Tensorflow Serving Build S2I

About

This S2I respository has template files used for building tensorflow wheel files for:

• Centos7
• Fedora28
• Fedora29
• Fedora30
• RHEL7.6

TODO:

• Fedora27
• RHEL7.5

NOTE: for RHEL7.5 you need a system with RHEL Subscription enabled.

For CPU

The files created from these templates are available at AICoE/tensorflow-wheels. And the instructions to use with pipfile are given here AICoE's TensorFlow Artifacts.

For GPU: TODO

GPU is not yet supported.

Bazel map with TensorFlow version

Bazel Version	Tensorflow version
0.24.1	1.14
0.21.0	1.13
0.18.0	1.12

Bazel build options

• TF_NEED_JEMALLOC: = 1
• TF_NEED_GCP: = 0
• TF_NEED_VERBS: = 0
• TF_NEED_HDFS: = 0
• TF_ENABLE_XLA: = 0
• TF_NEED_OPENCL: = 0
• TF_NEED_CUDA: = 1
• TF_NEED_MPI: = 0
• TF_NEED_GDR: = 0
• TF_NEED_S3: = 0
• TF_CUDA_VERSION: = 9.0
• TF_CUDA_COMPUTE_CAPABILITIES: = 3.0,3.5,5.2,6.0,6.1,7.0
• TF_CUDNN_VERSION: = 7
• TF_NEED_OPENCL_SYCL:= 0
• TF_NEED_TENSORRT:= 0
• TF_CUDA_CLANG:= 0
• GCC_HOST_COMPILER_PATH:= /usr/bin/gcc
• CUDA_TOOLKIT_PATH:= /usr/lib/cuda
• CUDNN_INSTALL_PATH:= /usr/lib/cuda
• TF_NEED_KAFKA:=0
• TF_NEED_OPENCL_SYCL:=0
• TF_DOWNLOAD_CLANG:=0
• TF_SET_ANDROID_WORKSPACE:=0
• TF_NEED_IGNITE: = 0
• TF_NEED_ROCM: = 0

Here is the default build command used to build tensorflow.

• CUSTOM_BUILD:=bazel build -c opt --local_resources 2048,2.0,1.0 --verbose_failures //tensorflow_serving/model_servers:tensorflow_model_server

• CUSTOM_BUILD_2:=bazel build -c opt --local_resources 2048,2.0,1.0 --verbose_failures tensorflow_serving/tools/pip_package:build_pip_package

Following should be left blank for a build job.

• TEST_LOOP:=
• BUILD_OPTS:=

Usage

To create a wheel file

set some environment values for convenience

# git token and repo
# valid values are 2.7,3.6,3.5
PYTHON_VERSION=3.6
GIT_TOKEN=
export GIT_RELEASE_REPO=

1. Create the templates

oc create -f tensorflow-build-image.json
oc create -f tensorflow-build-job.json
oc create -f tensorflow-build-dc.json

2. Create Tensorflow-Serving build image

oc new-app --template=tensorflow-serving-build-image  \
 --param=APPLICATION_NAME=tf-serving-fc29-build-image-${PYTHON_VERSION//.} \
 --param=S2I_IMAGE=registry.fedoraproject.org/f29/s2i-core   \
 --param=DOCKER_FILE_PATH=Dockerfile.fedora28 \
 --param=PYTHON_VERSION=$PYTHON_VERSION --param=BUILD_VERSION=1 --param=BAZEL_VERSION=0.21.0

oc new-app --template=tensorflow-serving-build-image  \
 --param=APPLICATION_NAME=tf-serving-centos7-build-image-${PYTHON_VERSION//.} \
 --param=S2I_IMAGE=openshift/base-centos7   \
 --param=DOCKER_FILE_PATH=Dockerfile.centos7 \
 --param=PYTHON_VERSION=$PYTHON_VERSION --param=BUILD_VERSION=1 --param=BAZEL_VERSION=0.21.0

The above command creates a tensorflow builder image APPLICATION_NAME:VERSION for specific OS.

The values for S2I_IMAGE are :

• Fedora28- registry.fedoraproject.org/f28/s2i-core
• Centos7- openshift/base-centos7

The values for DOCKER_FILE_PATH are :

• Fedora28- Dockerfile.fedora28
• Centos7- Dockerfile.centos7

3. Create Tensorflow wheel for CPU using the build image

oc new-app --template=tensorflow-serving-build-job  \
--param=APPLICATION_NAME=tf-serving-fc29-build-job-${PYTHON_VERSION//.} \
--param=BUILDER_IMAGESTREAM=tf-serving-fc29-build-image-${PYTHON_VERSION//.}:1  \
--param=PYTHON_VERSION=$PYTHON_VERSION     --param=GIT_TOKEN=$GIT_TOKEN \
--param=BAZEL_VERSION=0.21.0 --param=TF_GIT_BRANCH=r1.13 --param=CPU_LIMIT=36 --param=CPU_REQUESTS=36 --param=MEMORY_LIMIT=50Gi --param=MEMORY_REQUESTS=50Gi

oc new-app --template=tensorflow-serving-build-job  \
--param=APPLICATION_NAME=tf-serving-centos7-build-job-${PYTHON_VERSION//.} \
--param=BUILDER_IMAGESTREAM=tf-serving-centos7-build-image-${PYTHON_VERSION//.}:1  \
--param=PYTHON_VERSION=$PYTHON_VERSION     --param=GIT_TOKEN=$GIT_TOKEN \
--param=BAZEL_VERSION=0.21.0 --param=TF_GIT_BRANCH=r1.13 --param=CPU_LIMIT=36 --param=CPU_REQUESTS=36 --param=MEMORY_LIMIT=50Gi --param=MEMORY_REQUESTS=50Gi

OR

Import the template tensorflow-build-job.json into your namespace from Openshift UI. And then deploy from UI with appropriate values. Tensorflow wheel files will be pushed to $GIT_RELEASE_REPO using the token $GIT_TOKEN. (NOTE: This will ONLY work if the oauth token has scope of "repo". You can generate Personal API access token at https://github.com/settings/tokens. Minimal token scope is "repo".)

To create a DEV environment for debugging build issues :

oc new-app --template=tensorflow-serving-build-dc  \
--param=APPLICATION_NAME=tf-serving-build-dc-fc28-${PYTHON_VERSION//.} \
--param=BUILDER_IMAGESTREAM=tf-serving-fc28-build-image-${PYTHON_VERSION//.}:1  \
--param=PYTHON_VERSION=$PYTHON_VERSION     --param=GIT_TOKEN=$GIT_TOKEN \
--param=BAZEL_VERSION=0.18.0 --param=TF_GIT_BRANCH=r1.12

NOTE: BUILDER_IMAGESTREAM = APPLICATION_NAME:BUILD_VERSION from step 2.