ff.py

def backward(self, gradient=None):
    if not self.requires_grad:
        return
    
    if gradient is None:
        if self.shape == [1]:
            gradient = Tensor([1]) # dx/dx = 1 case
        else:
            raise RuntimeError("Gradient argument must be specified for non-scalar tensors.")

    if self.grad is None:
        self.grad = gradient

    else:
        self.grad += gradient

    if self.grad_fn is not None: # not a leaf
        grads = self.grad_fn.backward(gradient) # call the operation backward
        for tensor, grad in zip(self.grad_fn.input, grads):
            if isinstance(tensor, Tensor):
                tensor.backward(grad) # recursively call the backward again for the gradient expression (chain rule)