Force cuda since torch ask for a device, not if cuda is in fact avail…

[ccoulombe]

Force cuda since torch ask for a device, not if cuda is in fact available.

torch.cuda.is_available() is misleading since it actually checks for an available device, not if cuda is available.
Building on build-nodes where a gpu device is not necessarily available then fails. Force cuda through FORCE_CUDA=1 to allow building.

[bowenc0221]

Thanks for the PR, two questions:

1. FORCE_CUDA is not always set in environment variable, need to provide it with a default value
2. Do we still need CUDA_HOME when using FORCE_CUDA=1?

Maybe let's follow the same logic as detectron2 (and remove is_rocm_pytorch):
https://github.com/facebookresearch/detectron2/blob/9176db667837c57e7118d92a90f887565bd6d4c1/setup.py#L66-L68

[ccoulombe]

1. os.environ.get('FORCE_CUDA') returns None which evaluate to False when the env. variable is not set. So it not necessary to provide a different default value.

>>> import os
>>> os.environ['FORCE_CUDA'] = '1'
>>> bool(os.environ.get('FORCE_CUDA'))
True
>>> del os.environ['FORCE_CUDA']
>>> bool(os.environ.get('FORCE_CUDA'))
False

2. And yes, CUDA_HOME needs to be set and exists in order to build the code with GPU support.

The current code is equivalent to what detectron2 does

[bowenc0221]

Hi @ccoulombe, I've tested the PR and it works well. Please modify the error message to reflect your change as well (pls see my comment) and I will merge this PR. Thanks!

[ccoulombe]

Will do tomorrow!