Daily rc sync to master (#6771)

Co-authored-by: PietropaoloFrisoni <[email protected]>
Co-authored-by: Christina Lee <[email protected]>
Co-authored-by: GitHub Actions Bot <>

doc/releases/changelog-0.40.0.md

@@ -625,6 +625,9 @@ same information.
 
 <h3>Bug fixes 🐛</h3>
 
+* Adds validation so the device vjp is only used when the device actually supports it.
+  [(#6755)](https://github.com/PennyLaneAI/pennylane/pull/6755/)
+
 * `qml.counts` returns all outcomes when the `all_outcomes` argument is `True` and mid-circuit measurements are present.
   [(#6732)](https://github.com/PennyLaneAI/pennylane/pull/6732)
 

pennylane/workflow/resolution.py

@@ -86,9 +86,7 @@ def _use_tensorflow_autograph():
     return not tf.executing_eagerly()
 
 
-def _resolve_interface(
-    interface: Union[str, Interface, None], tapes: QuantumScriptBatch
-) -> Interface:
+def _resolve_interface(interface: Union[str, Interface], tapes: QuantumScriptBatch) -> Interface:
     """Helper function to resolve an interface based on a set of tapes.
 
     Args:
@@ -249,22 +247,24 @@ def _resolve_execution_config(
     ):
         updated_values["grad_on_execution"] = False
 
-    if execution_config.use_device_jacobian_product and isinstance(
-        device, qml.devices.LegacyDeviceFacade
-    ):
-        raise qml.QuantumFunctionError(
-            "device provided jacobian products are not compatible with the old device interface."
-        )
-
     if (
         "lightning" in device.name
-        and (transform_program and qml.metric_tensor in transform_program)
+        and transform_program
+        and qml.metric_tensor in transform_program
         and execution_config.gradient_method == "best"
     ):
         execution_config = replace(execution_config, gradient_method=qml.gradients.param_shift)
-
     execution_config = _resolve_diff_method(execution_config, device, tape=tapes[0])
 
+    if execution_config.use_device_jacobian_product and not device.supports_vjp(
+        execution_config, tapes[0]
+    ):
+        raise qml.QuantumFunctionError(
+            f"device_vjp=True is not supported for device {device},"
+            f" diff_method {execution_config.gradient_method},"
+            " and the provided circuit."
+        )
+
     if execution_config.gradient_method is qml.gradients.param_shift_cv:
         updated_values["gradient_keyword_arguments"]["dev"] = device
 

tests/test_qnode.py

@@ -873,6 +873,7 @@ def circuit(x, y):
 
         assert np.allclose(res, expected, atol=tol, rtol=0)
 
+    # pylint: disable=too-many-positional-arguments
     @pytest.mark.parametrize("dev_name", ["default.qubit", "default.mixed"])
     @pytest.mark.parametrize("first_par", np.linspace(0.15, np.pi - 0.3, 3))
     @pytest.mark.parametrize("sec_par", np.linspace(0.15, np.pi - 0.3, 3))
@@ -1170,6 +1171,27 @@ def decomposition(self) -> list:
         res = qml.execute([tape], dev)
         assert qml.math.get_interface(res) == "numpy"
 
+    def test_error_device_vjp_unsuppoprted(self):
+        """Test that an error is raised in the device_vjp is unsupported."""
+
+        class DummyDev(qml.devices.Device):
+
+            def execute(self, circuits, execution_config=qml.devices.ExecutionConfig()):
+                return 0
+
+            def supports_derivatives(self, execution_config=None, circuit=None):
+                return execution_config and execution_config.gradient_method == "vjp_grad"
+
+            def supports_vjp(self, execution_config=None, circuit=None) -> bool:
+                return execution_config and execution_config.gradient_method == "vjp_grad"
+
+        @qml.qnode(DummyDev(), diff_method="parameter-shift", device_vjp=True)
+        def circuit():
+            return qml.expval(qml.Z(0))
+
+        with pytest.raises(qml.QuantumFunctionError, match="device_vjp=True is not supported"):
+            circuit()
+
 
 class TestShots:
     """Unit tests for specifying shots per call."""

tests/workflow/test_resolve_execution_config.py

@@ -116,3 +116,30 @@ def test_jax_jit_interface():
     expected_mcm_config = MCMConfig(mcm_method="deferred", postselect_mode="fill-shots")
 
     assert resolved_config.mcm_config == expected_mcm_config
+
+
+# pylint: disable=unused-argument
+def test_no_device_vjp_if_not_supported():
+    """Test that an error is raised for device_vjp=True if the device does not support it."""
+
+    class DummyDev(qml.devices.Device):
+
+        def execute(self, circuits, execution_config=qml.devices.ExecutionConfig()):
+            return 0
+
+        def supports_derivatives(self, execution_config=None, circuit=None):
+            return execution_config and execution_config.gradient_method == "vjp_grad"
+
+        def supports_vjp(self, execution_config=None, circuit=None) -> bool:
+            return execution_config and execution_config.gradient_method == "vjp_grad"
+
+    config_vjp_grad = ExecutionConfig(use_device_jacobian_product=True, gradient_method="vjp_grad")
+    tape = qml.tape.QuantumScript()
+    # no error
+    _ = _resolve_execution_config(config_vjp_grad, DummyDev(), (tape,))
+
+    config_parameter_shift = ExecutionConfig(
+        use_device_jacobian_product=True, gradient_method="parameter-shift"
+    )
+    with pytest.raises(qml.QuantumFunctionError, match="device_vjp=True is not supported"):
+        _resolve_execution_config(config_parameter_shift, DummyDev(), (tape,))