In the example code for extending the autograd (link), in the backward function, grad_output is a Variable, but all the saved variables input, weight, bias are tensors since they got converted before the call of forward. If we invoke mm with mixture of Tensor and Variable, we will get a RuntimeError: mm(): argument 'mat2' (position 1) must be Variable, not torch.FloatTensor error. Should we convert grad_output to Tensor first and then convert the results back to Variable in backward?