Lazily load libcuda libnvrtc from c++ (#17317)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/16860
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17317

Differential Revision: D14157877

Pulled By: zdevito

fbshipit-source-id: c37aec2d77c2e637d4fc6ceffe2bd32901c70317

setup.py

@@ -638,31 +638,6 @@ if not IS_WINDOWS:
                    )
     extensions.append(DL)
 
-
-if USE_CUDA:
-    thnvrtc_link_flags = extra_link_args + [make_relative_rpath('lib')]
-    if IS_LINUX:
-        thnvrtc_link_flags = thnvrtc_link_flags + ['-Wl,--no-as-needed']
-    # these have to be specified as -lcuda in link_flags because they
-    # have to come right after the `no-as-needed` option
-    if IS_WINDOWS:
-        thnvrtc_link_flags += ['cuda.lib', 'nvrtc.lib']
-    else:
-        thnvrtc_link_flags += ['-lcuda', '-lnvrtc']
-    cuda_stub_path = [cuda_lib_path + '/stubs']
-    if IS_DARWIN:
-        # on macOS this is where the CUDA stub is installed according to the manual
-        cuda_stub_path = ["/usr/local/cuda/lib"]
-    THNVRTC = Extension("torch._nvrtc",
-                        sources=['torch/csrc/nvrtc.cpp'],
-                        language='c++',
-                        extra_compile_args=main_compile_args + extra_compile_args,
-                        include_dirs=[cwd],
-                        library_dirs=library_dirs + cuda_stub_path,
-                        extra_link_args=thnvrtc_link_flags,
-                        )
-    extensions.append(THNVRTC)
-
 # These extensions are built by cmake and copied manually in build_extensions()
 # inside the build_ext implementaiton
 extensions.append(