Fix state prep operation decomposition with LightningQubit (#661)

* Fixed LQ adjoint decomp

* Auto update version

* Trigger CI

* Apply suggestions from code review

Co-authored-by: Ali Asadi <[email protected]>

* Auto update version

* trigger ci

* trigger ci

* Added fix for MCM with adjoint

* Fixed LQ tests

---------

Co-authored-by: Dev version update bot <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Ali Asadi <[email protected]>
Co-authored-by: Vincent Michaud-Rioux <[email protected]>
Co-authored-by: Vincent Michaud-Rioux <[email protected]>

Author: 4 people

.github/CHANGELOG.md

@@ -50,6 +50,9 @@
 * Fix random `coverage xml` CI issues.
   [(#635)](https://github.com/PennyLaneAI/pennylane-lightning/pull/635)
 
+* `lightning.qubit` correctly decomposed state preparation operations with adjoint differentiation.
+  [(#661)](https://github.com/PennyLaneAI/pennylane-lightning/pull/661)
+
 ### Contributors
 
 This release contains contributions from (in alphabetical order):

pennylane_lightning/core/_version.py

@@ -16,4 +16,4 @@
    Version number (major.minor.patch[-label])
 """
 
-__version__ = "0.36.0-dev17"
+__version__ = "0.36.0-dev18"

pennylane_lightning/lightning_qubit/lightning_qubit.py

@@ -263,7 +263,7 @@ def _supports_adjoint(circuit):
 
     try:
         prog((circuit,))
-    except (qml.operation.DecompositionUndefinedError, qml.DeviceError):
+    except (qml.operation.DecompositionUndefinedError, qml.DeviceError, AttributeError):
         return False
     return True
 
@@ -282,7 +282,9 @@ def _add_adjoint_transforms(program: TransformProgram) -> None:
 
     name = "adjoint + lightning.qubit"
     program.add_transform(no_sampling, name=name)
-    program.add_transform(decompose, stopping_condition=adjoint_ops, name=name)
+    program.add_transform(
+        decompose, stopping_condition=adjoint_ops, name=name, skip_initial_state_prep=False
+    )
     program.add_transform(validate_observables, accepted_observables, name=name)
     program.add_transform(
         validate_measurements, analytic_measurements=adjoint_measurements, name=name

tests/new_api/test_device.py

@@ -86,6 +86,7 @@ def test_add_adjoint_transforms(self):
             decompose,
             stopping_condition=adjoint_ops,
             name=name,
+            skip_initial_state_prep=False,
         )
         expected_program.add_transform(validate_observables, accepted_observables, name=name)
         expected_program.add_transform(
@@ -248,6 +249,7 @@ def test_preprocess(self, adjoint):
                 decompose,
                 stopping_condition=adjoint_ops,
                 name=name,
+                skip_initial_state_prep=False,
             )
             expected_program.add_transform(validate_observables, accepted_observables, name=name)
             expected_program.add_transform(
@@ -554,6 +556,59 @@ def test_derivatives_no_trainable_params(self, dev, execute_and_derivatives, bat
         assert len(jac) == 1
         assert qml.math.shape(jac[0]) == (0,)
 
+    @pytest.mark.parametrize("execute_and_derivatives", [True, False])
+    @pytest.mark.parametrize(
+        "state_prep, params, wires",
+        [
+            (qml.BasisState, [1, 1], [0, 1]),
+            (qml.StatePrep, [0.0, 0.0, 0.0, 1.0], [0, 1]),
+            (qml.StatePrep, qml.numpy.array([0.0, 1.0]), [1]),
+        ],
+    )
+    @pytest.mark.parametrize(
+        "trainable_params",
+        [(0, 1, 2), (1, 2)],
+    )
+    def test_state_prep_ops(
+        self, dev, state_prep, params, wires, execute_and_derivatives, batch_obs, trainable_params
+    ):
+        """Test that a circuit containing state prep operations is differentiated correctly."""
+        qs = QuantumScript(
+            [state_prep(params, wires), qml.RX(1.23, 0), qml.CNOT([0, 1]), qml.RX(4.56, 1)],
+            [qml.expval(qml.PauliZ(1))],
+        )
+
+        config = ExecutionConfig(gradient_method="adjoint", device_options={"batch_obs": batch_obs})
+        program, new_config = dev.preprocess(config)
+        tapes, fn = program([qs])
+        tapes[0].trainable_params = trainable_params
+        if execute_and_derivatives:
+            res, jac = dev.execute_and_compute_derivatives(tapes, new_config)
+            res = fn(res)
+        else:
+            res, jac = (
+                fn(dev.execute(tapes, new_config)),
+                dev.compute_derivatives(tapes, new_config),
+            )
+
+        dev_ref = DefaultQubit(max_workers=1)
+        config = ExecutionConfig(gradient_method="adjoint")
+        program, new_config = dev_ref.preprocess(config)
+        tapes, fn = program([qs])
+        tapes[0].trainable_params = trainable_params
+        if execute_and_derivatives:
+            expected, expected_jac = dev_ref.execute_and_compute_derivatives(tapes, new_config)
+            expected = fn(expected)
+        else:
+            expected, expected_jac = (
+                fn(dev_ref.execute(tapes, new_config)),
+                dev_ref.compute_derivatives(tapes, new_config),
+            )
+
+        tol = 1e-5 if dev.c_dtype == np.complex64 else 1e-7
+        assert np.allclose(res, expected, atol=tol, rtol=0)
+        assert np.allclose(jac, expected_jac, atol=tol, rtol=0)
+
     def test_state_jacobian_not_supported(self, dev, batch_obs):
         """Test that an error is raised if derivatives are requested for state measurement"""
         qs = QuantumScript([qml.RX(1.23, 0)], [qml.state()], trainable_params=[0])