I am trying to run tensorflow on a cuda machine over the cloud where I do not have sudo privileges.
I installed tensorflow in a virtualenv but I am not able to point it towards a specific cuda version. Can this be only done when installing from source?
On my Ubuntu systems I use (to be executed in the same terminal as the one in which you run TensorFlow):
# CUDA 7.5: export PATH=/usr/local/cuda-7.5/bin:$PATH export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64:$LD_LIBRARY_PATH # CUDA 8.0: export PATH=/usr/local/cuda-8.0/bin:$PATH export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64:$LD_LIBRARY_PATH You may also want to point to CuDNN and CNMeM:
# CuDNN: export LD_LIBRARY_PATH=/scratch1/francky/cudnn/cuda/lib64:$LD_LIBRARY_PATH export CPATH=/scratch1/francky/cudnn/cuda/include:$CPATH export LIBRARY_PATH=/scratch1/francky/cudnn/cuda/lib64:$LD_LIBRARY_PATH # CNMeM: export LD_LIBRARY_PATH=/scratch1/francky/cudnn/cnmem/build:$LD_LIBRARY_PATH export CPATH=/scratch1/francky/cudnn/cnmem/include:$CPATH export LIBRARY_PATH=/scratch1/francky/cudnn/cnmem/build:$LD_LIBRARY_PATH FYI: Is there any point in using CNMeM when running TensorFlow?
Usually CUDA_HOME allows to solve such kind of issues. As for TensorFlow up to 0.12, the homepage gives detailed information.
Version 1.0 is available now, and the documentation has been updated (not re-read at this point), so there might be extra information there if you switched.
