Build pytorch v0.4.0 with CuDNN

1

I was trying to build pytorch 0.4.0 on my linux with cuda 9.1. However, after doing python setup.py install, I found that cudnn is not included given UserWarning: PyTorch was compiled without cuDNN support. To use cuDNN, rebuild PyTorch making sure the library is visible to the build system.

I did some check:

print(torch.backends.cudnn.version())
None
print(torch.cuda.is_available())
True

Further, in terminal

$CUDNN_LIB_DIR
-bash: /usr/local/cuda-9.1/lib64/: Is a directory
$CUDNN_INCLUDE_DIR
-bash: /usr/local/cuda-9.1/include/: Is a directory

Does anyone have some ideas on what I should do now?

2

libcudnn.so and cudnn.h have to be in their respective CUDNN_LIB_DIR and CUDNN_INCLUDE_DIR, make sure that this is the case.
Cudnn detection during compilation is done in this file https://github.com/pytorch/pytorch/blob/master/tools/setup_helpers/cudnn.py, make sure that cudnn is actually found when you compile.

1 Like
3

It’s weird that they are not in the corresponding cuda dir. Any thoughts? Also, I once tried to install pytorch with a provided binary file, and I did not get errors when running the code. How does that happen?
Thanks!

4

You have to separately install cudnn either by downloading it from nvidia’s site or from deb package http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64/ or via conda package. Cudnn is not a part of cuda toolkit. Pytorch binaries include cudnn, so when you are using pytorch binaries, you don’t have to separately install cudnn.