Guard CUDA calls with an explicit check

[st235]

If called on MacOS where CUDA is not available the method will crash with the following error:

  File "/nanodet/model/arch/one_stage_detector.py", line 50, in inference
    torch.cuda.synchronize()
  File "/python3.9/site-packages/torch/cuda/__init__.py", line 564, in synchronize
    _lazy_init()
  File "/python3.9/site-packages/torch/cuda/__init__.py", line 221, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

The proposed fix is to check if CUDA is available before accessing the API

[st235]

@RangiLyu can you, please, take a look?

[deshpandeneeraj]

I would suggest using next(self.parameters()).device instead. Since cuda can be initialized but still not used

[st235]

Thank you for the feedback.

I would stick with the proposed approach as it is consistent with the existing checks they have in the repo.
For example, torch.cuda.memory_reserved() / 1e9 if torch.cuda.is_available() else 0.

Since cuda can be initialized but still not used

I believe this case is less severe as it does not lead to runtime crash.

[st235]

@RangiLyu any luck you can take a look?

[RangiLyu]

@RangiLyu any luck you can take a look?

Sorry for the late reply! (I've been too busy lately)
Thanks a lot for your contribution!