Problem with model inference using accelerate

[muellerzr]

@akkasi

pred = torch.zeros(outputs['logits'].shape)

You need to move this to the GPU