Bug: Inconsistence of the channel of `input` and that of `self.gradInput` can also work!

[Naruto-Sasuke]

When building a custom layer, it turns out that the channel of self.gradInput can be larger than the input. For example:

local myLayer, parent = torch.class('nn.MyLayer', 'nn.Module')

function myLayer:__init()
    parent.__init(self)
end

function myLayer:updateOutput(input)
    self.c = input:size(2)
    self.output = input:repeatTensor(1, self.c*2, 1, 1)
    return self.output
end

function myLayer:updateGradInput(input, gradOutput)
       self.gradInput = gradOutput
      return self.gradInput  -- It works fine.
end

However, self.gradInput is twice the channel of input. Pytorch requires that self.gradInput must match the size of input, or it errors! Does Torch cut out self.gradInput to be the same size of input internal ? If so, how does it crop?