Parameter broadcast bug

PennyLane Help
3

Hey @Yan_Li! It doesn’t look like you’re using it, but we have a dedicated module in Pennylane for handling how Pennylane interfaces with Pytorch. You can check it out here: Turning quantum nodes into Torch Layers — PennyLane documentation

Currently, true parameter broadcasting within qnn.TorchLayer isn’t supported. I.e., you can still “broadcast”, but what happens under the hood is a serial execution, not parallel. So, it looks like broadcasting is happening, but really isn’t.

There’s a PR for this on our Github that will add support for true broadcasting if you’re interested in following its progress!

Regarding your specific error, adjoint doesn’t support parameter-broadcasting… so the error you’re getting is a little strange. In any case, we should have a more verbose and informative error message.

In the mean time, what you can do is:

  • Add qml.trasnforms.broadcast_expand to the qnode:
        @qml.transforms.broadcast_expand
        @qml.qnode(dev, interface="torch", diff_method=diff_method)
        def qnode(inputs,weights):
            # print("excute qnode ",weights.shape,inputs.shape)
            qml.AmplitudeEmbedding(features=inputs, wires=range(n_qubits),pad_with=0.,normalize=True) 
            # qml.Hadamard(wires=n_qubits-1)
            weights = weights.reshape(-1,2)
            qml.MPS(
                wires=range(n_qubits),
                n_block_wires=2,
                block=block,
                n_params_block=2,
                template_weights=weights,
            )
            r = qml.PauliZ(wires=n_qubits-1)
            # print("r ",r)
            return qml.expval(r)
  • switch to a different differentiation method (backprop)
  • switch to a different device that doesn’t think its should be supporting parameter broadcasting
1 Like