look into why the NVIDIA CUDA Dockerfile doesn't build pytorch

[soumith]

@devendrachaplot reports that starting from: compute.nvidia.com/nvidia/cuda pytorch has compile errors. Look into it.

[ngimel]

FWIW, with the following Dockerfile

FROM nvidia/cuda:8.0-cudnn5-devel-ubuntu14.04 

RUN apt-get update && apt-get install -y --no-install-recommends \
         build-essential \
         cmake \
         git \
         curl \
         ca-certificates \
         libjpeg-dev \
         libpng-dev &&\
     rm -rf /var/lib/apt/lists/*

RUN curl -o ~/miniconda.sh -O  https://repo.continuum.io/miniconda/Miniconda3-4.2.12-Linux-x86_64.sh  && \
     chmod +x ~/miniconda.sh && \
     ~/miniconda.sh -b -p /opt/conda && \     
     rm ~/miniconda.sh && \
     /opt/conda/bin/conda create -y --name pytorch-py35 python=3.5.2 numpy scipy ipython mkl&& \
     /opt/conda/bin/conda clean -ya 
ENV PATH /opt/conda/envs/pytorch-py35/bin:$PATH
RUN conda install --name pytorch-py35 -c soumith magma-cuda80
# This must be done before pip so that requirements.txt is available
WORKDIR /opt/pytorch
COPY . .

RUN cat requirements.txt | xargs -n1 pip install --no-cache-dir && \
    TORCH_CUDA_ARCH_LIST="3.5 5.2 6.0 6.1+PTX" \
    CMAKE_LIBRARY_PATH=/opt/conda/envs/pytorch-py35/lib \
    CMAKE_INCLUDE_PATH=/opt/conda/envs/pytorch-py35/include \
    pip install -v .

WORKDIR /workspace
RUN chmod -R a+w /workspace

I can build and run tests Ok.

[devendrachaplot]

This isn't extremely urgent for me, but if you want to reproduce the error, these are the steps I followed:
I used the docker compute.nvidia.com/nvidia/cuda.

In the docker, I ran:
wget https://repo.continuum.io/archive/Anaconda2-4.2.0-Linux-x86_64.sh
bash Anaconda2-4.2.0-Linux-x86_64.sh
export CMAKE_PREFIX_PATH=/root/anaconda2/
conda install numpy mkl setuptools cmake gcc cffi
conda install -c soumith magma-cuda80
git clone https://github.com/pytorch/pytorch.git
cd pytorch
pip install -r requirements.txt
python setup.py install

The error message:
6 errors detected in the compilation of "/tmp/tmpxft_00000e36_00000000-7_THCTensorSortShort.cpp1.ii".
CMake Error at THC_generated_THCTensorSortShort.cu.o.cmake:267 (message):
Error generating file
/pytorch/torch/lib/build/THC/CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortShort.cu.o

make[2]: *** [CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortShort.cu.o] Error 1
6 errors detected in the compilation of "/tmp/tmpxft_00000e5d_00000000-7_THCTensorSortDouble.cpp1.ii".
CMake Error at THC_generated_THCTensorSortDouble.cu.o.cmake:267 (message):
Error generating file
/pytorch/torch/lib/build/THC/CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortDouble.cu.o

make[2]: *** [CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortDouble.cu.o] Error 1
6 errors detected in the compilation of "/tmp/tmpxft_00000e4c_00000000-7_THCTensorSortChar.cpp1.ii".
CMake Error at THC_generated_THCTensorSortChar.cu.o.cmake:267 (message):
Error generating file
/pytorch/torch/lib/build/THC/CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortChar.cu.o

make[2]: *** [CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortChar.cu.o] Error 1
6 errors detected in the compilation of "/tmp/tmpxft_00000e78_00000000-7_THCTensorSortInt.cpp1.ii".
CMake Error at THC_generated_THCTensorSortInt.cu.o.cmake:267 (message):
Error generating file
/pytorch/torch/lib/build/THC/CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortInt.cu.o

make[2]: *** [CMakeFiles/THC.dir/generated/./THC_generated_THCTensorSortInt.cu.o] Error 1
make[1]: *** [CMakeFiles/THC.dir/all] Error 2
make: *** [all] Error 2

[ngimel]

It's better to specify a particular tag for the image rather than rely on :latest being what you want. Can you try Dockerfile above? If you want to build pytorch interactively within container, you can just run the RUN commands inside container.

[devendrachaplot]

It works with the Dockerfile mentioned by @ngimel. Thanks a lot!

[ngimel]

Just a heads up, now this dockerfile is broken because of #556

[soumith]

@ngimel fixed now

[apaszke]

Official dockerfiles are now merged in master (thanks to @ngimel).