🐛 Bug

torch.round() seems to be rounding differently for CPU and GPU tensors.
On CPU tensors it rounds half-to-even and on GPU it rounds away-from-zero as shown in the example below.

To Reproduce

CPU rounding:
>>> torch.Tensor([1.5, 2.5, 3.5, 4.5]).round()
tensor([2., 2., 4., 4.])

GPU rounding:
>>> torch.Tensor([1.5, 2.5, 3.5, 4.5]).cuda().round()
tensor([2., 3., 4., 5.], device=‘cuda:0’)

Expected behavior

I expected the function to behave the same on CPU/GPU, either returning [2., 2., 4., 4.] or [2., 3., 4., 5.] on both commands.
It could also be nice to have a way to control the rounding method in future releases, but that's a different issue.

Environment

• PyTorch Version (e.g., 1.0): 1.0
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source):
• Python version: 3.5.2
• CUDA/cuDNN version: 9
• GPU models and configuration: GeForce GTX 1060
• Any other relevant information:

Additional context