Issue description

When I take derivative of an element of an expanded tensor (or if such expression is a part of the graph) the derivative is incorrect.

Code example

import torch
from torch import Tensor
from torch.nn import Parameter

x = Parameter(Tensor([1]))
y = x.expand([100])
print(torch.autograd.grad(y[0], x)[0].item())
# outputs 100

The expected result is 1

System Info

PyTorch version: 0.4.0
Is debug build: No
CUDA used to build PyTorch: 8.0.61

OS: Debian GNU/Linux 9.3 (stretch)
GCC version: (Debian 6.3.0-18+deb9u1) 6.3.0 20170516
CMake version: version 3.7.2

Python version: 3.5
Is CUDA available: Yes
CUDA runtime version: 9.0.176
GPU models and configuration:
GPU 0: GeForce GTX 1080
GPU 1: GeForce GTX 1080

Nvidia driver version: 384.98
cuDNN version: Probably one of the following:
/usr/local/cuda-7.5/lib64/libcudnn.so.7.0.64
/usr/local/cuda-7.5/lib64/libcudnn_static.a
/usr/local/cuda-9.0/lib64/libcudnn.so.7.0.3
/usr/local/cuda-9.0/lib64/libcudnn_static.a
/usr/local/lib/python3.5/dist-packages/torch/lib/libcudnn-900fef33.so.7.0.5
/usr/local/matlab90/bin/glnxa64/libcudnn.so.7.0.64