend this route unless you have specific needs that are not addressed by running in a container.

TIP: This is also the easiest way to get TensorFlow Serving working with GPU support.

Installing using APT

Available binaries

The TensorFlow Serving ModelServer binary is available in two variants:

tensorflow-model-server: Fully optimized server that uses some platform specific compiler optimizations like SSE4 and AVX instructions. This should be the preferred option for most users, but may not work on some older machines.

tensorflow-model-server-universal: Compiled with basic optimizations, but doesn't include platform specific instruction sets, so should work on most if not all machines out there. Use this if tensorflow-model-server does not work for you. Note that the binary name is the same for both packages, so if you already installed tensorflow-model-server, you should first uninstall it using

apt-get remove tensorflow-model-server

Installation

  1. Add TensorFlow Serving distribution URI as a package source (one time setup)

    echo "deb [arch=amd64] http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | sudo tee /etc/apt/sources.list.d/tensorflow-serving.list && \
    curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | sudo apt-key add -
    
  2. Install and update TensorFlow ModelServer

    apt-get update && apt-get install tensorflow-model-server
    

Once installed, the binary can be invoked using the command tensorflow_model_server.

You can upgrade to a newer version of tensorflow-model-server with:

apt-get upgrade tensorflow-model-server

Building from source

The recommended approach to building from source is to use Docker. The TensorFlow Serving Docker development images encapsulate all the dependencies you need to build your own version of TensorFlow Serving.

For a listing of what these dependencies are, see the TensorFlow Serving Development Dockerfiles [CPU, GPU].

Installing Docker

General installation instructions are on the Docker site.

Clone the build script

After installing Docker, we need to get the source we want to build from. We will use Git to clone the master branch of TensorFlow Serving:

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

Build

In order to build in a hermetic environment with all dependencies taken care of, we will use the run_in_docker.sh script. This script passes build commands through to a Docker container. By default, the script will build with the latest nightly Docker development image.

TensorFlow Serving uses Bazel as its build tool. You can use Bazel commands to build individual targets or the entire source tree.

To build the entire tree, execute:

tools/run_in_docker.sh bazel build -c opt tensorflow_serving/...

Binaries are placed in the bazel-bin directory, and can be run using a command like:

bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server

To test your build, execute:

tools/run_in_docker.sh bazel test -c opt tensorflow_serving/...

See the basic tutorial and advanced tutorial for more in-depth examples of running TensorFlow Serving.

Building specific versions of TensorFlow Serving

If you want to build from a specific branch (such as a release branch), pass -b <branchname> to the git clone command.

We will also want to match the build environment for that branch of code, by passing the run_in_docker.sh script the Docker development image we'd like to use.

For example, to build version 1.10 of TensorFlow Serving:

$ git clone -b r1.10 https://github.com/tensorflow/serving.git
...
$ cd serving
$ tools/run_in_docker.sh -d tensorflow/serving:1.10-devel \
  bazel build tensorflow_serving/...
...
Optimized build

If you'd like to apply generally recommended optimizations, including utilizing platform-specific instruction sets for your processor, you can add --config=nativeopt to Bazel build commands when building TensorFlow Serving.

For example:

tools/run_in_docker.sh bazel build --config=nativeopt tensorflow_serving/...

It's also possible to compile using specific instruction sets (e.g. AVX). Wherever you see bazel build in the documentation, simply add the corresponding flag:

Instruction Set Flags
AVX --copt=mavx
AVX2 --copt=mavx2
FMA --copt=mfma
SSE 4.1 --copt=msse4.1
SSE 4.2 --copt=msse4.2
All supported by processor --copt=-march=native
Building with GPU Support

In order to build a custom version of TensorFlow Serving with GPU support, we recommend either building with the provided Docker images, or following the approach in the GPU Dockerfile.

TensorFlow Serving Python API PIP package

To run Python client code without the need to build the API, you can install the tensorflow-serving-api PIP package using:

pip install tensorflow-serving-api