Step-by-Step Guide to Installing TensorFlow 1.3.0 with GPU Support on CentOS 6.8

The author, an intern at Lianjia, documents the entire process of preparing a CentOS 6.8 machine for deep‑learning workloads, beginning with the acquisition of GPU servers (Tesla P100/V100) and outlining the two‑phase installation strategy.

Prerequisites : CentOS 6.8 (64‑bit) with gcc 4.4.7, Python 2.6.6, and internet access (or offline package preparation). The required software stack includes Java 8, gcc 4.9.4, Python 2.7, pip, setuptools, Bazel, NVIDIA driver, CUDA 8.0, cuDNN 6.0, and finally TensorFlow 1.3.0.

1. Install Java‑8 : Download the tar.gz from Oracle, extract, and add the bin directory to PATH.

2. Build and install gcc 4.9.4 :

wget http://ftp.gnu.org/gnu/gcc/gcc-4.9.4/gcc-4.9.4.tar.bz2

tar -xjf gcc-4.9.4.tar.bz2 && cd gcc-4.9.4

./contrib/download_prerequisites

mkdir build && cd build

../configure --prefix=/opt/gcc-4.9.4 --enable-languages=c,c++ --disable-multilib

make -j4

make install

Add the new gcc to PATH and set the required environment variables (CXX, CC, LDFLAGS, CXXFLAGS, etc.) in .bashrc.

3. Install Python 2.7 :

wget https://www.python.org/ftp/python/2.7.14/Python-2.7.14.tar.xz

tar -xf Python-2.7.14.tar.xz && cd Python-2.7.14

./configure --prefix=/usr/local

make -j4

make install

Update PATH or create symlinks so that python points to the new binary.

4. Fix yum’s Python dependency by editing /usr/bin/yum to use #!/usr/bin/python2.6 instead of the newer interpreter.

5. Upgrade pip and setuptools :

wget https://pypi.python.org/packages/.../setuptools‑…‑py2.py3-none-any.whl

python setup.py install

pip install --upgrade pip

Configure a fast domestic mirror in ~/.pip/pip.conf.

6. Install Bazel 0.5.3 :

wget https://github.com/bazelbuild/bazel/releases/download/0.5.3/bazel-0.5.3-dist.zip

unzip bazel-0.5.3-dist.zip -d bazel-0.5.3

cd bazel-0.5.3 && ./compile.sh

Copy the resulting binary to /usr/local/bin.

7. Install NVIDIA driver (Tesla P100): download NVIDIA-Linux-x86_64-384.66.run, make it executable, run it, and reboot.

8. Install CUDA 8.0 :

chmod +x cuda_8.0.61_375.26_linux.run

./cuda_8.0.61_375.26_linux.run

Select “no” for the bundled driver, accept the defaults, and add /usr/local/cuda-8.0/bin to PATH and /usr/local/cuda-8.0/lib64 to LD_LIBRARY_PATH.

9. Install cuDNN 6.0 for CUDA 8.0:

tar -zxvf cudnn-8.0-linux-x64-v6.0.tgz

cp cuda/include/cudnn.h /usr/local/cuda/include/

cp cuda/lib64/libcudnn* /usr/local/cuda/lib64/

Set read permissions on the copied files.

10. Build TensorFlow 1.3.0 from source :

git clone https://github.com/tensorflow/tensorflow

cd tensorflow && git checkout r1.3

./configure

During configuration answer “n” to jemalloc, “y” to CUDA support, and accept the defaults for other options.

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

If the build succeeds, package the wheel:

bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

Install it with:

pip install /tmp/tensorflow_pkg/tensorflow-1.3.0-py2-none-any.whl

11. Verification : Run python -c "import tensorflow as tf; print(tf.__version__)" and ensure no import errors. Use nvidia-smi and CUDA sample deviceQuery to confirm GPU visibility.

Common errors and fixes are listed throughout the guide, including missing kernel source, Bazel build failures due to missing patch, C++ compilation issues, and the need to disable jemalloc during TensorFlow configuration.

By following these steps, the author successfully created a functional TensorFlow GPU environment on CentOS 6.8, and shares the experience to help others avoid similar pitfalls.