Support building with GPU support even when CUDA is not currently available

[vnavkal]

Currently we use this conditional to determine whether to build with GPU support:

DCNv2/setup.py

Line 33 in c7f778f

if torch.cuda.is_available() and CUDA_HOME is not None:

But it would be nice to be able to build the library with GPU support even in environments where CUDA is not available. In particular, I'd like to be able to do this while building a docker image.

If this sounds like a reasonable change, I'm happy to submit a PR for it. I think it would look quite similar to pytorch/vision#927.

[zhaotf16]

Hi, I met the same problem. Did you solve it? It would be better to be able to build on a login-node without GPU and to run on a gpu-node

[vnavkal]

@zhaotf16 I solved it by modifying the conditional to allow being overridden by a FORCE_CUDA environment variable, as in the torchvision commit I linked to. (I did this only in my copy of the repo and didn't submit a PR for that change.)