[FIXED] How to build CUDA custom C++ extension for PyTorch without CUDA?

Issue

I was tasked with creating a CI workflow for building a PyTorch CUDA extension for this application. Up until now, the application was deployed by creating the target AWS VM with a CUDA GPU, pushing all the sources there and running setup.py, but instead I want to do the build in our CI system and deploy pre-built binaries to the production environment.

When running setup.py in the CI system, I get the error "No CUDA GPUs are available" - which is true, there are no CUDA GPUs in the CI system. Is there a way to just build the CUDA extension without a CUDA GPU available?

This is the error message:

gcc -pthread -shared -B /usr/local/miniconda/envs/build/compiler_compat -L/usr/local/miniconda/envs/build/lib -Wl,-rpath=/usr/local/miniconda/envs/build/lib -Wl,--no-as-needed -Wl,--sysroot=/ /app/my-app/build/temp.linux-x86_64-3.6/my-extension/my-module.o -L/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/lib -lc10 -ltorch -ltorch_cpu -ltorch_python -o build/lib.linux-x86_64-3.6/my-extension/my-module.cpython-36m-x86_64-linux-gnu.so
building 'my-extension.my-module._cuda_ext' extension
creating /app/my-app/build/temp.linux-x86_64-3.6/my-extension/src
Traceback (most recent call last):
  File "setup.py", line 128, in <module>
    'build_ext': BuildExtension
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/setuptools/__init__.py", line 153, in setup
    return distutils.core.setup(**attrs)
  File "/usr/local/miniconda/envs/build/lib/python3.6/distutils/core.py", line 148, in setup
    dist.run_commands()
  File "/usr/local/miniconda/envs/build/lib/python3.6/distutils/dist.py", line 955, in run_commands
    self.run_command(cmd)
  File "/usr/local/miniconda/envs/build/lib/python3.6/distutils/dist.py", line 974, in run_command
    cmd_obj.run()
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/setuptools/command/build_ext.py", line 79, in run
    _build_ext.run(self)
  File "/usr/local/miniconda/envs/build/lib/python3.6/distutils/command/build_ext.py", line 339, in run
    self.build_extensions()
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/utils/cpp_extension.py", line 653, in build_extensions
    build_ext.build_extensions(self)
  File "/usr/local/miniconda/envs/build/lib/python3.6/distutils/command/build_ext.py", line 448, in build_extensions
    self._build_extensions_serial()
  File "/usr/local/miniconda/envs/build/lib/python3.6/distutils/command/build_ext.py", line 473, in _build_extensions_serial
    self.build_extension(ext)
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/setuptools/command/build_ext.py", line 196, in build_extension
    _build_ext.build_extension(self, ext)
  File "/usr/local/miniconda/envs/build/lib/python3.6/distutils/command/build_ext.py", line 533, in build_extension
    depends=ext.depends)
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/utils/cpp_extension.py", line 468, in unix_wrap_ninja_compile
    cuda_post_cflags = unix_cuda_flags(cuda_post_cflags)
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/utils/cpp_extension.py", line 377, in unix_cuda_flags
    cflags + _get_cuda_arch_flags(cflags) +
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/utils/cpp_extension.py", line 1407, in _get_cuda_arch_flags
    capability = torch.cuda.get_device_capability()
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/cuda/__init__.py", line 291, in get_device_capability
    prop = get_device_properties(device)
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/cuda/__init__.py", line 296, in get_device_properties
    _lazy_init()  # will define _get_device_properties
  File "/usr/local/miniconda/envs/build/lib/python3.6/site-packages/torch/cuda/__init__.py", line 172, in _lazy_init
    torch._C._cuda_init()
RuntimeError: No CUDA GPUs are available

I'm not very familiar with CUDA and only half proficient in Python (I'm here as the "ops" part of "devops").

Solution

It is not a complete solution, as I lack details to completely figure out a solution. But it should help you or your teammates.

---

So first based on the source code, it is not required to reach torch._C._cuda_init() if you have CUDA arch flags set.

This means the pytorch is trying to figure out the CUDA arch because it is not specified by the user.

Here is a related thread. As you can see, setting the TORCH_CUDA_ARCH_LIST environment to something that fits your environment should work for you.

Answered By - Sraw