torch.ipynb

1 - the error comes from the fact that you modify in place a tensor for which the gradient is required.
2 - You can just by using another tensor do the same thing, as demonstrated below:

import torch
x = torch.tensor([[1., 2.], [3., 4.]], requires_grad=True)
z = torch.zeros_like(x)
for i in range(2):
    for j in range(2):
        z[i, j] = x[i, j] ** 2
y = torch.mean(x)
print(y)
y.backward()
x.grad

tensor(2.5000)
Out[6]:
tensor([[ 0.2500, 0.2500],
[ 0.2500, 0.2500]])