Question about torch.nn.DataParallel module in backward pass

Hi there,
I would like to use torch.nn.DataParallel for intermediate layers in my model. the structure looks like following:

class MyModel(nn.Module):
    def __init__(self):
        ... <some other layers>
        self.linear = nn.DataParallel(nn.Linear(100,100))
        self.linear2 = nn.DataParallel(nn.Linear(50,500))

    def forward(self, x):"cuda:0") # assume moved to gpu-0
        ... <some other operation on inputs>
        x = self.linear(x)
        x = self.linear2(x)
        return x

While I have several concerns:

  1. the output of the nn.DataParallel by default is the gpu-0, will it cause extract overhead? because of the memory copy among different GPUs, assume we have 8 GPUs for Data parallel.
  2. When we calculate the loss, if the loss value is calculated on GPU-0, how does intermediate nn.DataParallel layer do the right backward pass?
  3. does pytorch allow cross device backward pass? (e.g. I have a tensor A at cuda:0, later I copied A and moved it to cuda:1, B = A; A ='cuda:1'), later when I do backward on variable A, will the gradient computation propagate to B at cuda:0?