Could NOT find CUDA when installing cutorch

[Vietanh Hosy]

Hi,

I encounter this problem when trying to install Cutorch using command: 

luarocks install cutorch

CMake Error at /usr/local/share/cmake-3.2/Modules/FindPackageHandleStandardArgs.cmake:138 (message):
  Could NOT find CUDA: Found unsuitable version "5.0", but required is at
  least "5.5" (found /usr)
Call Stack (most recent call first):
  /usr/local/share/cmake-3.2/Modules/FindPackageHandleStandardArgs.cmake:372 (_FPHSA_FAILURE_MESSAGE)
  /usr/local/share/cmake-3.2/Modules/FindCUDA.cmake:855 (find_package_handle_standard_args)
  CMakeLists.txt:7 (FIND_PACKAGE)

I'm sure that there is no Cuda version in /usr and /usr/local/cuda links to cuda-7.0. I could not find any possible cause of this problem, hence I don't know how to solve it.

I even manually edit the FindCUDA.cmake file, to specify CUDA_TOOLKIT_ROOT_DIR before trying to find:

 855 set(CUDA_TOOLKIT_ROOT_DIR /usr/local/cuda-7.0)

 856 message(${CUDA_TOOLKIT_ROOT_DIR})

 857 find_package_handle_standard_args(CUDA

 858   REQUIRED_VARS 

 859     CUDA_TOOLKIT_ROOT_DIR 

 860     CUDA_NVCC_EXECUTABLE

 861     CUDA_INCLUDE_DIRS

 862     CUDA_CUDART_LIBRARY

 863   VERSION_VAR             

 864     CUDA_VERSION                                                                                

 865   )

And now comes the error message:

CMake Error at /usr/local/share/cmake-3.2/Modules/FindPackageHandleStandardArgs.cmake:138 (message):

  Could NOT find CUDA: Found unsuitable version "5.0", but required is at

  least "5.5" (found /usr/local/cuda-7.0)

I don't know what happened with cutorch/CMake and I cannot make things work.

Please advice. Many thanks.

[alban desmaison]

What does "/usr/local/cuda/bin/nvcc --version" gives you?

[Vietanh Hosy]

Hi,

It's 7.0

nvcc: NVIDIA (R) Cuda compiler driver

Copyright (c) 2005-2015 NVIDIA Corporation

Built on Mon_Feb_16_22:59:02_CST_2015

Cuda compilation tools, release 7.0, V7.0.27

I doubt it could be some problem with CMake :(

--
You received this message because you are subscribed to a topic in the Google Groups "torch7" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/torch7/QkRk8eKAJLw/unsubscribe.
To unsubscribe from this group and all its topics, send an email to torch7+un...@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at https://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

[alban desmaison]

If I am not mistaken, we use the builtin cmake cuda detection here https://github.com/torch/cutorch/blob/master/CMakeLists.txt#L7

Make sure that you do not have a CMAKE_LIBRARY_PATH pointing to an older version of CUDA.

The simplest would be to make sure that you removed all other version of cuda on your computer :)

On Thursday, February 18, 2016 at 10:21:15 AM UTC, Vietanh Hosy wrote:

Hi,

It's 7.0

nvcc: NVIDIA (R) Cuda compiler driver

Copyright (c) 2005-2015 NVIDIA Corporation

Built on Mon_Feb_16_22:59:02_CST_2015

Cuda compilation tools, release 7.0, V7.0.27

I doubt it could be some problem with CMake :(

[Vietanh Hosy]

Hi,

I have found the problem. Somehow there is a /usr/bin/nvcc file that is of Cuda 5.0. It seems that CMake tries to find from that folder before going to CUDA_TOOLKIT_ROOT_DIRS.

Simply removing this file solves the problem.

Thanks.

[Michael Xiongfei Wei]

I removed the nvcc file in /usr/bin but I got another error:

Installing https://raw.githubusercontent.com/torch/rocks/master/cutorch-scm-1.rockspec...

Using https://raw.githubusercontent.com/torch/rocks/master/cutorch-scm-1.rockspec... switching to 'build' mode

Cloning into 'cutorch'...

remote: Counting objects: 107, done.

remote: Compressing objects: 100% (103/103), done.

remote: Total 107 (delta 8), reused 44 (delta 2), pack-reused 0

Receiving objects: 100% (107/107), 157.40 KiB | 0 bytes/s, done.

Resolving deltas: 100% (8/8), done.

Checking connectivity... done.

cmake -E make_directory build && cd build && cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH="/home/ubuntu/torch/install/bin/.." -DCMAKE_INSTALL_PREFIX="/home/ubuntu/torch/install/lib/luarocks/rocks/cutorch/scm-1" && make -j$(getconf _NPROCESSORS_ONLN) install

-- The C compiler identification is GNU 4.8.4

-- The CXX compiler identification is GNU 4.8.4

-- Check for working C compiler: /usr/bin/cc

-- Check for working C compiler: /usr/bin/cc -- works

-- Detecting C compiler ABI info

-- Detecting C compiler ABI info - done

-- Check for working CXX compiler: /usr/bin/c++

-- Check for working CXX compiler: /usr/bin/c++ -- works

-- Detecting CXX compiler ABI info

-- Detecting CXX compiler ABI info - done

-- Found Torch7 in /home/ubuntu/torch/install

CMake Error at /usr/share/cmake-2.8/Modules/FindCUDA.cmake:548 (message):

  Specify CUDA_TOOLKIT_ROOT_DIR

Call Stack (most recent call first):

  CMakeLists.txt:7 (FIND_PACKAGE)

-- Configuring incomplete, errors occurred!

See also "/tmp/luarocks_cutorch-scm-1-153/cutorch/build/CMakeFiles/CMakeOutput.log".

Error: Build error: Failed building.

[Michael Xiongfei Wei]

[shaoguang guo]

Hi,
I think he best solutions is not to remove the link file , especilly when you just a comman user.
Just add an option when compile
-DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-6.0
no matter what 'nvcc --version' output.

This is the method when I try to install software on a supercomputer, which I have no root privilege.

Hope to help.


在 2016年2月19日星期五 UTC+8上午3:13:14，Vietanh Hosy写道：

[John Pertoft]

I'm getting similar problems. 

I have cuda7.5 installed at /usr/local/cuda-7.5 and /usr/local/cuda is a symlink to it.

nvcc --version says 7.5

I have removed the /usr/bin/nvcc (which did say version 5.5).

What else can I try?

Error message:

Warning: unmatched variable LUALIB

CMake Error at /usr/share/cmake-3.2/Modules/FindPackageHandleStandardArgs.cmake:138 (message):

  Could NOT find CUDA: Found unsuitable version "5.5", but required is at

  least "6.5" (found /usr)

Call Stack (most recent call first):

  /usr/share/cmake-3.2/Modules/FindPackageHandleStandardArgs.cmake:372 (_FPHSA_FAILURE_MESSAGE)

  /home/john/torch/install/share/cmake/torch/FindCUDA.cmake:1009 (find_package_handle_standard_args)

  CMakeLists.txt:7 (FIND_PACKAGE)

Error: Build error: Failed building.

[Vietanh Hosy]

Hi John,

I think there is nvcc 5.5 somewhere. You can search for all the nvcc file in your system, then check the version of each file and try changing name, etc.

I can only suggest that approach.

[John Pertoft]

Thanks for a quick reply!

I searched for all instances of nvcc on the system and found another of version 5.5 in /usr/lib/nvidia-cuda-toolkit/bin which I renamed temporarily. I have root access but other people are using the computer as well so I didn't want to remove it in case someone uses it. I still got the same error when installing though.

Is there any environment variable that causes this problem or circumvents it?

[John Pertoft]

Solved it by bypassing the FindCUDA.cmake script and just hardcoding all the CUDA variables instead.

[hope yao]

Hi John,

I also encountered the problem as you and still I'm stuck here. 

1. I deleted /usr/bin/nvcc

2. I tried to rename all NVCC file I found in my computer.

3. I also tried to add SET(CUDA_TOOLKIT_ROOT_DIR "/usr/local/cuda") in the FindCUDA.cmake file, but it seems that this file will be overwrite in runing install.sh

But non of them worked.

I'm trying to follow your solution, but I don't quite understand how to "bypassing the FindCUDA". Could you please give me more detailed information how you solved this? Your kind help would be highly appreciated.

[hope yao]

Okay. Now I solved it. There are two FindCUDA.cmake. That line needs to be added to torch/cmake/3.6/Modules/FindCUDA.cmake. The other file is copied from this one.

[ty Du]

Hi yao,

Could you please specify where to add SET(CUDA_TOOLKIT_ROOT_DIR "/usr/local/cuda") in FindCUDA.cmake? I'm new to cmake and torch. I'm also suffering from the ' Could NOT find CUDA' problem.

Thanks a lot!

在 2016年8月8日星期一 UTC+8上午4:55:25，hope yao写道：