🐛 Bug

Problem occurs when building multiple extension, and giving the same dictionnary for extra_compile_args

To Reproduce

Very easy to reproduce with https://github.com/pytorch/extension-cpp/tree/master/cuda

change the setup.py to have a second fake extension and add extra_compile_args.
this version of setup.py won't work, but the second proposed version will:

args = {'nvcc':[], 'cxx':[]}

setup(
    name='lltm_cuda',
    ext_modules=[
        CUDAExtension('lltm_cuda', [
            'lltm_cuda.cpp',
            'lltm_cuda_kernel.cu',
        ], extra_compile_args=args),
        CUDAExtension('lltm_cuda2', [
            'lltm_cuda.cpp',
            'lltm_cuda_kernel.cu',
        ], extra_compile_args=args),
    ],
    cmdclass={
        'build_ext': BuildExtension
    })

This outputs (although it compiles)

<command-line>:0:0: warning: "TORCH_EXTENSION_NAME" redefined
<command-line>:0:0: note: this is the location of the previous definition

and importing first module gets the error

In [1]: import lltm_cuda                                                                                                                                                                 
---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
<ipython-input-1-15b747285754> in <module>
----> 1 import lltm_cuda

ImportError: dynamic module does not define module export function (PyInit_lltm_cuda)

args = {'nvcc':[], 'cxx':[]}

setup(
    name='lltm_cuda',
    ext_modules=[
        CUDAExtension('lltm_cuda', [
            'lltm_cuda.cpp',
            'lltm_cuda_kernel.cu',
        ], extra_compile_args=args),
        CUDAExtension('lltm_cuda2', [
            'lltm_cuda.cpp',
            'lltm_cuda_kernel.cu',
        ], extra_compile_args=copy.deepcopy(args)),
    ],
    cmdclass={
        'build_ext': BuildExtension
    })

This works fine and both modules can be imported.

Expected behavior

Same behaviour between the two versions of the install script.

Additional context

The problem comes from the fact that a simple shallow copy of dictionnary is not enough, since compile args are nested into two lists, one for C++ and one for cuda.
Recommended actions should be either change this line
https://github.com/pytorch/pytorch/blob/master/torch/utils/cpp_extension.py#L373
to make a copy.deepcopy instead of a simple copy.copy, or change this line https://github.com/pytorch/pytorch/blob/master/torch/utils/cpp_extension.py#L376 to make a second copy.copy .

First solution is less verbose but might be less secure, since we can't control the depth of the copy. I'm personally more inclined to first solution as I don't see a way to unintentionally copy too much with a simple extra_compile_args dictionnary.