Developing with Docker
Pulling a development image
For a development environment where you can build TensorFlow Serving, you can try:
docker pull tensorflow/serving:latest-devel
For a development environment where you can build TensorFlow Serving with GPU support, use:
docker pull tensorflow/serving:latest-devel-gpu
See the Docker Hub tensorflow/serving repo for other versions of images you can pull.
After pulling one of the development Docker images, you can run it while opening the gRPC port (8500):
docker run -it -p 8500:8500 --gpus all tensorflow/serving:latest-devel
Testing the development environment
To test a model, from inside the container try:
# train the mnist model python tensorflow_serving/example/mnist_saved_model.py /tmp/mnist_model # serve the model tensorflow_model_server --port=8500 --model_name=mnist --model_base_path=/tmp/mnist_model/ & # test the client python tensorflow_serving/example/mnist_client.py --num_tests=1000 --server=localhost:8500
We currently maintain the following Dockerfiles:
Dockerfile, which is a minimal VM with TensorFlow Serving installed.
Dockerfile.gpu, which is a minimal VM with TensorFlow Serving with GPU support to be used with
Dockerfile.devel, which is a minimal VM with all of the dependencies needed to build TensorFlow Serving.
Dockerfile.devel-gpu, which is a minimal VM with all of the dependencies needed to build TensorFlow Serving with GPU support.
Building a container from a Dockerfile
If you'd like to build your own Docker image from a Dockerfile, you can do so by running the Docker build command:
docker build --pull -t $USER/tensorflow-serving .
docker build --pull -t $USER/tensorflow-serving-gpu -f Dockerfile.gpu .
docker build --pull -t $USER/tensorflow-serving-devel -f Dockerfile.devel .
docker build --pull -t $USER/tensorflow-serving-devel-gpu -f Dockerfile.devel-gpu .
TIP: Before attempting to build an image, check the Docker Hub tensorflow/serving repo to make sure an image that meets your needs doesn't already exist.
Building from sources consumes a lot of RAM. If RAM is an issue on your system,
you may limit RAM usage by specifying
invoking Bazel. See the
for more information. You can use this same mechanism to tweak the optmizations
you're building TensorFlow Serving with. For example:
docker build --pull --build-arg TF_SERVING_BUILD_OPTIONS="--copt=-mavx \ --cxxopt=-D_GLIBCXX_USE_CXX11_ABI=0 --local_ram_resources=2048" -t \ $USER/tensorflow-serving-devel -f Dockerfile.devel .
Running a container
This assumes you have built the
To run the container opening the gRPC port (8500):
docker run -it -p 8500:8500 $USER/tensorflow-serving-devel
TIP: If you're running a GPU image, be sure to run using the NVIDIA runtime
From here, you can follow the instructions for testing a development environment.
Building an optimized serving binary
When running TensorFlow Serving's ModelServer, you may notice a log message that looks like this:
I external/org_tensorflow/tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
This indicates that your ModelServer binary isn't fully optimized for the CPU its running on. Depending on the model you are serving, further optimizations may not be necessary. However, building an optimized binary is straight-forward.
When building a Docker image from the provided
Dockerfile.devel-gpu files, the ModelServer binary will be built with the flag
-march=native. This will cause Bazel to build a ModelServer binary with all of
the CPU optimizations the host you're building the Docker image on supports.
To create a serving image that's fully optimized for your host, simply:
Clone the TensorFlow Serving project
git clone https://github.com/tensorflow/serving cd serving
Build an image with an optimized ModelServer
docker build --pull -t $USER/tensorflow-serving-devel \ -f tensorflow_serving/tools/docker/Dockerfile.devel .
For GPU: `
docker build --pull -t $USER/tensorflow-serving-devel-gpu \ -f tensorflow_serving/tools/docker/Dockerfile.devel-gpu .
Build a serving image with the development image as a base
docker build -t $USER/tensorflow-serving \ --build-arg TF_SERVING_BUILD_IMAGE=$USER/tensorflow-serving-devel \ -f tensorflow_serving/tools/docker/Dockerfile .
Your new optimized Docker image is now
$USER/tensorflow-serving, which you can use just as you would the standard
docker build -t $USER/tensorflow-serving-gpu \ --build-arg TF_SERVING_BUILD_IMAGE=$USER/tensorflow-serving-devel-gpu \ -f tensorflow_serving/tools/docker/Dockerfile.gpu .
Your new optimized Docker image is now
$USER/tensorflow-serving-gpu, which you can use just as you would the standard