Issue description

when it use torch._C._nn.upsample_nearest2d(input, _output_size(2)), it's comes with

RuntimeError: CUDA error: invalid configuration argument.

The problem is only occur in torch-nightly( 1.2.0.dev20190702).

Error log

File "/home/xdjf/XXXX/XXXXX/XXXXXXXX.py", line 74, in _upsample_add
return F.interpolate(x, size=(H,W), mode='nearest') + y
File "/home/XXXX/.conda/envs/py36torch120/lib/python3.6/site-packages/torch/nn/functional.py", line 2545, in interpolate
return torch._C._nn.upsample_nearest2d(input, _output_size(2))
RuntimeError: CUDA error: invalid configuration argument

You can get the script and run it with:
torch-nightly( 1.2.0.dev20190702)

import torch
import torch.nn.functional as F

x=torch.rand(4,256,1,1).cuda()
y=torch.rand(4,256,3,3).cuda()

F.interpolate(x, size=(3,3), mode='nearest') + y

• PyTorch:
• conda install pytorch-nightly cudatoolkit=10.0 -c pytorch
• OS:Ubuntu16.04 LTS
• PyTorch version: torch-nightly( 1.2.0.dev20190702)
• Python version:3.6
• CUDA/cuDNN version:9.0 or 10.0
• GPU models and configuration: GTX 1080ti * 4