Run out of memory

Test code:

```
encoder = nn.GRU(100, 100)
encoder.cuda()

hidden = Variable(torch.randn(1, 1, 100)).cuda()
output = ''

num = 0
while True:
    output, hidden = encoder(hidden, hidden)
    print num + 1
    num +=1
```


The GPU memory usage keeps growing, and ends up in out of memory.