Good idea to check if duplicate python packages installed
pip freeze
and pip uninstall any prior version you had and pip install the version we tried to install. E.g. on conda you might need to do:
pip uninstall numpy
pip install numpy==1.13.1 # or whatever version was attempted to be installed by the wheel
After pip installing the wheel, make sure you use a fresh bash environment to ensure the python cache is not used.
e.g. Warning: h2o4gpu_kmeans_lib shared object (dynamic library) /home/tomk/anaconda3/envs/h2o4gpu/lib/python3.6/site-packages/h2o4gpu/libs/../../ch2o4gpu_gpu.so failed to load.
This usually means some other library is missing. One can run:
cd <your python environment path>/site-packages/
e.g. for pyenv: /home/$USER/.pyenv/versions/3.6.1/lib/python3.6/site-packages/
ldd ch2o4gpu_gpu.so
To see if things are missing. Make sure you installed cuda and it's linked correctly to /usr/local/cuda . And make sure you set the environment variables related to cuda.
And check if any missing dependencies
If compiled with icc (default if present) and have conda, need to do:
conda install --no-dep -c intel icc_rt