Building TensorFlow for CentOS 6

Posted on June 30, 2016April 1, 2021 by Kuan-Yi Li

You should no longer need this as you can install and run TensorFlow smoothly with Anaconda now!

Just install Anaconda and install TensorFlow by running

conda install tensorflow

conda install tensorflow-gpu

and you’re good to go since Continuum Analytics have applied our patch.

Note:
Do NOT install TensorFlow package from conda-forge as that is for newer version of Linux distros.

Introduction

The official-released binary packages of TensorFlow are built for newer version of Linux distros. Here is how to build TensorFlow binary package for CentOS 6.

Install Dependencies

Admin privilege is required here.

OpenJDK (for Bazel)

yum install java-1.8.0-openjdk-devel

Developer Toolset (for Bazel and TensorFlow)

C++11 compatible compiler is required during the building process. Follow the instructions in this article to install Developer Toolset. It is worth noting that, in this case, installing just devtoolset-6-toolchain would be sufficient.

Python 2.7 (for TensorFlow)

This step also requires Software Collections installed. Therefore, if you’ve installed Developer Toolset by following the instructions above, you should have no problem installing these dependencies.

yum install python27 python27-numpy python27-python-devel python27-python-wheel

Build and Install Bazel

Enter Build Environment

Enter Software Collection environment with Developer Toolset.

scl enable devtoolset-6 bash

Prepare Source Code

Method 1 (Recommended)

The standard way of compiling a release version of Bazel from source is to use a distribution archive. From version 0.4.1, bootstrapping Bazel from GitHub repository is no longer supported as only the distribution archives contain generated artifacts required for direct compilation. Download bazel-<VERSION>-dist.zip from the release page.

# download distribution archive
wget https://github.com/bazelbuild/bazel/releases/download/0.9.0/bazel-0.9.0-dist.zip
unzip bazel-0.9.0-dist.zip -d bazel-0.9.0-dist
cd bazel-0.9.0-dist

Method 2 (Only Works for Version <= 0.4.0)

# clone source code repository
git clone https://github.com/bazelbuild/bazel.git
cd bazel

# select version (optional)
git checkout 0.4.0

Build and Install

# compile
./compile.sh

# install
mkdir -p ~/bin
cp output/bazel ~/bin/

# exit from Software Collection environment
exit

Build TensorFlow with Bazel

Enter Build Environment

Enter Software Collection environment with Developer Toolset and Python 2.7.

scl enable devtoolset-6 python27 bash

Prepare Source Code

# clone source code repository
git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow

# select version (optional)
git checkout v1.5.0

Since GNU C library version in CentOS 6 is less than 2.17, a slight modification needs to be applied before compilation.

Modify tf_extension_linkopts function in tensorflow/tensorflow.bzl from

def tf_extension_linkopts():
  return []  # No extension link opts

def tf_extension_linkopts():
  return ["-lrt"]

Build

After above modification, we can now start building!

Here we disable jemalloc when configuring — otherwise the build will fail.

# configure workspace
export \
    PYTHON_BIN_PATH=/opt/rh/python27/root/usr/bin/python \
    PYTHON_LIB_PATH=/opt/rh/python27/root/usr/lib/python2.7/site-packages \
    TF_NEED_JEMALLOC=0 \
    TF_NEED_GCP=0 \
    TF_NEED_HDFS=0 \
    TF_NEED_S3=0 \
    TF_ENABLE_XLA=0 \
    TF_NEED_GDR=0 \
    TF_NEED_VERBS=0 \
    TF_NEED_OPENCL=0 \
    TF_NEED_CUDA=0 \
    TF_NEED_MPI=0 \
    CC_OPT_FLAGS="-march=native"
./configure

# build
~/bin/bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

# exit from Software Collection environment
exit

Finishing Up

Finally, you will found a tensorflow-1.5.0-cp27-none-linux_x86_64.whl (filename may vary upon versions) in /tmp/tensorflow_pkg/. With this self-built binary package, we can now deploy TensorFlow to CentOS 6 by pip.

Enter Python Environment on Deployment Target

Enter Python 2.7 environment either by Software Collection or by tools like pyenv.

# by Software Collection
scl enable python27 bash

# or by pyenv
pyenv shell anaconda2-5.0.1

Install

Depending on your environment setup, admin privilege might be needed here.

pip install tensorflow-1.5.0-cp27-none-linux_x86_64.whl

Note:
sudo clears environment variables before running commands for security reasons, causing Software Collection environment, which relies on environment variables, to dysfunction. If you are installing TensorFlow package in Software Collection environment through sudo, you might find this command helpful.

sudo scl enable python27 -- pip install tensorflow-1.5.0-cp27-none-linux_x86_64.whl

It sets up Software Collection environment variables AFTER sudo cleans them up.

References

Posted in Technical.

11 Comments

syyuan

August 5, 2016 at 06:23 Reply

Hi, I followed your method and successfully intalled tensorflow on CentOS. Thank you so much! Did you try to install it with GPU support?
- Kuan-Yi Li
  
  August 10, 2016 at 22:15 Reply
  
  Glad you found it helpful. I didn’t install TensorFlow with GPU support since I currently don’t have compatible hardware at hand. However, I believe it should be easy if you want GPU feature enabled. Just install CUDA and build the pip package with GPU support (which means adding --config=cuda to the build command).
Hangi Kim

August 18, 2016 at 08:42 Reply

Dear Kuan-Yi Li
Thank you a lot for you article.
It is very helpful to me.

I suggest add the following lines
.
yum install python27-setuptools
scl enable devtoolset-4 python27 bash
easy_install pip
- Kuan-Yi Li
  
  August 18, 2016 at 16:30 Reply
  
  Glad you found it helpful. In regard to pip, package python27 requires python27-python-setuptools and python27-python-pip, so installing python27 should get you covered.
Igor

December 2, 2016 at 02:37 Reply

Huge thanks! The ‘return [“-lrt”]’ is wonderful! I was finally able to make tensorflow r0.11 work on my CentOS 6.5 with glibc version 2.12.

This post seem to be the only place, which explicitly mentions that this is the fix for glibc-version related issues. Now, after understanding that this is the fix, I can see plenty of posts, mentioning this option in github-issues posts. I saw them before, but they were not looking like something related to “ImportError: /lib64/libc.so.6: version `GLIBC_2.14′ not found” for me…
Michael Weiner

July 9, 2017 at 01:47 Reply

Thank you for the post, the tip was appreciated, however i am having an issue compiling this on an HPC running CentOS 6.5 and python 2.7.12. I built bazel with no issues, but when i go to build the pip package it fails with the following:

ERROR: /root/temp/tensorflow/tensorflow/core/BUILD:1395:1: Executing genrule //tensorflow/core:version_info_gen failed: bash failed: error executing command
(cd /root/.cache/bazel/_bazel_root/b217a9ee56a7642547b5289e2480d3f7/execroot/org_tensorflow && \
exec env – \
LD_LIBRARY_PATH=/cm/shared/apps/python/2.7.12/lib:/cm/shared/apps/intel/parallelstudio/2017/compilers_and_libraries_2017/linux/mkl/lib/intel64/:/cm/shared/apps/intel/parallelstudio/2017/compilers_and_libraries_2017/linux/lib/intel64:/cm/shared/apps/gcc/4.8.2/lib:/cm/shared/apps/gcc/4.8.2/lib64:/cm/shared/apps/slurm/14.03.0/lib64/slurm:/cm/shared/apps/slurm/14.03.0/lib64 \
PATH=/cm/shared/apps/bazel/0.5.2:/cm/shared/apps/python/2.7.12/bin:/cm/shared/apps/gcc/4.8.2/bin:/cm/shared/apps/slurm/14.03.0/sbin:/cm/shared/apps/slurm/14.03.0/bin:/cm/local/apps/cluster-tools/bin:/cm/local/apps/cmd/sbin:/cm/local/apps/cmd/bin:/cm/shared/apps/cmgui:/usr/lib64/qt-3.3/bin:/usr/local/sbin:/usr/local/bin:/sbin:/bin:/usr/sbin:/usr/bin:/opt/ibutils/bin:/opt/dell/srvadmin/bin:/opt/dell/srvadmin/sbin:/root/bin:/opt/dell/mdstoragesoftware/mdstoragemanager/jre/bin \
/bin/bash -c ‘source external/bazel_tools/tools/genrule/genrule-setup.sh; tensorflow/tools/git/gen_git_source.py –generate tensorflow/tools/git/gen/spec.json tensorflow/tools/git/gen/head tensorflow/tools/git/gen/branch_ref “bazel-out/host/genfiles/tensorflow/core/util/version_info.cc”‘): com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.
Traceback (most recent call last):
File “tensorflow/tools/git/gen_git_source.py”, line 30, in
import argparse
ImportError: No module named argparse
Target //tensorflow/tools/pip_package:build_pip_package failed to build
INFO: Elapsed time: 52.002s, Critical Path: 17.53s

And i can confirm that argparse is indeed installed:
[root@cc-dclrilog51 tensorflow]# python2.7
Python 2.7.12 (default, Oct 25 2016, 10:39:13)
[GCC Intel(R) C++ gcc 4.4.7 mode] on linux2
Type “help”, “copyright”, “credits” or “license” for more information.
>>> import argparse
>>> print argparse.__version__
1.1
Does anyone have a clue how to fix this?
- Kuan-Yi Li
  
  July 10, 2017 at 10:49 Reply
  
  argparse is a standard library since Python 2.7 so it is very much likely your build tool is still using CentOS 6 system built-in Python, which is Python 2.6. One of the reasons I used Software Collection environment is that it not only installs Python 2.7 but also overrides system built-in Python when enabled.
  
  You may type which python or python -V to find out which version of Python is active before TensorFlow compilation.
  
  To further demonstrate the difference, below is what I got from my build server.
  
  [centos@tensorflow-centos6 ~]$ which python /usr/bin/python [centos@tensorflow-centos6 ~]$ python -V Python 2.6.6 [centos@tensorflow-centos6 ~]$ scl enable devtoolset-6 python27 bash [centos@tensorflow-centos6 ~]$ which python /opt/rh/python27/root/usr/bin/python [centos@tensorflow-centos6 ~]$ python -V Python 2.7.13
Luigi Cerulo

September 20, 2017 at 22:07 Reply

hi all, I obtained tensorflow working with GPU using devtoolset-4 instead of devtoolset-6.
Greg

December 2, 2017 at 00:44 Reply

I tried the conda route, but I still get the version ‘GLIBC_2.16’ not found error. It seems that build is still necessary for the C++11 compatible compiler side of things?
- Kuan-Yi Li
  
  December 11, 2017 at 04:57 Reply
  
  I took some time to test it out, the conda route worked fine. However, I did find a possible caveat: You should install tensorflow package only from default channel, NOT conda-forge channel. Run conda config --get channels to check whether you have any auxiliary channel enabled, and if so, priority matters.
Fleek IT Solutions

July 20, 2022 at 14:53 Reply

Thank you for the great article. It really helped me to TensorFlow on CentOS successfully