Skip to content

Instantly share code, notes, and snippets.

@jessebenson

jessebenson/building_tensorflow.md

Forked from kmhofmann/building_tensorflow.md
Created March 20, 2018 04:09
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save jessebenson/c8ab4735bb3698c1f0303c089fb0ad79 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save jessebenson/c8ab4735bb3698c1f0303c089fb0ad79 to your computer and use it in GitHub Desktop.
Download ZIP

Revisions

  1. @kmhofmann kmhofmann revised this gist Feb 7, 2018. 1 changed file with 8 additions and 4 deletions.
    12 changes: 8 additions & 4 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -105,15 +105,19 @@ necessary.

    * Leave out `--config=cuda` when not building with GPU support.

    * Due to [this issue](https://github.com/tensorflow/tensorflow/issues/15492), you might need to add the argument
    `--incompatible_load_argument_is_label=false` if you see an error similar to
    `name 'sycl_library_path' is not defined`.

    * TensorFlow 1.5: As [mentioned here](https://github.com/tensorflow/tensorflow/issues/16694), you might have to add
    `--action_env=LD_LIBRARY_PATH=${LD_LIBRARY_PATH}` to the call to Bazel, in case you encounter linking errors related
    to CUDA/cuDNN.

    * Add `-c dbg --strip=never` in case you do not want debug symbols to be stripped (e.g. for debugging purposes).
    Usually, you won't need to add this option.

    * Add `--compilation_mode=dbg` to build in debug instead of release mode, i.e. without optimizations.
    You shouldn't do this unless you really want to.

    * Due to [this issue](https://github.com/tensorflow/tensorflow/issues/15492), you might need to add the argument
    `--incompatible_load_argument_is_label=false` if you see an error similar to
    `name 'sycl_library_path' is not defined`.

    * We still need to build a Python package using the now generated build_pip_package script.

  2. @kmhofmann kmhofmann revised this gist Jan 26, 2018. 1 changed file with 28 additions and 11 deletions.
    39 changes: 28 additions & 11 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -4,7 +4,8 @@ The official instructions on building TensorFlow are here: https://www.tensorflo

    ## Prerequisites

    We are assuming a build with CUDA support, as well as including SIMD optimizations (SSE3, SSE4, AVX, AVX2, FMA), on a Debian-like system (e.g. Ubuntu Linux).
    We are assuming a build with CUDA support, as well as including SIMD optimizations (SSE3, SSE4, AVX, AVX2, FMA), on a
    Debian-like system (e.g. Ubuntu Linux).

    On new systems, one will have to install CUDA, CuDNN, plus the following dependencies:

    @@ -21,7 +22,8 @@ Good to know: The compute capabilities for

    ## Installing Bazel

    Bazel is Google's own build system, required to build TensorFlow. Building TensorFlow usually requires an up-to-date version of Bazel; there is a good chance that whatever your package manager provides will be outdated.
    Bazel is Google's own build system, required to build TensorFlow. Building TensorFlow usually requires an up-to-date
    version of Bazel; there is a good chance that whatever your package manager provides will be outdated.

    There are various ways to obtain a build of Bazel (see https://bazel.build/versions/master/docs/install-ubuntu.html).

    @@ -35,7 +37,10 @@ Recommended by Google, but you need to be comfortable adding another APT package

    ### Option 2: Building from source

    In case you prefer building from source, it's unfortunately not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.
    In case you prefer building from source, it's unfortunately not as easy as cloning the Git repository and typing `make`.
    Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains
    some already pre-generated files. With one such installation in place, one could build Bazel straight from the
    repository source, but that's probably not necessary.

    So we will go with building a distribution build, which is reasonably straightforward:

    @@ -44,15 +49,17 @@ So we will go with building a distribution build, which is reasonably straightfo
    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.9.0/bazel-0.9.0-dist.zip

    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in its own directory...
    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in
    its own directory...

    $ unzip bazel-0.9.0-dist.zip

    * Compile Bazel

    $ bash ./compile.sh

    * The output executable is now located in `output/bazel`. Add a `PATH` entry to your `.bashrc`, or just export it in your current shell:
    * The output executable is now located in `output/bazel`. Add a `PATH` entry to your `.bashrc`, or just export it in
    your current shell:

    $ export PATH=`pwd`/output:$PATH

    @@ -74,7 +81,8 @@ You should now be able to call the `bazel` executable from anywhere on your file

    $ deactivate

    * Clone the sources, and check out the desired branch. At the time of writing, 1.5.0 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.5.0 was the latest version; adjust if
    necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    @@ -95,11 +103,17 @@ You should now be able to call the `bazel` executable from anywhere on your file

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

    Leave out `--config=cuda` when not building with GPU support.
    * Leave out `--config=cuda` when not building with GPU support.

    Add `-c dbg --strip=never` in case you do not want debug symbols to be stripped (e.g. for debugging purposes). Usually, you won't need to add this option.
    * Add `-c dbg --strip=never` in case you do not want debug symbols to be stripped (e.g. for debugging purposes).
    Usually, you won't need to add this option.

    Add `--compilation_mode=dbg` to build in debug instead of release mode, i.e. without optimizations. You shouldn't do this unless you really want to.
    * Add `--compilation_mode=dbg` to build in debug instead of release mode, i.e. without optimizations.
    You shouldn't do this unless you really want to.

    * Due to [this issue](https://github.com/tensorflow/tensorflow/issues/15492), you might need to add the argument
    `--incompatible_load_argument_is_label=false` if you see an error similar to
    `name 'sycl_library_path' is not defined`.

    * We still need to build a Python package using the now generated build_pip_package script.

    @@ -115,9 +129,12 @@ That should be it!

    ### Notes on building with unsupported system

    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA 8 (still) requires a GCC compiler from the GCC 5.x series. In this case:
    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS
    Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA 8 (still) requires
    a GCC compiler from the GCC 5.x series. In this case:

    * Build GCC 5.x from source and install, preferably user-local. (https://github.com/kmhofmann/build_stuff provides a script to build various versions of GCC from source.)
    * Build GCC 5.x from source and install, preferably user-local. (https://github.com/kmhofmann/build_stuff provides
    a script to build various versions of GCC from source.)

    Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:

  3. @kmhofmann kmhofmann revised this gist Jan 26, 2018. 1 changed file with 3 additions and 3 deletions.
    6 changes: 3 additions & 3 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -74,11 +74,11 @@ You should now be able to call the `bazel` executable from anywhere on your file

    $ deactivate

    * Clone the sources, and check out the desired branch. At the time of writing, 1.4.1 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.5.0 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout v1.4.1
    $ git checkout v1.5.0

    * Run the configuration script

    @@ -107,7 +107,7 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.

    $ pip install /tmp/tensorflow_pkg/tensorflow-1.4.1-cp35-cp35m-linux_x86_64.whl
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.5.0-cp35-cp35m-linux_x86_64.whl

    (Adjust the file name, if necessary.)

  4. @kmhofmann kmhofmann revised this gist Dec 21, 2017. 1 changed file with 3 additions and 3 deletions.
    6 changes: 3 additions & 3 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -39,14 +39,14 @@ In case you prefer building from source, it's unfortunately not as easy as cloni

    So we will go with building a distribution build, which is reasonably straightforward:

    * Download a distribution package from the releases page. The current version at the time of writing was 0.7.0.
    * Download a distribution package from the releases page. The current version at the time of writing was 0.9.0.

    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.7.0/bazel-0.7.0-dist.zip
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.9.0/bazel-0.9.0-dist.zip

    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in its own directory...

    $ unzip bazel-0.7.0-dist.zip
    $ unzip bazel-0.9.0-dist.zip

    * Compile Bazel

  5. @kmhofmann kmhofmann revised this gist Dec 15, 2017. 1 changed file with 32 additions and 18 deletions.
    50 changes: 32 additions & 18 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -4,18 +4,21 @@ The official instructions on building TensorFlow are here: https://www.tensorflo

    ## Prerequisites

    We are assuming a build with CUDA support, as well as including SIMD optimizations (SSE3, SSE4, AVX, AVX2, FMA).

    Good to know: The compute capabilities for
    * Maxwell TITAN X: `5.2`
    * Pascal TITAN X (2016): `6.1`
    We are assuming a build with CUDA support, as well as including SIMD optimizations (SSE3, SSE4, AVX, AVX2, FMA), on a Debian-like system (e.g. Ubuntu Linux).

    On new systems, one will have to install CUDA, CuDNN, plus the following dependencies:

    $ sudo apt-get install python3-numpy python3-dev python3-pip python3-wheel libcupti-dev

    (Leave out `libcupti-dev` when not building with GPU support.)

    Good to know: The compute capabilities for
    * Maxwell TITAN X: `5.2`
    * Pascal TITAN X (2016): `6.1`
    * GeForce GTX 1080 Ti: `6.1`

    (See [here](https://developer.nvidia.com/cuda-gpus) for the full list.)

    ## Installing Bazel

    Bazel is Google's own build system, required to build TensorFlow. Building TensorFlow usually requires an up-to-date version of Bazel; there is a good chance that whatever your package manager provides will be outdated.
    @@ -57,24 +60,20 @@ You should now be able to call the `bazel` executable from anywhere on your file

    ## Installing TensorFlow

    ### Notes on building with unsupported system
    ### Building TensorFlow

    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA 8 (still) requires a GCC compiler from the GCC 5.x series. In this case:
    * Create a Python 3 virtualenv, if you have not done this yet. For example:

    * Build GCC 5.x from source and install, preferably user-local. (https://github.com/kmhofmann/build_stuff provides a script to build various versions of GCC from source.)
    $ virtualenv -p python3 --system-site-packages ~/.virtualenvs/tf_dev

    Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:

    $ export PATH=$HOME/local/bin:$PATH
    $ export LD_LIBRARY_PATH=$HOME/local/lib64:$LD_LIBRARY_PATH

    Ensure that `gcc --version` is the desired version.
    * Activate your respective Python 3 based virtual environment.

    $ source ~/.virtualenvs/tf_dev/bin/activate
    * Proceed with building, as described below.
    * This can later be deactivated with

    ### Building TensorFlow
    $ deactivate

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.4.1 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    @@ -112,4 +111,19 @@ You should now be able to call the `bazel` executable from anywhere on your file

    (Adjust the file name, if necessary.)

    That should be it!
    That should be it!

    ### Notes on building with unsupported system

    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA 8 (still) requires a GCC compiler from the GCC 5.x series. In this case:

    * Build GCC 5.x from source and install, preferably user-local. (https://github.com/kmhofmann/build_stuff provides a script to build various versions of GCC from source.)

    Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:

    $ export PATH=$HOME/local/bin:$PATH
    $ export LD_LIBRARY_PATH=$HOME/local/lib64:$LD_LIBRARY_PATH

    Ensure that `gcc --version` is the desired version.

    * Proceed with building, as described below.
  6. @kmhofmann kmhofmann revised this gist Dec 15, 2017. 1 changed file with 6 additions and 4 deletions.
    10 changes: 6 additions & 4 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -75,11 +75,11 @@ You should now be able to call the `bazel` executable from anywhere on your file
    ### Building TensorFlow

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.4.0 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.4.1 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout v1.4.0
    $ git checkout v1.4.1

    * Run the configuration script

    @@ -98,15 +98,17 @@ You should now be able to call the `bazel` executable from anywhere on your file

    Leave out `--config=cuda` when not building with GPU support.

    Add `-c dbg --strip=never` in case you do not want debug symbols to be stripped (e.g. for debugging purposes). (Usually, you won't need to add this option.)
    Add `-c dbg --strip=never` in case you do not want debug symbols to be stripped (e.g. for debugging purposes). Usually, you won't need to add this option.

    Add `--compilation_mode=dbg` to build in debug instead of release mode, i.e. without optimizations. You shouldn't do this unless you really want to.

    * We still need to build a Python package using the now generated build_pip_package script.

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

    * Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.

    $ pip install /tmp/tensorflow_pkg/tensorflow-1.4.0-cp35-cp35m-linux_x86_64.whl
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.4.1-cp35-cp35m-linux_x86_64.whl

    (Adjust the file name, if necessary.)

  7. @kmhofmann kmhofmann revised this gist Dec 14, 2017. 1 changed file with 3 additions and 1 deletion.
    4 changes: 3 additions & 1 deletion building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -96,7 +96,9 @@ You should now be able to call the `bazel` executable from anywhere on your file

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

    (Leave out `--config=cuda` when not building with GPU support.)
    Leave out `--config=cuda` when not building with GPU support.

    Add `-c dbg --strip=never` in case you do not want debug symbols to be stripped (e.g. for debugging purposes). (Usually, you won't need to add this option.)

    * We still need to build a Python package using the now generated build_pip_package script.

  8. @kmhofmann kmhofmann revised this gist Nov 7, 2017. 1 changed file with 9 additions and 9 deletions.
    18 changes: 9 additions & 9 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -36,14 +36,14 @@ In case you prefer building from source, it's unfortunately not as easy as cloni

    So we will go with building a distribution build, which is reasonably straightforward:

    * Download a distribution package from the releases page. The current version at the time of writing was 0.5.3.
    * Download a distribution package from the releases page. The current version at the time of writing was 0.7.0.

    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.5.3/bazel-0.5.3-dist.zip
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.7.0/bazel-0.7.0-dist.zip

    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in its own directory...

    $ unzip bazel-0.5.3-dist.zip
    $ unzip bazel-0.7.0-dist.zip

    * Compile Bazel

    @@ -70,16 +70,16 @@ You should now be able to call the `bazel` executable from anywhere on your file

    Ensure that `gcc --version` is the desired version.

    * Proceed with building, as described below. In case of errors, you may need to apply [#9482](https://github.com/tensorflow/tensorflow/pull/9482), either by checking out a version >1.1.0 below (once released), building straight from the `master` branch, or applying the patch manually (it's small).
    * Proceed with building, as described below.

    ### Building TensorFlow

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.3.0 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.4.0 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout v1.3.0
    $ git checkout v1.4.0

    * Run the configuration script

    @@ -92,9 +92,9 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Compile TensorFlow using Bazel.

    I got the special build options from here: http://stackoverflow.com/questions/41293077/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions
    This command will build Tensorflow using optimized settings for the current machine architecture.

    $ bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
    $ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

    (Leave out `--config=cuda` when not building with GPU support.)

    @@ -104,7 +104,7 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.

    $ pip install /tmp/tensorflow_pkg/tensorflow-1.3.0-cp35-cp35m-linux_x86_64.whl
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.4.0-cp35-cp35m-linux_x86_64.whl

    (Adjust the file name, if necessary.)

  9. @kmhofmann kmhofmann revised this gist Aug 17, 2017. 1 changed file with 6 additions and 6 deletions.
    12 changes: 6 additions & 6 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -36,14 +36,14 @@ In case you prefer building from source, it's unfortunately not as easy as cloni

    So we will go with building a distribution build, which is reasonably straightforward:

    * Download a distribution package from the releases page. The current version at the time of writing was 0.5.1.
    * Download a distribution package from the releases page. The current version at the time of writing was 0.5.3.

    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.5.1/bazel-0.5.1-dist.zip
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.5.3/bazel-0.5.3-dist.zip

    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in its own directory...

    $ unzip bazel-0.5.1-dist.zip
    $ unzip bazel-0.5.3-dist.zip

    * Compile Bazel

    @@ -75,11 +75,11 @@ You should now be able to call the `bazel` executable from anywhere on your file
    ### Building TensorFlow

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.2.0 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.3.0 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout v1.2.0
    $ git checkout v1.3.0

    * Run the configuration script

    @@ -104,7 +104,7 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.

    $ pip install /tmp/tensorflow_pkg/tensorflow-1.2.0-cp35-cp35m-linux_x86_64.whl
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.3.0-cp35-cp35m-linux_x86_64.whl

    (Adjust the file name, if necessary.)

  10. @kmhofmann kmhofmann revised this gist Aug 17, 2017. 1 changed file with 3 additions and 1 deletion.
    4 changes: 3 additions & 1 deletion building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -61,7 +61,9 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA 8 (still) requires a GCC compiler from the GCC 5.x series. In this case:

    * Build GCC 5.x from source and install, preferably user-local. Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:
    * Build GCC 5.x from source and install, preferably user-local. (https://github.com/kmhofmann/build_stuff provides a script to build various versions of GCC from source.)

    Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:

    $ export PATH=$HOME/local/bin:$PATH
    $ export LD_LIBRARY_PATH=$HOME/local/lib64:$LD_LIBRARY_PATH
  11. @kmhofmann kmhofmann revised this gist Jun 16, 2017. 1 changed file with 9 additions and 7 deletions.
    16 changes: 9 additions & 7 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -24,24 +24,26 @@ There are various ways to obtain a build of Bazel (see https://bazel.build/versi

    ### Option 1: Using the Bazel APT repository

    Recommended by Google, but you need to be comfortable adding another APT package souce.

    $ echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee /etc/apt/sources.list.d/bazel.list
    $ curl https://bazel.build/bazel-release.pub.gpg | sudo apt-key add -
    $ sudo apt-get update && sudo apt-get install bazel

    ### Option 2: Building from source

    I usually prefer building from source, but it's not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.
    In case you prefer building from source, it's unfortunately not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.

    So we will go with building a distribution build, which is reasonably straightforward:

    * Download a distribution package from the releases page. The current version at the time of writing was 0.4.5.
    * Download a distribution package from the releases page. The current version at the time of writing was 0.5.1.

    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.4.5/bazel-0.4.5-dist.zip
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.5.1/bazel-0.5.1-dist.zip

    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in its own directory...

    $ unzip bazel-0.4.5-dist.zip
    $ unzip bazel-0.5.1-dist.zip

    * Compile Bazel

    @@ -71,11 +73,11 @@ You should now be able to call the `bazel` executable from anywhere on your file
    ### Building TensorFlow

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.1.0 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.2.0 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout r1.1
    $ git checkout v1.2.0

    * Run the configuration script

    @@ -100,7 +102,7 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.

    $ pip install /tmp/tensorflow_pkg/tensorflow-1.1.0-cp35-cp35m-linux_x86_64.whl
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.2.0-cp35-cp35m-linux_x86_64.whl

    (Adjust the file name, if necessary.)

  12. @kmhofmann kmhofmann revised this gist May 2, 2017. 1 changed file with 2 additions and 2 deletions.
    4 changes: 2 additions & 2 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -61,8 +61,8 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Build GCC 5.x from source and install, preferably user-local. Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:

    $ export PATH=%HOME/local/bin:$PATH
    $ export LD_LIBRARY_PATH=%HOME/local/lib64:$LD_LIBRARY_PATH
    $ export PATH=$HOME/local/bin:$PATH
    $ export LD_LIBRARY_PATH=$HOME/local/lib64:$LD_LIBRARY_PATH

    Ensure that `gcc --version` is the desired version.

  13. @kmhofmann kmhofmann revised this gist May 2, 2017. 1 changed file with 2 additions and 2 deletions.
    4 changes: 2 additions & 2 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -57,7 +57,7 @@ You should now be able to call the `bazel` executable from anywhere on your file

    ### Notes on building with unsupported system

    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA requires a GCC compiler from the GCC 5.x series. In this case:
    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA 8 (still) requires a GCC compiler from the GCC 5.x series. In this case:

    * Build GCC 5.x from source and install, preferably user-local. Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:

    @@ -71,7 +71,7 @@ You should now be able to call the `bazel` executable from anywhere on your file
    ### Building TensorFlow

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.1 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.1.0 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
  14. @kmhofmann kmhofmann revised this gist May 2, 2017. 1 changed file with 15 additions and 0 deletions.
    15 changes: 15 additions & 0 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -55,6 +55,21 @@ You should now be able to call the `bazel` executable from anywhere on your file

    ## Installing TensorFlow

    ### Notes on building with unsupported system

    * Additional steps may be necessary when building TensorFlow on an officially unsupported system (e.g. on a non-LTS Ubuntu version). For example, Ubuntu versions >16.04 ship with GCC 6 as default compiler, but CUDA requires a GCC compiler from the GCC 5.x series. In this case:

    * Build GCC 5.x from source and install, preferably user-local. Add the respective paths to the `PATH` and `LD_LIBRARY_PATH` variables, e.g.:

    $ export PATH=%HOME/local/bin:$PATH
    $ export LD_LIBRARY_PATH=%HOME/local/lib64:$LD_LIBRARY_PATH

    Ensure that `gcc --version` is the desired version.

    * Proceed with building, as described below. In case of errors, you may need to apply [#9482](https://github.com/tensorflow/tensorflow/pull/9482), either by checking out a version >1.1.0 below (once released), building straight from the `master` branch, or applying the patch manually (it's small).

    ### Building TensorFlow

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.1 was the latest version; adjust if necessary.

  15. @kmhofmann kmhofmann revised this gist May 1, 2017. 1 changed file with 3 additions and 3 deletions.
    6 changes: 3 additions & 3 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -56,11 +56,11 @@ You should now be able to call the `bazel` executable from anywhere on your file
    ## Installing TensorFlow

    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.0 was the latest version; adjust if necessary.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.1 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout r1.0
    $ git checkout r1.1

    * Run the configuration script

    @@ -85,7 +85,7 @@ You should now be able to call the `bazel` executable from anywhere on your file

    * Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.

    $ pip install /tmp/tensorflow_pkg/tensorflow-1.0.0-cp35-cp35m-linux_x86_64.whl
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.1.0-cp35-cp35m-linux_x86_64.whl

    (Adjust the file name, if necessary.)

  16. @kmhofmann kmhofmann revised this gist May 1, 2017. 1 changed file with 10 additions and 2 deletions.
    12 changes: 10 additions & 2 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -20,9 +20,17 @@ On new systems, one will have to install CUDA, CuDNN, plus the following depende

    Bazel is Google's own build system, required to build TensorFlow. Building TensorFlow usually requires an up-to-date version of Bazel; there is a good chance that whatever your package manager provides will be outdated.

    There are various ways to obtain a build of Bazel described here. One of the options is installing it via custom APT repository (see https://bazel.build/versions/master/docs/install-ubuntu.html), but I personally don't like adding all kinds of APT sources.
    There are various ways to obtain a build of Bazel (see https://bazel.build/versions/master/docs/install-ubuntu.html).

    I prefer building from source, but it's not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.
    ### Option 1: Using the Bazel APT repository

    $ echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee /etc/apt/sources.list.d/bazel.list
    $ curl https://bazel.build/bazel-release.pub.gpg | sudo apt-key add -
    $ sudo apt-get update && sudo apt-get install bazel

    ### Option 2: Building from source

    I usually prefer building from source, but it's not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.

    So we will go with building a distribution build, which is reasonably straightforward:

  17. @kmhofmann kmhofmann revised this gist May 1, 2017. 1 changed file with 1 addition and 2 deletions.
    3 changes: 1 addition & 2 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -60,9 +60,8 @@ You should now be able to call the `bazel` executable from anywhere on your file

    You can leave most defaults, but do specify the following (or similar):

    jemalloc -> Y
    CUDA support -> Y
    CUDA compute capability -> 5.1,6.1
    CUDA compute capability -> 5.2,6.1

    * Compile TensorFlow using Bazel.

  18. @kmhofmann kmhofmann revised this gist Apr 16, 2017. 1 changed file with 3 additions and 3 deletions.
    6 changes: 3 additions & 3 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -26,14 +26,14 @@ I prefer building from source, but it's not as easy as cloning the Git repositor

    So we will go with building a distribution build, which is reasonably straightforward:

    * Download a distribution package from the releases page. The current version at the time of writing was 0.4.4.
    * Download a distribution package from the releases page. The current version at the time of writing was 0.4.5.

    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.4.4/bazel-0.4.4-dist.zip
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.4.5/bazel-0.4.5-dist.zip

    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in its own directory...

    $ unzip bazel-0.4.4-dist.zip
    $ unzip bazel-0.4.5-dist.zip

    * Compile Bazel

  19. @kmhofmann kmhofmann revised this gist Apr 16, 2017. 1 changed file with 1 addition and 1 deletion.
    2 changes: 1 addition & 1 deletion building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -20,7 +20,7 @@ On new systems, one will have to install CUDA, CuDNN, plus the following depende

    Bazel is Google's own build system, required to build TensorFlow. Building TensorFlow usually requires an up-to-date version of Bazel; there is a good chance that whatever your package manager provides will be outdated.

    There are various ways to obtain a build of Bazel described here. One of the options is installing it via custom APT repository, but I personally don't like adding all kinds of APT sources.
    There are various ways to obtain a build of Bazel described here. One of the options is installing it via custom APT repository (see https://bazel.build/versions/master/docs/install-ubuntu.html), but I personally don't like adding all kinds of APT sources.

    I prefer building from source, but it's not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.

  20. @kmhofmann kmhofmann revised this gist Apr 10, 2017. 1 changed file with 62 additions and 37 deletions.
    99 changes: 62 additions & 37 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -4,57 +4,82 @@ The official instructions on building TensorFlow are here: https://www.tensorflo

    ## Prerequisites

    We are assuming a build with CUDA support, and compute capabilities 5.2 (Maxwell TITAN X) and 6.1 (Pascal TITAN X),
    as well as including SIMD optimizations (SSE3, SSE4, AVX, AVX2, FMA).
    We are assuming a build with CUDA support, as well as including SIMD optimizations (SSE3, SSE4, AVX, AVX2, FMA).

    Good to know:
    The compute capabilities for
    Good to know: The compute capabilities for
    * Maxwell TITAN X: `5.2`
    * Pascal TITAN X (2016): `6.1`

    On new systems, one will have to install CUDA, CuDNN, plus the following dependencies:
    $ sudo apt-get install python3-numpy python3-dev python3-pip python3-wheel libcupti-dev
    (Leave out libcupti-dev when not building with GPU support.)

    $ sudo apt-get install python3-numpy python3-dev python3-pip python3-wheel libcupti-dev

    (Leave out `libcupti-dev` when not building with GPU support.)

    ## Installing Bazel

    Bazel is Google's own build system, required to build TensorFlow. Building TensorFlow usually requires an up-to-date version of Bazel; there is a good chance that whatever your package manager provides will be outdated.

    There are various ways to obtain a build of Bazel described here. One of the options is installing it via custom APT repository, but I (Michael) personally don't like adding all kinds of APT sources.
    There are various ways to obtain a build of Bazel described here. One of the options is installing it via custom APT repository, but I personally don't like adding all kinds of APT sources.

    I prefer building from source, but it's not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be built with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.

    So we will go with building a distribution build, which is reasonably straightforward:

    * Download a distribution package from the releases page. The current version at the time of writing was 0.4.4.

    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.4.4/bazel-0.4.4-dist.zip

    I prefer building from source, but it's not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be build with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.
    * Unzip the sources. This being a zip file, the files are stored without containing folder. Glad we already put it in its own directory...

    So we will go with building a distribution build, which is straightforward:
    $ unzip bazel-0.4.4-dist.zip

    Download a distribution package from the releases page. The current version at the time of writing was 0.4.4.
    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.4.4/bazel-0.4.4-dist.zip
    Unzip the sources. This being a zip file, the files are stored without containing folder. (sad) Glad we put it in its own directory...
    $ unzip bazel-0.4.4-dist.zip
    Compile Bazel
    $ bash ./compile.sh
    The output executable is now located in output/bazel. Add a PATH entry to your .bashrc, or just export it in your current shell:
    $ export PATH=`pwd`/output:$PATH
    You should now be able to call the bazel executable from anywhere on your filesystem.
    * Compile Bazel

    $ bash ./compile.sh

    * The output executable is now located in `output/bazel`. Add a `PATH` entry to your `.bashrc`, or just export it in your current shell:

    $ export PATH=`pwd`/output:$PATH

    You should now be able to call the `bazel` executable from anywhere on your filesystem.

    ## Installing TensorFlow

    Activate your respective Python 3 based virtual environment!
    Clone the sources, and check out the desired branch. At the time of writing, 1.0 was the latest version.
    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout r1.0
    Run the configuration script
    $ ./configure
    You can leave most defaults, but do specify the following:
    jemalloc -> Y
    CUDA support -> Y
    CUDA compute capability -> 5.1,6.1
    Compile TensorFlow using Bazel. I got the special build options from here.
    $ bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
    (Leave out --config=cuda when not building with GPU support.)
    We still need to build a Python package using the now generated build_pip_package script.
    $ bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
    Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.0.0-cp35-cp35m-linux_x86_64.whl
    * Activate your respective Python 3 based virtual environment.
    * Clone the sources, and check out the desired branch. At the time of writing, 1.0 was the latest version; adjust if necessary.

    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout r1.0

    * Run the configuration script

    $ ./configure

    You can leave most defaults, but do specify the following (or similar):

    jemalloc -> Y
    CUDA support -> Y
    CUDA compute capability -> 5.1,6.1

    * Compile TensorFlow using Bazel.

    I got the special build options from here: http://stackoverflow.com/questions/41293077/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions

    $ bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

    (Leave out `--config=cuda` when not building with GPU support.)

    * We still need to build a Python package using the now generated build_pip_package script.

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

    * Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.

    $ pip install /tmp/tensorflow_pkg/tensorflow-1.0.0-cp35-cp35m-linux_x86_64.whl

    (Adjust the file name, if necessary.)

    That should be it!
  21. @kmhofmann kmhofmann created this gist Apr 10, 2017.
    60 changes: 60 additions & 0 deletions building_tensorflow.md
    View file
    Original file line number Diff line number Diff line change
    @@ -0,0 +1,60 @@
    # Building TensorFlow from source

    The official instructions on building TensorFlow are here: https://www.tensorflow.org/install/install_sources

    ## Prerequisites

    We are assuming a build with CUDA support, and compute capabilities 5.2 (Maxwell TITAN X) and 6.1 (Pascal TITAN X),
    as well as including SIMD optimizations (SSE3, SSE4, AVX, AVX2, FMA).

    Good to know:
    The compute capabilities for

    On new systems, one will have to install CUDA, CuDNN, plus the following dependencies:
    $ sudo apt-get install python3-numpy python3-dev python3-pip python3-wheel libcupti-dev
    (Leave out libcupti-dev when not building with GPU support.)

    ## Installing Bazel

    Bazel is Google's own build system, required to build TensorFlow. Building TensorFlow usually requires an up-to-date version of Bazel; there is a good chance that whatever your package manager provides will be outdated.

    There are various ways to obtain a build of Bazel described here. One of the options is installing it via custom APT repository, but I (Michael) personally don't like adding all kinds of APT sources.

    I prefer building from source, but it's not as easy as cloning the Git repository and typing `make`. Recent versions of Bazel can only be build with Bazel, unless one downloads a distribution source build, which contains some already pre-generated files. With one such installation in place, one could build Bazel straight from the repository source, but that's probably not necessary.

    So we will go with building a distribution build, which is straightforward:

    Download a distribution package from the releases page. The current version at the time of writing was 0.4.4.
    $ mkdir bazel && cd bazel
    $ wget https://github.com/bazelbuild/bazel/releases/download/0.4.4/bazel-0.4.4-dist.zip
    Unzip the sources. This being a zip file, the files are stored without containing folder. (sad) Glad we put it in its own directory...
    $ unzip bazel-0.4.4-dist.zip
    Compile Bazel
    $ bash ./compile.sh
    The output executable is now located in output/bazel. Add a PATH entry to your .bashrc, or just export it in your current shell:
    $ export PATH=`pwd`/output:$PATH
    You should now be able to call the bazel executable from anywhere on your filesystem.

    ## Installing TensorFlow

    Activate your respective Python 3 based virtual environment!
    Clone the sources, and check out the desired branch. At the time of writing, 1.0 was the latest version.
    $ git clone https://github.com/tensorflow/tensorflow
    $ cd tensorflow
    $ git checkout r1.0
    Run the configuration script
    $ ./configure
    You can leave most defaults, but do specify the following:
    jemalloc -> Y
    CUDA support -> Y
    CUDA compute capability -> 5.1,6.1
    Compile TensorFlow using Bazel. I got the special build options from here.
    $ bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
    (Leave out --config=cuda when not building with GPU support.)
    We still need to build a Python package using the now generated build_pip_package script.
    $ bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
    Now there will be a package inside /tmp/tensorflow_pkg, which can be installed with pip.
    $ pip install /tmp/tensorflow_pkg/tensorflow-1.0.0-cp35-cp35m-linux_x86_64.whl
    (Adjust the file name, if necessary.)

    That should be it!