Merge branch 'master' into parameter-broadcasting-1

doc/releases/changelog-dev.md

@@ -309,6 +309,9 @@
   instead of the controlled version of the diagonal unitary.
   [(#2525)](https://github.com/PennyLaneAI/pennylane/pull/2525)
 
+* Updated the gradients fix [(#2485)](https://github.com/PennyLaneAI/pennylane/pull/2485) to only apply to the `strawberryfields.gbs` device, since
+  the original logic was breaking some devices. [(#2595)](https://github.com/PennyLaneAI/pennylane/pull/2595)
+
 <h3>Deprecations</h3>
 
 <h3>Documentation</h3>
@@ -328,5 +331,4 @@
 This release contains contributions from (in alphabetical order):
 
 Amintor Dusko, Chae-Yeun Park, Christian Gogolin, Christina Lee, David Wierichs, Edward Jiang, Guillermo Alonso-Linaje,
-Jay Soni, Juan Miguel Arrazola, Maria Schuld, Mikhail Andrenkov, Soran Jahangiri, Utkarsh Azad
-
+Jay Soni, Juan Miguel Arrazola, Maria Schuld, Mikhail Andrenkov, Samuel Banning, Soran Jahangiri, Utkarsh Azad

pennylane/interfaces/autograd.py

@@ -234,11 +234,10 @@ def grad_fn(dy):
         return_vjps = [
             qml.math.to_numpy(v, max_depth=_n) if isinstance(v, ArrayBox) else v for v in vjps
         ]
-        if device.capabilities().get("provides_jacobian", False):
-            # in the case where the device provides the jacobian,
-            # the output of grad_fn must be wrapped in a tuple in
-            # order to match the input parameters to _execute.
-            return (return_vjps,)
+        if device.short_name == "strawberryfields.gbs":  # pragma: no cover
+            # TODO: remove this exceptional case once the source of this issue
+            # https://github.com/PennyLaneAI/pennylane-sf/issues/89 is determined
+            return (return_vjps,)  # pragma: no cover
         return return_vjps
 
     return grad_fn

tests/interfaces/test_autograd.py

@@ -1141,3 +1141,37 @@ def circuit(v):
 
         d_out = d_circuit(params)
         assert np.allclose(d_out, np.array([1.0, 2.0, 3.0, 4.0]))
+
+    def test_custom_jacobians_2(self):
+        """Test computing the gradient using the parameter-shift
+        rule with a device that provides a jacobian"""
+
+        class MyQubit(DefaultQubit):
+            @classmethod
+            def capabilities(cls):
+                capabilities = super().capabilities().copy()
+                capabilities.update(
+                    provides_jacobian=True,
+                )
+                return capabilities
+
+            def jacobian(self, *args, **kwargs):
+                raise NotImplementedError()
+
+        dev = MyQubit(wires=2)
+
+        @qml.qnode(dev, diff_method="parameter-shift", mode="backward")
+        def qnode(a, b):
+            qml.RY(a, wires=0)
+            qml.RX(b, wires=1)
+            qml.CNOT(wires=[0, 1])
+            return [qml.expval(qml.PauliZ(0)), qml.expval(qml.PauliY(1))]
+
+        a = np.array(0.1, requires_grad=True)
+        b = np.array(0.2, requires_grad=True)
+
+        res = qml.jacobian(qnode)(a, b)
+        expected = ([-np.sin(a), np.sin(a) * np.sin(b)], [0, -np.cos(a) * np.cos(b)])
+
+        assert np.allclose(res[0], expected[0])
+        assert np.allclose(res[1], expected[1])