Pytorch only support cuDNN 6.x or above, but another program on my computer needs cuDNN 5.1. So there are 2 versions on my computer and paths are: /usr/local/cuda-8.0/lib64/cuDNNv5 /usr/local/cuda-8.0/lib64/cuDNNv6
However, even I add /usr/local/cuda-8.0/lib64/cuDNNv6 in LD_LIBRARY_PATH and CUDNN_LIB_DIR, cudnn is not found when I install pytorch.
Any one can give some help? How to specify CUDNN path for installing pytorch.
Can you try setting CUDNN_INCLUDE_DIR to the include folder and CUDNN_LIB_DIR to the lib folder?
I’m not sure what’s under your cuDNNv6 folder but in general the CUDNN_LIB_DIR should have libcudnn.so (of some form) and CUDNN_INCLUDE_DIR should have cudnn.h
CC and CXX must be GCC-5 for CUDA-8 and can be GCC-6 for CUDA-9 (I don’t think GCC-7 will work for any of those).
Don’t forget the WITH_CUDA and WITH_CUDNN
I am using GeForce GTX 1080, and pytorch can be installed successfully without CUDNN.
I modified tools/setup_helpers/cudnn.py to print environment path of cudnn, and got None.
You should also consider looking into Docker. It is the best thing ever for keeping different libraries and programs running in their own separate containers, with just a 0.1% performance penalty.
Regarding Docker, on my side I had mitigated experience while using it for DL competition environment on AWS, it was really slow and a pain to setup. Personally I use a LXC container dedicated to DL (Docker used to use LXC tech in the past) with no perf penalty. Your mileage may vary but any kind of container is good to make sure your DL environment doesn’t pollute/is not polluted by your main system, especially Python version updates. And backup-ing a container is relatively straightforward.