Investigating and Enabling Autograd Support

[Bmete7]

We need to first investigate, if MQSSPennylaneAdapter is able to run in Autograd mode and can use parameter-shift method:

import pennylane as qml
from pennylane import numpy as np
from mqss.pennylane_adapter.device import MQSSPennylaneDevice

dev = MQSSPennylaneDevice(wires=2, token=MQSS_TOKEN, backends=MQSS_BACKENDS)

@qml.qnode(dev, interface="autograd", diff_method="parameter-shift", shots = 1024)
def quantum_function_autograd(x, y):
    qml.RZ(x, wires=0)
    qml.CNOT(wires=[0, 1])
    qml.RY(y, wires=1)
    qml.CNOT(wires=[1, 0])
    qml.RX(x, wires=1)
    return qml.expval(qml.PauliX(0) @ qml.PauliZ(1)) # Here, qubits 0 and 1 are being measured, in bases X and Z, respectively. 

params = np.array([0.1, 0.2], requires_grad=True)
results = qml.gradients.param_shift(quantum_function_autograd)(params[0], params[1])

To-Do's:

• Figuring out whether a scheme similar to described above works in latest stable version of MQSS Pennylane Adapter
• If it can't, implement the required changes on the device level.

[Bmete7]

FYI @annkay108
Additional bug tracing and user reports shared:

I just tried to run my QNN with the updated MQSS-Pennylane adapter. The good news is that the forward pass on the device now worked fine. 😊
On the other side, unfortunately, I still get an error related to the backward pass, which apparently fails at a null_postprocessing check. See the full error trace below.
As you already mentioned, this concerns the parameter-shift method.

Some more details:
I am using the device
@qml.qnode(

MQSSPennylaneDevice(wires=2, token=os.getenv("<LRZ_API_TOKEN>"), backends=”EQE1”),
interface="torch",
diff_method="parameter-shift"
)

I verified, that after one forward pass, the output of my quantum model looks like this:

 ` [tensor(-0.9400, dtype=torch.float64, grad_fn=<ExecuteTapesBackward>), tensor(0.5600, dtype=torch.float64, grad_fn=<ExecuteTapesBackward>)].`

Meaning that for gradient calculation in backpropagation, it will call Pennylanes ExecuteTapesBackward, which is correct (no gradient call assignment missing).
The list of post-processing results that pennylane checks in the backward pass are apparently empty.
Since it works with the pennylane simulator (qml.device("default.qubit")), there must be then an issue in the adapter about passing on the gradients?

Do you have an idea, what is the problem here?