How to build Tensorflow 1.4 with CUDNN 5.0?

I'm trying to install Tensorflow 1.4 from sources with CUDA 8.0 and CUDNN 5.0.5, on Centos 7. It's indicated in the documentation that it should work with CUDNN 3 and higher. I'm working in a virtual env with Python 3.4.5, using Bazel 0.7.0, with GCC 4.9. During the configuration, I've set CUDNN version to 5.0.5 and the library has been found.

Unfortunately, it doesn't work and ends up with an error that seems to indicate that CUDNN v6 is needed (I may be wrong on the cause of the error).

Here is the command I'm using:

bazel build --config=opt --config=cuda --config=mkl //tensorflow/tools/pip_package:build_pip_packag

Here is the error:

ERROR: /home/localuser/tensorflow/tensorflow/stream_executor/BUILD:52:1: C++ compilation of rule '//tensorflow/stream_executor:cuda_platform' failed (Exit 1).
tensorflow/stream_executor/cuda/cuda_dnn.cc: In member function 'cudnnStatus_t perftools::gputools::cuda::wrap::WrapperShim__cudnnSetRNNDescriptor_v6::operator()(perftools::gputools::cuda::CUDAExecutor*, Args ...)':
tensorflow/stream_executor/cuda/cuda_dnn.cc:140:30: error: '::cudnnSetRNNDescriptor_v6' has not been declared
       cudnnStatus_t retval = ::__name(args...);                    \
                              ^
tensorflow/stream_executor/cuda/cuda_dnn.cc:235:3: note: in expansion of macro 'PERFTOOLS_GPUTOOLS_CUDNN_WRAP'
   __macro(cudnnSetRNNDescriptor_v6)                           \
   ^
tensorflow/stream_executor/cuda/cuda_dnn.cc:240:1: note: in expansion of macro 'CUDNN_DNN_ROUTINE_EACH_R5'
 CUDNN_DNN_ROUTINE_EACH_R5(PERFTOOLS_GPUTOOLS_CUDNN_WRAP)

Here is the result of the configure (.tf_configure.bazelrc):

build --action_env PYTHON_BIN_PATH="/home/localuser/venvs/tensorflow14-opt/bin/python"
build --action_env PYTHON_LIB_PATH="/home/localuser/venvs/tensorflow14-opt/lib/python3.4/site-packages"
build --define PYTHON_BIN_PATH="/home/localuser/venvs/tensorflow14-opt/bin/python"
build --define PYTHON_LIB_PATH="/home/localuser/venvs/tensorflow14-opt/lib/python3.4/site-packages"
build --force_python=py3
build --host_force_python=py3
build --python_path="/home/localuser/venvs/tensorflow14-opt/bin/python"
test --force_python=py3
test --host_force_python=py3
test --define PYTHON_BIN_PATH="/home/localuser/venvs/tensorflow14-opt/bin/python"
test --define PYTHON_LIB_PATH="/home/localuser/venvs/tensorflow14-opt/lib/python3.4/site-packages"
run --define PYTHON_BIN_PATH="/home/localuser/venvs/tensorflow14-opt/bin/python"
run --define PYTHON_LIB_PATH="/home/localuser/venvs/tensorflow14-opt/lib/python3.4/site-packages"
build:gcp --define with_gcp_support=true
build:hdfs --define with_hdfs_support=true
build:s3 --define with_s3_support=true
build:xla --define with_xla_support=true
build:gdr --define with_gdr_support=true
build:verbs --define with_verbs_support=true
build --action_env TF_NEED_OPENCL="0"
build --action_env TF_NEED_CUDA="1"
build --action_env CUDA_TOOLKIT_PATH="/usr/local/cuda"
build --action_env TF_CUDA_VERSION="8.0"
build --action_env CUDNN_INSTALL_PATH="/usr/local/cuda-8.0"
build --action_env TF_CUDNN_VERSION="5.0.5"
build --action_env TF_CUDA_COMPUTE_CAPABILITIES="3.7,3.7,3.7,3.7"
build --action_env TF_CUDA_CLANG="0"
build --action_env GCC_HOST_COMPILER_PATH="/opt/rh/devtoolset-3/root/usr/bin/gcc"
build --config=cuda
test --config=cuda
build:opt --cxxopt=-march=native --copt=-march=native
build:mkl --define using_mkl=true
build:mkl -c opt
build:mkl --copt="-DEIGEN_USE_VML"
build:monolithic --define framework_shared_object=false
build --define framework_shared_object=true

Is there anything I can do to install TF 1.4 using CUDNN 5 ?

Solution

The problem was a bug in Tensorflow where they were indeed using CUDNN V6 functions in a V5 build. This is now fixed by applying PR #12355 to the branch 1.4 (the PR is only applied to master).