What are the major changes for GPU programs?

On CentOS 7, the default CUDA library is version 10.1 with cuDNN v7.5.1; there is no separate cuDNN module. The cuda/8.0 (with no associated cuDNN library), cuda/9.0 with cuDNN v7.4.1, cuda/9.2 with cuDNN v7.5.1 and cuda/10.0 with cuDNN v7.5.1 modules are available as required by some deep-learning packages such as TensorFlow and Pytorch.

The following deep-learning libraries are available under CentOS 7:

  • darknet/436
  • netket/1.0.4
  • pytorch/1.0
  • pytorch/1.1
  • tensorflow/1.13.1
  • theano/1.0.2

The torch and torch7 modules have been deprecated and are no longer available. If you need to access torch or torch7, please contact us at help@scc.bu.edu and we will help.

The following GPU enabled chemistry software packages are available with with upgraded CUDA libraries:

  • gromacs/2018.3
  • gromacs/2019.1
  • hoomd/2.4.1
  • namd/2.13_mc-cuda
  • openmm/7.2.2

Troubleshooting

To load the cuda/10.1 module:

scc1% module load cuda/10.1

If your jobs were using GPU enabled software under CentOS 6, please check out the new versions of the software and the dependencies under CentOS 7 by doing the following steps.

Try to load the new versions of the software packages you use. Here is an example trying to load pytorch:

scc1% module load pytorch/1.1
-------------------------------------------------------------------------------
ERROR: pytorch/1.1 requires several additional modules. Run the following
commands to load all of the dependencies (including this module):

  module load python3/3.7.0
  module load cuda/10.1
  module load pytorch/1.1
-------------------------------------------------------------------------------

Check out and load the new dependencies according to the message on the terminal:

scc1% module load python3/3.7.0
scc1% module load cuda/10.1
scc1% module load pytorch/1.1