tf_serving_install.md

Install TensorFlow Serving with Intel® Extension for TensorFlow*

TensorFlow Serving is an open-source system designed by Google that acts as a bridge between trained machine learning models and the applications that need to use them, streamlining the process of deploying and serving models in a production environment while maintaining efficiency and scalability.

Install Model Server

Install using Docker

A good way to get started using TensorFlow Serving with Intel® Extension for TensorFlow* is with Docker containers.

• Install Docker on Ubuntu 22.04

  sudo apt install docker
  sudo apt install docker.io
  

• Pull docker image

  # For CPU
  docker pull intel/intel-extension-for-tensorflow:serving-cpu
  
  # For GPU
  docker pull intel/intel-extension-for-tensorflow:serving-gpu 
  

Build from source

Tips:

• We recommend you put the source code of Intel® Extension for TensorFlow*, TensorFlow, and TensorFlow Serving in the same folder.
• Replace related paths with those on your machine.

1. Build Intel® Extension for TensorFlow* C++ library

Refer to Intel® Extension for TensorFlow* for C++ to build Intel® Extension for TensorFlow* C++ library

Note: When following this installation guide, you only need to build the Intel® Extension for TensorFlow* C++ library. You can ignore the other steps.

The generated libitex_cpu_cc.so or libitex_gpu_cc.so binary are found in the intel_extension_for_tensorflow/bazel-bin/itex/ directory.

2. Build TensorFlow Serving

• Patch TensorFlow

  • Get TensorFlow with commit id specified by TensorFlow Serving: https://github.com/tensorflow/serving/blob/master/WORKSPACE#L28

    # Exit intel-extension-for-tensorflow source code folder
    cd ..
    
    # clone TensorFlow
    git clone https://github.com/tensorflow/tensorflow
    
    # checkout specific commit id
    cd tensorflow
    git checkout xxxxx
    

  • Add alwayslink=1 for kernels_experimental library in local tensorflow/tensorflow/c/BUILD file:

    tf_cuda_library(
        name = "kernels_experimental",
        srcs = ["kernels_experimental.cc"],
        hdrs = ["kernels_experimental.h"],
        copts = tf_copts(),
        visibility = ["//visibility:public"],
        deps = [
            ...
        ] + if_not_mobile([
            ...
        ]),
        alwayslink=1, # add this line
    )
    

• Patch TensorFlow Serving

  • Get TensorFlow Serving source code

    # Exit tensorflow source code folder
    cd ..
    
    git clone https://github.com/tensorflow/serving
    

  • Patch TensorFlow Serving

    cd serving
    git checkout r2.14
    git apply ../intel-extension-for-tensorflow/third_party/tf_serving/serving_plugin.patch
    

• Build TensorFlow Serving

  bazel build --copt="-Wno-error=stringop-truncation" --config=release //tensorflow_serving/model_servers:tensorflow_model_server
  

  The generated tensorflow_model_server will be found in the serving/bazel-bin/tensorflow_serving/model_servers/ directory.

Build Docker image from Dockerfile

Refer to Intel® Extension for TensorFlow* Serving Docker Container Guide to build docker image from dockerfile.

Run sample

• Train and export TensorFlow model

  cd serving
  rm -rf /tmp/mnist
  python tensorflow_serving/example/mnist_saved_model.py /tmp/mnist
  

  Now let's take a look at the export directory. You should find a directory named 1 that contains saved_models.pb file and variables folder.

  ls /tmp/mnist
  1
  
  ls /tmp/mnist/1
  saved_model.pb variables
  

• Load exported model with TensorFlow ModelServer plugged with Intel® Extension for TensorFlow*

  • Use Docker from Docker Hub

    # For CPU
    docker run \
      -it \
      --rm \
      -p 8500:8500 \
      -e MODEL_NAME=mnist \
      -v /tmp/mnist:/models/mnist \
      intel/intel-extension-for-tensorflow:serving-cpu
    
    # For GPU
    docker run \
      -it \
      --rm \
      -p 8500:8500 \
      -e MODEL_NAME=mnist \
      -v /tmp/mnist:/models/mnist \
      --device /dev/dri/ \
      -v /dev/dri/by-path/:/dev/dri/by-path/ \
      intel/intel-extension-for-tensorflow:serving-gpu
    

    You will see:

    plugin library "/itex/bazel-bin/itex/libitex_cpu_cc.so" load successfully!
    
    plugin library "/itex/bazel-bin/itex/libitex_gpu_cc.so" load successfully!
    

  • Use tensorflow_model_server built from source

    # cd tensorflow_model_server binary folder
    
     # For CPU
    ./tensorflow_model_server \
      --port=8500 \
      --rest_api_port=8501 \
      --model_name=mnist \
      --model_base_path=/tmp/mnist \
      --tensorflow_plugins=path_to_libitex_cpu_cc.so
    
    # For GPU
    # source oneapi environment
    source oneapi_install_path/compiler/latest/env/vars.sh
    source oneapi_install_path/mkl/latest/env/vars.sh
    
    ./tensorflow_model_server \
      --port=8500 \
      --rest_api_port=8501 \
      --model_name=mnist \
      --model_base_path=/tmp/mnist \
      --tensorflow_plugins=path_to_libitex_gpu_cc.so
    

    You will see:

    plugin library "path_to_libitex_cpu_cc.so/libitex_cpu_cc.so" load successfully!
    
    plugin library "path_to_libitex_gpu_cc.so/libitex_gpu_cc.so" load successfully!
    

  • Use Docker built from dockerfile

    cd intel-extension-for-tensorflow source code folder
    
    cd docker/tensorflow-serving
    
    export MODEL_NAME=mnist
    export MODEL_DIR=/tmp/mnist
    
    ./run.sh [cpu/gpu]
    

    You will see:

    plugin library "/itex/itex-bazel-bin/bin/itex/libitex_cpu_cc.so" load successfully!
    
    plugin library "/itex/itex-bazel-bin/bin/itex/libitex_gpu_cc.so" load successfully!
    

• Test the server

  pip install tensorflow-serving-api
  
  cd serving
  python tensorflow_serving/example/mnist_client.py --num_tests=1000 --server=127.0.0.1:8500
  

  You will see:

  ...
  
  Inference error rate: xx.xx%
  

Refer to TensorFlow Serving Guides to learn more about how to use TensorFlow Serving.