CUDA enabled torchvision silently broken builds if not built from a host with a physical GPU.

[Micket]

It still links to all the CUDA libraries, so at a glance it doesn't look obviously broken. But, apparently a vital CUDA backend for torchvision::nms is missing.

The following code will reveal if the issue is present with the installation;

import torch
from torchvision.ops import boxes as box_ops

N = 10

device = torch.device('cuda:0')
boxes = torch.rand((N,4), device=device)
scores = torch.rand((N), device=device)
idxs = torch.zeros((N), device=device)
iou_threshold = 0.5
box_ops.batched_nms(boxes.float(), scores, idxs, iou_threshold)

I think we should either try to detect if it's about to be a problem, or fix it somehow (perhaps some extra flags can coerce torchvision to build the backend regardless.

[branfosj]

Switch to using CMakePythonPackage for the easyblock?

Also, setting env.setvar('FORCE_CUDA', '1') builds the C++ bindings on a non-GPU node. We might be able to use -DWITH_CUDA=on.