Closed t-kalinowski closed 1 year ago
Beginning with TF 2.12, cudnn can now be provided via a pip package that contains libcudnn.so.8. Unfortunately, TF cannot find it if it's not on the LD_LIBRARY_PATH.
https://pypi.org/project/nvidia-cudnn-cu11/
Upstream docs indicate this is the new recommended approach:
pip install nvidia-cudnn-cu11==8.6.0.163 CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)")) export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/:$CUDNN_PATH/lib
We will have to include hooks in the R package to add these entries to LD_LIBRARY_PATH.
LD_LIBRARY_PATH
Beginning with TF 2.12, cudnn can now be provided via a pip package that contains libcudnn.so.8. Unfortunately, TF cannot find it if it's not on the LD_LIBRARY_PATH.
https://pypi.org/project/nvidia-cudnn-cu11/
Upstream docs indicate this is the new recommended approach:
We will have to include hooks in the R package to add these entries to
LD_LIBRARY_PATH.