Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
Installing MaTEx TensorFlow GPU
Clone this wiki locally
We have provided a set of scripts for easy installation on GPU clusters. We expect that minimal changes would be required to the script for installation.
For using GPU enabled TensorFlow, you would need to set up the CUDNN_HOME, CUDA_HOME, and MPI_HOME environment variables to point where the CUDNN library, the CUDA SDK, and MPI home directory should be found.
As an example, for bash shells:
$ export CUDA_HOME=/where/nvcc/resides $ export CUDNN_HOME=/where/cudnn/resides $ export MPI_HOME=/where/mpi_library/resides
Once the environment variables have been setup:
Installation for bash shells
$ cd matex/src/deeplearning/tensorflow/gpu/py3.x $ source ./install_mpi_tf.sh
Installation for C-shells
$ cd matex/src/deeplearning/tensorflow/gpu/py3.x $ source ./install_mpi_tf.csh
You will be in a virtual python environment that encapsulates the TensorFlow changes (Your shell prompt should look differently).
Note for DGX-1 and CUDA 8 users
The default GPU wheel uses CUDA 7.5. However, we have provided an option to install the wheel that supports CUDA version 8 and CUDNN version 5.1. The parameter "8" should be provided as shown in the example below:
$ cd gpu/py3.x $ source ./install_mpi_tf.csh 8
Alternatively, the following script can also be used for CUDA 8 and CUDNN version 5.1:
$ cd gpu/py3.x $ source ./install_mpi_tf_cuda8_0.csh 8