Convert ndarray of tensors into tuple of tensors with gradients

I have a nd array of tensors with gradeints:
lis = [0,1,2]
ae = torch.tensor(lis, dtype = torch.float, requires_grad=True)
aqw =np.array([ae[0], ae[1],ae[2]])

I need to convert into a tensor with gradients intact. I tried torch.from_numpy but it losses the grads and I want to have a use case like this, I do not want to perform my computations in torch but in numpy

Numpy does not provide an autograd engine, so you won’t be able to automatically differentiate your operations in numpy.
If you want to use numpy methods in the forward pass, you could implement the backward manually by writing a custom torch.autograd.Function as described here.