Resnet50 Inference [1, 3, 480, 1024] use 11G GPU Memory

[LossNAN]

When I use the ResNet-50 classification network with an image size of [1, 3, 480, 1024] for inference using the eval mode, the GPU memory consumption is close to 11 GB. Is this normal?

with torch.no_grad():
    self.model = init_model(config, checkpoint)
    
    tensor = torch.randn(1, 3, 480, 1024).to('cuda:0').half()
    data = {'img': tensor}
    
    for i in range(10):
        _ = self.model(return_loss=False, **data)