Clean up how interface is handled in QNode and qml.execute

doc/releases/changelog-dev.md

@@ -7,18 +7,14 @@
 <h3>Improvements 🛠</h3>
 
 * `qml.qchem.excitations` now optionally returns fermionic operators.
-  [(#6171)](https://github.com/PennyLaneAI/pennylane/pull/6171)
+ [(#6171)](https://github.com/PennyLaneAI/pennylane/pull/6171)
 
 * The `diagonalize_measurements` transform now uses a more efficient method of diagonalization 
  when possible, based on the `pauli_rep` of the relevant observables.
  [#6113](https://github.com/PennyLaneAI/pennylane/pull/6113/)
 
-<h4>Capturing and representing hybrid programs</h4>
-
-* Differentiation of hybrid programs via `qml.grad` can now be captured into plxpr.
- When evaluating a captured `qml.grad` instruction, it will dispatch to `jax.grad`,
- which differs from the Autograd implementation of `qml.grad` itself.
- [(#6120)](https://github.com/PennyLaneAI/pennylane/pull/6120)
+* The `Hermitian` operator now has a `compute_sparse_matrix` implementation.
+ [(#6225)](https://github.com/PennyLaneAI/pennylane/pull/6225)
 
 <h4>Capturing and representing hybrid programs</h4>
 
@@ -102,12 +98,19 @@
 * The ``qml.Qubitization`` template now orders the ``control`` wires first and the ``hamiltonian`` wires second, which is the expected according to other templates.
  [(#6229)](https://github.com/PennyLaneAI/pennylane/pull/6229)
 
-* <h3>Contributors ✍️</h3>
+* Fixes a bug where a circuit using the `autograd` interface sometimes returns nested values that are not of the `autograd` interface.
+ [(#6225)](https://github.com/PennyLaneAI/pennylane/pull/6225)
+
+* Fixes a bug where a simple circuit with no parameters or only builtin/numpy arrays as parameters returns autograd tensors.
+ [(#6225)](https://github.com/PennyLaneAI/pennylane/pull/6225)
+
+<h3>Contributors ✍️</h3>
 
 This release contains contributions from (in alphabetical order):
 
 Guillermo Alonso,
 Utkarsh Azad,
+Astral Cai,
 Lillian M. A. Frederiksen,
 Christina Lee,
 William Maxwell,

pennylane/devices/execution_config.py

@@ -17,7 +17,7 @@
 from dataclasses import dataclass, field
 from typing import Optional, Union
 
-from pennylane.workflow import SUPPORTED_INTERFACES
+from pennylane.workflow import SUPPORTED_INTERFACE_NAMES
 
 
 @dataclass
@@ -110,9 +110,9 @@ def __post_init__(self):
 
  Note that this hook is automatically called after init via the dataclass integration.
  """
- if self.interface not in SUPPORTED_INTERFACES:
+ if self.interface not in SUPPORTED_INTERFACE_NAMES:
  raise ValueError(
- f"Unknown interface. interface must be in {SUPPORTED_INTERFACES}, got {self.interface} instead."
+ f"Unknown interface. interface must be in {SUPPORTED_INTERFACE_NAMES}, got {self.interface} instead."
  )
 
  if self.grad_on_execution not in {True, False, None}:

pennylane/devices/legacy_facade.py

@@ -25,6 +25,7 @@
 import pennylane as qml
 from pennylane.measurements import MidMeasureMP, Shots
 from pennylane.transforms.core.transform_program import TransformProgram
+from pennylane.workflow.execution import INTERFACE_MAP
 
 from .device_api import Device
 from .execution_config import DefaultExecutionConfig
@@ -328,18 +329,18 @@ def _create_temp_device(self, batch):
  if interface == "numpy":
  return self._device
 
- mapped_interface = qml.workflow.execution.INTERFACE_MAP.get(interface, interface)
+ interface = INTERFACE_MAP.get(interface, interface)
 
  backprop_interface = self._device.capabilities().get("passthru_interface", None)
- if mapped_interface == backprop_interface:
+ if interface == backprop_interface:
  return self._device
 
  backprop_devices = self._device.capabilities().get("passthru_devices", None)
 
  if backprop_devices is None:
  raise qml.DeviceError(f"Device {self} does not support backpropagation.")
 
- if backprop_devices[mapped_interface] == self._device.short_name:
+ if backprop_devices[interface] == self._device.short_name:
  return self._device
 
  if self.target_device.short_name != "default.qubit.legacy":
@@ -367,7 +368,7 @@ def _create_temp_device(self, batch):
  )
  # we already warned about backprop device switching
  new_device = qml.device(
- backprop_devices[mapped_interface],
+ backprop_devices[interface],
  wires=self._device.wires,
  shots=self._device.shots,
  ).target_device
@@ -396,25 +397,24 @@ def _validate_backprop_method(self, tape):
  return False
  params = tape.get_parameters(trainable_only=False)
  interface = qml.math.get_interface(*params)
+ if interface != "numpy":
+ interface = INTERFACE_MAP.get(interface, interface)
 
  if tape and any(isinstance(m.obs, qml.SparseHamiltonian) for m in tape.measurements):
  return False
- if interface == "numpy":
- interface = None
- mapped_interface = qml.workflow.execution.INTERFACE_MAP.get(interface, interface)
 
  # determine if the device supports backpropagation
  backprop_interface = self._device.capabilities().get("passthru_interface", None)
 
  if backprop_interface is not None:
  # device supports backpropagation natively
- return mapped_interface in [backprop_interface, "Numpy"]
+ return interface in [backprop_interface, "numpy"]
  # determine if the device has any child devices that support backpropagation
  backprop_devices = self._device.capabilities().get("passthru_devices", None)
 
  if backprop_devices is None:
  return False
- return mapped_interface in backprop_devices or mapped_interface == "Numpy"
+ return interface in backprop_devices or interface == "numpy"
 
  def _validate_adjoint_method(self, tape):
  # The conditions below provide a minimal set of requirements that we can likely improve upon in

pennylane/devices/qubit/simulate.py

@@ -922,7 +922,7 @@ def _(original_measurement: ExpectationMP, measures): # pylint: disable=unused-
  for v in measures.values():
  if not v[0] or v[1] is tuple():
  continue
- cum_value += v[0] * v[1]
+ cum_value += qml.math.multiply(v[0], v[1])
  total_counts += v[0]
  return cum_value / total_counts
 
@@ -935,7 +935,7 @@ def _(original_measurement: ProbabilityMP, measures): # pylint: disable=unused-
  for v in measures.values():
  if not v[0] or v[1] is tuple():
  continue
- cum_value += v[0] * v[1]
+ cum_value += qml.math.multiply(v[0], v[1])
  total_counts += v[0]
  return cum_value / total_counts
 

pennylane/ops/qubit/observables.py

@@ -137,6 +137,10 @@ def compute_matrix(A: TensorLike) -> TensorLike: # pylint: disable=arguments-di
  Hermitian._validate_input(A)
  return A
 
+ @staticmethod
+ def compute_sparse_matrix(A) -> csr_matrix: # pylint: disable=arguments-differ
+ return csr_matrix(Hermitian.compute_matrix(A))
+
  @property
  def eigendecomposition(self) -> dict[str, TensorLike]:
  """Return the eigendecomposition of the matrix specified by the Hermitian observable.

pennylane/workflow/__init__.py

@@ -56,6 +56,6 @@
 
 """
 from .construct_batch import construct_batch, get_transform_program
-from .execution import INTERFACE_MAP, SUPPORTED_INTERFACES, execute
+from .execution import INTERFACE_MAP, SUPPORTED_INTERFACE_NAMES, execute
 from .qnode import QNode, qnode
 from .set_shots import set_shots

pennylane/workflow/execution.py

@@ -51,12 +51,9 @@
  "autograd",
  "numpy",
  "torch",
- "pytorch",
  "jax",
- "jax-python",
  "jax-jit",
  "tf",
- "tensorflow",
 }
 
 SupportedInterfaceUserInput = Literal[
@@ -78,30 +75,29 @@
 ]
 
 _mapping_output = (
- "Numpy",
+ "numpy",
  "auto",
  "autograd",
  "autograd",
  "numpy",
  "jax",
- "jax",
+ "jax-jit",
  "jax",
  "jax",
  "torch",
  "torch",
  "tf",
  "tf",
- "tf",
- "tf",
+ "tf-autograph",
+ "tf-autograph",
 )
+
 INTERFACE_MAP = dict(zip(get_args(SupportedInterfaceUserInput), _mapping_output))
 """dict[str, str]: maps an allowed interface specification to its canonical name."""
 
-#: list[str]: allowed interface strings
-SUPPORTED_INTERFACES = list(INTERFACE_MAP)
+SUPPORTED_INTERFACE_NAMES = list(INTERFACE_MAP)
 """list[str]: allowed interface strings"""
 
-
 _CACHED_EXECUTION_WITH_FINITE_SHOTS_WARNINGS = (
  "Cached execution with finite shots detected!\n"
  "Note that samples as well as all noisy quantities computed via sampling "
@@ -135,41 +131,41 @@ def _get_ml_boundary_execute(
  pennylane.QuantumFunctionError if the required package is not installed.
 
  """
- mapped_interface = INTERFACE_MAP[interface]
  try:
- if mapped_interface == "autograd":
+ if interface == "autograd":
  from .interfaces.autograd import autograd_execute as ml_boundary
 
- elif mapped_interface == "tf":
- if "autograph" in interface:
- from .interfaces.tensorflow_autograph import execute as ml_boundary
+ elif interface == "tf-autograph":
+ from .interfaces.tensorflow_autograph import execute as ml_boundary
 
-  ml_boundary = partial(ml_boundary, grad_on_execution=grad_on_execution)
+ ml_boundary = partial(ml_boundary, grad_on_execution=grad_on_execution)
 
-  else:
-  from .interfaces.tensorflow import tf_execute as full_ml_boundary
+ elif interface == "tf":
+ from .interfaces.tensorflow import tf_execute as full_ml_boundary
 
-  ml_boundary = partial(full_ml_boundary, differentiable=differentiable)
+ ml_boundary = partial(full_ml_boundary, differentiable=differentiable)
 
- elif mapped_interface == "torch":
+ elif interface == "torch":
  from .interfaces.torch import execute as ml_boundary
 
  elif interface == "jax-jit":
  if device_vjp:
  from .interfaces.jax_jit import jax_jit_vjp_execute as ml_boundary
  else:
  from .interfaces.jax_jit import jax_jit_jvp_execute as ml_boundary
- else: # interface in {"jax", "jax-python", "JAX"}:
+
+ else: # interface is jax
  if device_vjp:
  from .interfaces.jax_jit import jax_jit_vjp_execute as ml_boundary
  else:
  from .interfaces.jax import jax_jvp_execute as ml_boundary
 
  except ImportError as e: # pragma: no-cover
  raise qml.QuantumFunctionError(
- f"{mapped_interface} not found. Please install the latest "
- f"version of {mapped_interface} to enable the '{mapped_interface}' interface."
+ f"{interface} not found. Please install the latest "
+ f"version of {interface} to enable the '{interface}' interface."
  ) from e
+
  return ml_boundary
 
 
@@ -263,12 +259,22 @@ def _get_interface_name(tapes, interface):
 
  Returns:
  str: Interface name"""
+
+ if interface not in SUPPORTED_INTERFACE_NAMES:
+ raise qml.QuantumFunctionError(
+ f"Unknown interface {interface}. Interface must be one of {SUPPORTED_INTERFACE_NAMES}."
+ )
+
+ interface = INTERFACE_MAP[interface]
+
  if interface == "auto":
  params = []
  for tape in tapes:
  params.extend(tape.get_parameters(trainable_only=False))
  interface = qml.math.get_interface(*params)
- if INTERFACE_MAP.get(interface, "") == "tf" and _use_tensorflow_autograph():
+ if interface != "numpy":
+ interface = INTERFACE_MAP[interface]
+ if interface == "tf" and _use_tensorflow_autograph():
  interface = "tf-autograph"
  if interface == "jax":
  try: # pragma: no cover
@@ -439,6 +445,7 @@ def cost_fn(params, x):
 
  ### Specifying and preprocessing variables ####
 
+ _interface_user_input = interface
  interface = _get_interface_name(tapes, interface)
  # Only need to calculate derivatives with jax when we know it will be executed later.
  if interface in {"jax", "jax-jit"}:
@@ -460,7 +467,11 @@ def cost_fn(params, x):
  )
 
  # Mid-circuit measurement configuration validation
- mcm_interface = interface or _get_interface_name(tapes, "auto")
+ # If the user specifies `interface=None`, regular execution considers it numpy, but the mcm
+ # workflow still needs to know if jax-jit is used
+ mcm_interface = (
+ _get_interface_name(tapes, "auto") if _interface_user_input is None else interface
+ )
  finite_shots = any(tape.shots for tape in tapes)
  _update_mcm_config(config.mcm_config, mcm_interface, finite_shots)
 
@@ -479,12 +490,12 @@ def cost_fn(params, x):
  cache = None
 
  # changing this set of conditions causes a bunch of tests to break.
- no_interface_boundary_required = interface is None or config.gradient_method in {
+ no_interface_boundary_required = interface == "numpy" or config.gradient_method in {
  None,
  "backprop",
  }
  device_supports_interface_data = no_interface_boundary_required and (
- interface is None
+ interface == "numpy"
  or config.gradient_method == "backprop"
  or getattr(device, "short_name", "") == "default.mixed"
  )
@@ -497,9 +508,9 @@ def cost_fn(params, x):
  numpy_only=not device_supports_interface_data,
  )
 
- # moved to its own explicit step so it will be easier to remove
+ # moved to its own explicit step so that it will be easier to remove
  def inner_execute_with_empty_jac(tapes, **_):
- return (inner_execute(tapes), [])
+ return inner_execute(tapes), []
 
  if interface in jpc_interfaces:
  execute_fn = inner_execute
@@ -522,7 +533,7 @@ def inner_execute_with_empty_jac(tapes, **_):
  and getattr(device, "short_name", "") in ("lightning.gpu", "lightning.kokkos")
  and interface in jpc_interfaces
  ): # pragma: no cover
- if INTERFACE_MAP[interface] == "jax" and "use_device_state" in gradient_kwargs:
+ if "jax" in interface and "use_device_state" in gradient_kwargs:
  gradient_kwargs["use_device_state"] = False
 
  jpc = LightningVJPs(device, gradient_kwargs=gradient_kwargs)
@@ -563,7 +574,7 @@ def execute_fn(internal_tapes) -> tuple[ResultBatch, tuple]:
  config: the ExecutionConfig that specifies how to perform the simulations.
  """
  numpy_tapes, _ = qml.transforms.convert_to_numpy_parameters(internal_tapes)
- return (device.execute(numpy_tapes, config), tuple())
+ return device.execute(numpy_tapes, config), tuple()
 
  def gradient_fn(internal_tapes):
  """A partial function that wraps compute_derivatives method of the device.
@@ -612,7 +623,7 @@ def gradient_fn(internal_tapes):
 
  # trainable parameters can only be set on the first pass for jax
  # not higher order passes for higher order derivatives
- if interface in {"jax", "jax-python", "jax-jit"}:
+ if "jax" in interface:
  for tape in tapes:
  params = tape.get_parameters(trainable_only=False)
  tape.trainable_params = qml.math.get_trainable_indices(params)