Keeps referencing wrong version of gcc when building from source

al3x · June 10, 2018, 11:37pm

I’m trying to build pytorch from source, having just installed MAGMA due to the error

No CUDA implementation of 'potrf'. Install MAGMA and rebuild cutorch

I’m doing this on a 2013 macbook, so I’m aware support is limited, however I did it before so I have faith it will work again!

The problem I’m having is when I try and call setup.py, it keeps referencing a different version of gcc. I have both 8.1.0 and 5.3.0 – 5 is the one I want. I’ve set up environment variables corresponding to this version in ~/.bash_profile

$ echo $CC
/usr/local/bin/gcc-5
$ echo $CXX
/usr/local/bin/g++-5

I’ve also set up symbolic links to /usr/bin/gcc and /usr/bin/g++ in case it was calling them.

However, when I run the command CC=gcc-5 CXX=g++-5 DCUDA_HOST_COMPILER=/usr/local/bin/gcc-5 python setup.py install

I get the error:

-- Caffe2: CUDA detected: 8.0
-- Caffe2: CUDA nvcc is: /usr/local/cuda/bin/nvcc
-- Caffe2: CUDA toolkit directory: /usr/local/cuda
-- Caffe2: Header version is: 8.0
-- Found cuDNN: v6.0.21  (include: /usr/local/cuda/include, library: /usr/local/cuda/lib/libcudnn.6.dylib)
CMake Error at cmake/public/cuda.cmake:321 (message):
  CUDA 8.0 is not compatible with GCC version >= 6.  Use the following option
  to use another version (for example):

    -DCUDA_HOST_COMPILER=/usr/bin/gcc-5

Call Stack (most recent call first):
  cmake/Dependencies.cmake:401 (include)
  CMakeLists.txt:204 (include)


-- Configuring incomplete, errors occurred!
See also "/Users/askates/Documents/GitRepos/pytorch/build/CMakeFiles/CMakeOutput.log".
See also "/Users/askates/Documents/GitRepos/pytorch/build/CMakeFiles/CMakeError.log".

and when I look at CMakeOutput.log I see it’s referencing /usr/local/bin/gcc-8

How can I make it reference gcc-5 instead of gcc-8? What am I missing?

al3x · June 11, 2018, 10:15am

I seem to have gotten it to work with cmake CMakeLists.txt, however I now get the error:

CMake Error at third_party/sleef/CMakeLists.txt:27 (message):
  SLEEF does not allow in-source builds.

  You can refer to doc/build-with-cmake.md for instructions on how provide a
  separate build directory.  Note: Please remove autogenerated file
  `CMakeCache.txt` and directory `CMakeFiles` in the current directory.


-- Configuring incomplete, errors occurred!

I’ve tried building sleef separately and if I run

$ mkdir third_party/sleef/build_dir
$ cd third_party/sleef/build_dir
$ cmake ..

It builds successfully, but when I try and run setup.py I get this:

-- The CXX compiler identification is unknown
-- The C compiler identification is unknown
CMake Error at CMakeLists.txt:6 (project):
  The CMAKE_CXX_COMPILER:

    /usr/local/bin/g++-8

  is not a full path to an existing compiler tool.

  Tell CMake where to find the compiler by setting either the environment
  variable "CXX" or the CMake cache entry CMAKE_CXX_COMPILER to the full path
  to the compiler, or to the compiler name if it is in the PATH.


CMake Error at CMakeLists.txt:6 (project):
  The CMAKE_C_COMPILER:

    /usr/local/bin/gcc-8

  is not a full path to an existing compiler tool.

  Tell CMake where to find the compiler by setting either the environment
  variable "CC" or the CMake cache entry CMAKE_C_COMPILER to the full path to
  the compiler, or to the compiler name if it is in the PATH.


-- Configuring incomplete, errors occurred!

tom · June 11, 2018, 12:10pm

I’m not sure what the official way is, but here is what has consistently worked for me for a long time by now:

I have set up two directories with symlinks I add to my path. In ~/pytorch/compiler I have

g++ -> /usr/bin/g++-5
gcc -> /usr/bin/gcc-5
x86_64-linux-gnu-g++ -> /usr/bin/x86_64-linux-gnu-g++-5
x86_64-linux-gnu-gcc -> /usr/bin/x86_64-linux-gnu-gcc-5

and because I like ccache (I use the Debian/unstable ccache seems to do OK) I also have ~/pytorch/compilers-with-ccache

g++ -> /usr/bin/ccache
g++-5 -> /usr/bin/ccache
gcc -> /usr/bin/ccache
gcc-5 -> /usr/bin/ccache
x86_64-linux-gnu-g++ -> /usr/bin/ccache
x86_64-linux-gnu-g++-5 -> /usr/bin/ccache
x86_64-linux-gnu-gcc -> /usr/bin/ccache
x86_64-linux-gnu-gcc-5 -> /usr/bin/ccache

Then I compile PyTorch with

PATH=~/pytorch/compilers-with-ccache/:~/pytorch/compilers/:$PATH python3 setup.py bdist_wheel

to get a whl in the dist directory.
Finally, I install with sudo pip3 -U --no-deps dist/something.whl because my system Python 3 is with pip3.

You might have to clean (python3 setup.py clean) before rerunning setup.py.

Best regards

Thomas

al3x · June 11, 2018, 1:09pm

I just tried that – I don’t have a ccache directory though so I missed that out. I also added all the others just in case:

$ mkdir compilers; cd compilers
$ ln -s /usr/local/bin/g++-5 g++
$ ln -s /usr/local/bin/gcc-5 gcc
$ ln -s /usr/local/bin/gcc-ar-5 gcc-ar
$ ln -s /usr/local/bin/gcc-nm-5 gcc-nm
$ ln -s /usr/local/bin/gcc-ranlib-5 gcc-ranlib
$ cd ..
$ PATH=~/Documents/GitRepos/pytorch/compilers/:$PATH python setup.py bdist_wheel

I get this again:

-- The CXX compiler identification is unknown
CMake Error at CMakeLists.txt:6 (project):
  The CMAKE_CXX_COMPILER:

    /usr/local/bin/g++-8

  is not a full path to an existing compiler tool.

  Tell CMake where to find the compiler by setting either the environment
  variable "CXX" or the CMake cache entry CMAKE_CXX_COMPILER to the full path
  to the compiler, or to the compiler name if it is in the PATH.


-- Configuring incomplete, errors occurred!

I don’t know why it’s defaulting to g++-8 still…

$ compilers/gcc --version
gcc (GCC) 5.3.0
Copyright (C) 2015 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

al3x · June 11, 2018, 4:37pm

I got the whole project again, overwriting all local changes. Your instructions seemed to work up until I got to

[ 48%] Linking CXX shared library ../lib/libcaffe2.dylib
ld: library not found for -lgcc_s
collect2: error: ld returned 1 exit status
make[2]: *** [lib/libcaffe2.dylib] Error 1
make[1]: *** [caffe2/CMakeFiles/caffe2.dir/all] Error 2
make: *** [all] Error 2
Failed to run 'bash tools/build_pytorch_libs.sh --with-cuda --with-nnpack caffe2 nanopb libshm THD'

tom · June 12, 2018, 6:03am

So I know nothing at all about this (I missed that you are on OS X , but googling this error message, people apparently do funny manual symlinking with some success.

Best regards

Thomas