unsupported Python type error from set_default_tensor_type()

Hi,
I've updated pytorch to the lastest master (`a54407a7`) from github, but got an error when I use `torch.set_default_tensor_type("torch.cuda.FloatTensor")`:
```
terminate called after throwing an instance of 'std::invalid_argument'
  what():  unsupported Python type
Aborted (core dumped)
```

What to do for workaround? Thanks!
Jin