Error importing torch on machines without cuda

[jmspereira]

Hello!

First of all, thank you for the amazing Dockerfile template to build PyTorch from the source!
I was able to use the wheel produced in environments that have CUDA installed, however, in environments that do not have CUDA I am not able of executing the command 'import torch', saying correctly that I do not have CUDA.

This is a large problem for me, the possibility of importing torch in environments that do not have CUDA. Do you know what is the configuration in the build that is causing this? If I use PyTorch from the wheels published by them, I am able of importing torch.

Thank you again!

[veritas9872]

Hello, the reason is because of the USE_CUDA, USE_CUDNN, and FORCE_CUDA flags in the PyTorch and subsidiary library builds. I may update the Dockerfile to have a global USE_CUDA option later. Also, the TORCH_CUDA_ARCH_LIST flag should not be given. However, the speed-up gained by using a source-build on CPU is probably limited. Other methods such as using the C++ API or the JIT compiler in TorchScript are much more effective.

[veritas9872]

Current main branch now has a USE_CUDA flag.

[jmspereira]

Thank you for the answer!

Possibly I didn't explain myself well, sorry about that! My goal is to have a build that supports CUDA, but doesn't fail to run if CUDA isn't installed. As with the PyTorch provided wheels (e.g., https://download.pytorch.org/whl/cu110/torch-1.7.1%2Bcu110-cp38-cp38-linux_x86_64.whl). With this wheel I can execute PyTorch in CPU, even if I do not have CUDA installed.

[veritas9872]

Hello. I will close this issue for now but please feel free to reopen it if any other issues occur.