[Fix] fix torch allocator resouce releasing

[AllentDan]

Reproduce codes in master branch:

from mmdeploy.backend.tensorrt.wrapper import TRTWrapper
trt_model = TRTWrapper(
    '/path_to/cls/resnet/end2end.engine',
    ['output'])
input = torch.rand(1, 3, 224, 224).cuda()
output_trt = trt_model.forward({'input': input})
print(output_trt)

Error might triggered:

/mmdeploy/utils/logging.py(28): get_logger
Exception caught in deallocate(): TypeError: 'NoneType' object is not callable

or

codes/mmdeploy/mmdeploy/backend/tensorrt/torch_allocator.py(61): deallocate

[ERROR] Exception caught in deallocate(): AttributeError: 'NoneType' object has no attribute 'caching_allocator_delete'

[grimoire]

Unable to reproduce the error.
Why would logging and torch.cuda be None?

[AllentDan]

Unable to reproduce the error. Why would logging and torch.cuda be None?

Don't know either. Just happened at the end of the program. Would it be possible that torch.cuda and logging are released at that time?

[grimoire]

More investigations are required.

[AllentDan]

After some experiments, we may have following conclusions:

1. resources including torch and logging may be released randomly before we call deallocate
2. this kind of releasing does not delete the allocated cache since we can still delete the cache manually.

[grimoire]

Could report to TensorRT team.