HPC fix April 2016 - Githubissues

AndreasMadsen / my-setup

Guides for myself and other, on how some of my stuff is configured.

6 stars 5 forks source link

HPC fix April 2016 #3

Closed AndreasMadsen closed 8 years ago

AndreasMadsen commented 8 years ago

use k40
parallelize all make calls
upgrade to qt5 (changes was required anyway)
upgrade pyopencl (requires new ctags and gcc 4.8.2+)
upgrade libgpuarray (requires new cmake and cuda 6.5+)
upgrade to cuda 7.0
upgrade theano, lasange
upgrade netcdf4

Fixes: #2

AndreasMadsen commented 8 years ago

The theano documentation says that is is only necessary to set LD_LIBRARY_PATH etc.. But it also says that it will automatically detect cudnn support by looking in the CUDA_PATH it may be enough to force theano to use cudnn by using a .theanorc flag if the other environment are set correctly.

AndreasMadsen commented 8 years ago

There might be an issue with using k40_interactive, but the HPC queue is unstable at the moment so I can't really tell.

AndreasMadsen commented 8 years ago

Adding cuDNN support this was looks to be possible, however I got the following error:

RuntimeError: You enabled cuDNN, but we aren't able to use it: Version is too old. Update to v5, was 3007.

Note I had to add:

module load cudnn
export CPATH=/appl/cudnn/v3.0-prod/include:$CPATH
export LIBRARY_PATH=$LD_LIBRARY_PATH

and this to ~/.theanorc

[dnn]
enabled = True