Installation Guide¶

Note

Since Tensor Comprehensions is still a research project in its infancy it is recommended to build the master from source.

Conda installation¶

You can install the latest released package for tensor comprehensions as follows:

conda install -y -c pytorch -c tensorcomp tensor_comprehensions

Build from source¶

Note

In order to uniformize and simplify the build system we had to make choices. TC is currently only officially supported on Ubuntu 16.04 with gcc 5.4.0 and takes most its dependencies from conda packages. Other configurations may work too but are not yet officially supported. For more information about the setup we use to build the conda dependencies see the following Dockerfile. Our main goal with this decision is to make the build procedure extremely simple, both reproducible internally and extensible to new targets in the future. In particular the gcc-4 / gcc-5 ABI switch is not something we want to concern ourselves with at this point, we go for gcc-5.4.0.

Prerequisites¶

Building TC from source requires gmp, cmake (v3.10 or higher), automake and libtool. It is generally a good idea to look at the Dockerfile which shows the environment we use for testing and building our packages. To install on Ubuntu16.04:

sudo apt-get install libgmp3-dev cmake automake libtool

Conda from scratch (first time configuration)¶

Choose and set an INSTALLATION_PATH then run the following:

wget https://repo.anaconda.com/archive/Anaconda3-5.1.0-Linux-x86_64.sh -O anaconda.sh && \
chmod +x anaconda.sh && \
./anaconda.sh -b -p ${INSTALLATION_PATH} && \
rm anaconda.sh

. ${INSTALLATION_PATH}/bin/activate
conda update -y -n base conda

Create a new environment in which TC will be built and install core dependencies:

conda create -y --name tc_build python=3.6
conda activate tc_build
conda install -y pyyaml mkl-include pytest
conda install -y -c nicolasvasilache llvm-trunk halide

Then install the PyTorch 0.4.0 version that corresponds to your system binaries (e.g. cuda 9.0):

conda install -y -c pytorch pytorch=0.4.0 torchvision cuda90
conda remove -y cudatoolkit --force
Note

As of PyTorch 0.4, PyTorch links cuda libraries dynamically and it pulls cudatoolkit. However cudatoolkit can never replace a system installation because it cannot package libcuda.so (which comes with the driver, not the toolkit). As a consequence cudatoolkit only contains redundant libraries and we remove it explicitly. In a near future, the unified PyTorch + Caffe2 build system will link everything statically and stop pulling the cudatoolkit dependency.

Note

As of PyTorch 0.4.1, the ATen library is no shipped in libATen.so anymore. It is now instead part of libcaffe2.so. In shipping binaries, the PyTorch team opted to use the older GCC 4.9.2 C++ ABI. TC only supports the newer ABI. As a consequence TC must link with the libraries provided by the pytorch=0.4.0 package.

Activate conda in your current terminal¶

Once the first time configuration above has been completed, one should activate conda in each new terminal window explicitly (it is discouraged to add this to your .bashrc or equivalent)

. ${INSTALLATION_PATH}/bin/activate
conda activate tc_build

Build TC with dependencies supplied by conda¶

git clone http://www.github.com/facebookresearch/TensorComprehensions --recursive
cd TensorComprehensions
CLANG_PREFIX=$(${CONDA_PREFIX}/bin/llvm-config --prefix) ./build.sh

You may need to pass the environment variable CUDA_TOOLKIT_ROOT_DIR pointing to your cuda installation (this is required for FindCUDA.cmake to find your cuda installation and can be omitted on most systems). When required, passing CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda is generally sufficient.

Test locally¶

Run C++ tests:

./test.sh

Install the TC Python package locally to /tmp for smoke checking:

python setup.py install --prefix=/tmp
export PYTHONPATH=${PYTHONPATH}:$(find /tmp/lib -name site-packages)

Run Python smoke checks:

python -c 'import torch'
python -c 'import tensor_comprehensions'

Run Python tests:

python ./test_python/test_tc.py -v

At this point, if things work as expected you can venture installing as follows (it is always a good idea to record installed files for easy removal):

python setup.py install --record tc_files.txt

Advanced / development mode installation¶

Optional dependencies¶

Optionally if you want to use Caffe2 (this is necessary for building the C++ benchmarks since Caffe2 is our baseline):

conda install -y -c conda-forge eigen
conda install -y -c nicolasvasilache caffe2

Cudnn version 7.1 in Caffe2 / dev mode¶

Note

As of PyTorch 0.4, we need to package our own Caffe2. The curent PyTorch + Caffe2 build system links cudnn dynamically. The version of cudnn that is linked dynamically is imposed on us by the docker image supported by NVIDIA (see Dockerfile). For now this cudnn version is cudnn 7.1.

If for some reason, one cannot install cudnn 7.1 system-wide, one may resort to the following:

conda install -c anaconda cudnn
conda remove -y cudatoolkit --force
Note

cudnn pulls a cudatoolkit dependency but this can never replace a system installation because it cannot package libcuda.so (which comes with the driver, not the toolkit). As a consequence cudatoolkit only contains redundant libraries and we remove it explicitly. In a near future, the unified PyTorch + Caffe2 build system will link everything statically and we will not need to worry about cudnn anymore.