pip install not working due to cuda.h missing error

[ycemsubakan]

I have tried pip install pytorch-fft, on two different machines with gpus, and on both I am getting this error:
venv/local/lib/python2.7/site-packages/torch/utils/ffi/../../lib/include/THC/THCGeneral.h:9:18: fatal error: cuda.h: No such file or directory

I tried python 2.7 and 3.6. I also tried to install without using a virtual environment.

P.S.: I first got the error which says the package cffi was missing. I installed that package, and now I am getting this error.

I would appreciate your help!

[riceric22]

You also need to have cuda installed, cuda.h is a dependency from cuda

[ycemsubakan]

I know, I am using cuda 8.0 For some reason it is not accessing cuda.h . (Pytorch runs on the gpu with cuda 8.0)

[riceric22]

Make sure cuda.h is in your path

[ycemsubakan]

Okay, I have added this to my path:
/usr/local/cuda-8.0/targets/x86_64-linux/include/cuda.h

And it still doesn't work, I am getting the same error. Is this the exact path I am supposed to add?

I think the error says that the installation looks at this directory and cannot find cuda.h:
~/venv3.6/lib/python3.6/site-packages/torch/lib/include/THC

And there is no cuda.h under this directory... What should I do?

[riceric22]

Try following the instructions on the cuda quick start guide, they describe how you can update your corresponding path variables (you'll need both path and ld_library_path)

http://docs.nvidia.com/cuda/cuda-quick-start-guide/index.html

[ycemsubakan]

Path and ld_library_path are properly set-up for me. As I have said, my pytorch works perfectly fine with cuda.

[riceric22]

Can you try making a soft link from /usr/local/cuda/ to /usr/local/cuda-8.0/ ?

[ycemsubakan]

Alright, creating the soft link worked. Thank you!

I am closing the issue.

[jzbontar]

I had the same issue. I resolved it by exporting some environmental variables:

$ export C_INCLUDE_PATH=${CUDA_HOME}/include:${C_INCLUDE_PATH}
$ export LIBRARY_PATH=${CUDA_HOME}/lib64:$LIBRARY_PATH
$ pip install pytorch-fft

[stefbraun]

@ycemsubakan , how did you create the soft link? can you give the command please?

[wonderingabout]

the fix for me was to install sudo apt-get install nvidia-cuda-toolkit after sudo apt-get install -y cuda-9.0

also on a brand new ubuntu it is not needed to install nvidia-384, just install cuda-9.0 and graphics driver is included in it

all steps from here :
https://medium.com/@mishra.thedeepak/cuda-and-cudnn-installation-for-tensorflow-gpu-79beebb356d2
and here (change to 3.0.4) :
https://kezunlin.me/post/dacc4196/

now pycuda works (for tensorrt 3.0.4 tar file install on ubuntu 16.04 - cuda 9.0 deb - cudnn 7.0.5 deb)