Manually Downloading Dependency Packages for TF Serving Build

Dependency packages need to be downloaded from the network during TF Serving build. If the download fails due to network problems, the error message shown in Figure 1 is displayed.

Figure 1 TF Serving compilation error

The following describes the solution.

1. Download the required dependency packages via the following links and upload them to any directory on the server (for example, ${HOME}):

   tensorflow, rules_closure, bazel-skylib, rapidjson, abseil-cpp, libevent, and llvm

   The downloaded dependency packages can be used only after being renamed, as shown in the following table.

   Table 1 Renaming dependency packages

   Dependency	Before	After
   tensorflow	tensorflow-590d6eef7e91a6a7392c8ffffb7b58f2e0c8bc6b.tar.gz	590d6eef7e91a6a7392c8ffffb7b58f2e0c8bc6b.tar.gz
   rules_closure	rules_closure-316e6133888bfc39fb860a4f1a31cfcbae485aef.tar.gz	316e6133888bfc39fb860a4f1a31cfcbae485aef.tar.gz
   bazel-skylib	bazel-skylib-0.7.0.tar.gz	0.7.0.tar.gz
   rapidjson	rapidjson-1.1.0.zip	v1.1.0.zip
   abseil-cpp	abseil-cpp-36d37ab992038f52276ca66b9da80c1cf0f57dc2.tar.gz	36d37ab992038f52276ca66b9da80c1cf0f57dc2.tar.gz
   libevent	libevent-release-2.1.8-stable.zip	release-2.1.8-stable.zip
   llvm	llvm-7a7e03f906aada0cf4b749b51213fe5784eeff84.tar.gz	7a7e03f906aada0cf4b749b51213fe5784eeff84.tar.gz

2. Add the --distdir parameter when compiling TF Serving as follows:

   bazel --output_user_root=/opt/tf_serving build -c opt --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0" --distdir=${HOME}/tensorflow_serving/model_servers:tensorflow_model_server