Skip to content

Release 0.43.0

Latest
Latest

Choose a tag to compare

View all tags
@andrijapau andrijapau released this 15 Oct 18:03
v0.43.0
5dc163b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

New features since last release

A brand new resource estimation module 📖

A new toolkit dedicated to resource estimation is now available in the pennylane.estimator module! The functionality therein is designed to rapidly and flexibly estimate the quantum resources required to execute programs written at different levels of abstraction. This new module includes the following features:

  • A new pennylane.estimator.estimate.estimate function allows users to estimate the quantum resources required to execute a circuit or operation with respect to a given gate set and configuration. (#8203) (#8205) (#8275) (#8227) (#8279) (#8288) (#8311) (#8313) (#8360)

    The pennylane.estimator.estimate.estimate function can be used on circuits written at different levels of detail to get high-level estimates of gate counts and additional wires fast. For workflows that are already defined in detail, like executable QNodes, the pennylane.estimator.estimate.estimate function works as follows:

    import pennylane as qml
import pennylane.estimator as qre

dev = qml.device("null.qubit")

@qml.qnode(dev)
def circ():
    for w in range(2):
        qml.Hadamard(wires=w)
    qml.CNOT(wires=[0,1])
    qml.RX(1.23*np.pi, wires=0)
    qml.RY(1.23*np.pi, wires=1)
    qml.QFT(wires=[0, 1, 2])
    return qml.state()
    >>> res = qre.estimate(circ)()
>>> print(res)
--- Resources: ---
  Total wires: 3
  algorithmic wires: 3
  allocated wires: 0
    zero state: 0
    any state: 0
  Total gates : 408
  'T': 396,
  'CNOT': 9,
  'Hadamard': 3

    If exact argument values and other details to operators are unknown or not available, pennylane.estimator.estimate.estimate can also be used on new lightweight representations of PennyLane operations that require minimal information to obtain high-level estimates. As part of this release, many operations in PennyLane now have a corresponding lightweight version that inherits from a new class called pennylane.estimator.resource_operator.ResourceOperator, which can be found in the pennylane.estimator module.

    For example, the lightweight representation of QFT is qre.QFT. By simply specifying the number of wires it acts on, we can obtain resource estimates:

    >>> qft = qre.QFT(num_wires=3)
>>> res = qre.estimate(qft)
>>> print(res)
--- Resources: ---
  Total wires: 3
    algorithmic wires: 3
    allocated wires: 0
      zero state: 0
      any state: 0
  Total gates : 408
  'T': 396,
  'CNOT': 9,
  'Hadamard': 3

    One can create a circuit comprising these operations with similar syntax as defining a QNode, but with far less detail. Here is an example of a circuit with 50 (logical) algorithmic qubits, which includes a pennylane.estimator.templates.QROMStatePreparation acting on 48 qubits. Defining this state preparation for execution would require a state vector of length $2^{48}$ (see qml.QROMStatePreparation), but we are able to estimate the required resources with only metadata, bypassing this computational barrier. Even at this scale, the resource estimate is computed in a fraction of a second!

    def my_circuit():
    qre.QROMStatePreparation(num_state_qubits=48)
    for w in range(2):
        qre.Hadamard(wires=w)
    qre.QROM(num_bitstrings=32, size_bitstring=8, restored=False)
    qre.CNOT(wires=[0,1])
    qre.RX(wires=0)
    qre.RY(wires=1)
    qre.QFT(num_wires=30)
    return
    >>> res = qre.estimate(my_circuit)()
>>> print(res)
--- Resources: ---
 Total wires: 129
  algorithmic wires: 50
  allocated wires: 79
    zero state: 71
    any state: 8
  Total gates : 2.702E+16
  'Toffoli': 1.126E+15,
  'T': 5.751E+4,
  'CNOT': 2.027E+16,
  'X': 2.252E+15,
  'Z': 32,
  'S': 64,
  'Hadamard': 3.378E+15

    Here is a summary of the lightweight operations made available in this release. A complete list can be found in the pennylane.estimator module.

    • pennylane.estimator.ops.Identity, pennylane.estimator.ops.GlobalPhase, and various non-parametric operators and single-qubit parametric operators. (#8240) (#8242) (#8302)
    • Various controlled single and multi qubit operators. (#8243)
    • pennylane.estimator.ops.Controlled, and pennylane.estimator.ops.Adjoint as symbolic operators. (#8252) (#8349)
    • pennylane.estimator.ops.Pow, pennylane.estimator.ops.Prod, pennylane.estimator.ops.ChangeOpBasis, and parametric multi-qubit operators. (#8255)
    • Templates including pennylane.estimator.templates.SemiAdder, pennylane.estimator.templates.QFT, pennylane.estimator.templates.AQFT, pennylane.estimator.templates.BasisRotation, pennylane.estimator.templates.Select, pennylane.estimator.templates.QROM, pennylane.estimator.templates.SelectPauliRot, pennylane.estimator.templates.QubitUnitary, pennylane.estimator.templates.ControlledSequence, pennylane.estimator.templates.QPE, pennylane.estimator.templates.IterativeQPE, pennylane.estimator.templates.MPSPrep, pennylane.estimator.templates.QROMStatePreparation, pennylane.estimator.templates.UniformStatePrep, pennylane.estimator.templates.AliasSampling, pennylane.estimator.templates.IntegerComparator, pennylane.estimator.templates.SingleQubitComparator, pennylane.estimator.templates.TwoQubitComparator, pennylane.estimator.templates.RegisterComparator, pennylane.estimator.templates.SelectTHC, pennylane.estimator.templates.PrepTHC, and pennylane.estimator.templates.QubitizeTHC. (#8300) (#8305) (#8309)

    For defining your own customized lightweight resource operations that integrate with features in the pennylane.estimator module, check out the documentation for pennylane.estimator.resource_operator.ResourceOperator.

  • Users can define customized configurations to be used during resource estimation using the new pennylane.estimator.resource_config.ResourceConfig class. This enables the seamless analysis of tradeoffs between resources required and quantities like individual gate precisions or different gate decompositions. (#8259)

    In the following example, a pennylane.estimator.resource_config.ResourceConfig is used to modify the default precision of single qubit rotations, and T counts are compared between different configurations.

    def my_circuit():
    qre.RX(wires=0)
    qre.RY(wires=1)
    qre.RZ(wires=2)
    return

my_rc = qre.ResourceConfig()
res1 = qre.estimate(my_circuit, config=my_rc)()
my_rc.set_single_qubit_rot_precision(1e-2)
res2 = qre.estimate(my_circuit, config=my_rc)()
    >>> t1 = res1.gate_counts['T']
>>> t2 = res2.gate_counts['T']
>>> print(t1, t2)
132 51

  • Hamiltonians are often both expensive to compute and to analyze, but the amount of information required to estimate the resources of Hamiltonian simulation can be surprisingly small in comparison. The pennylane.estimator.compact_hamiltonian.CDFHamiltonian, pennylane.estimator.compact_hamiltonian.THCHamiltonian, pennylane.estimator.compact_hamiltonian.VibronicHamiltonian, and pennylane.estimator.compact_hamiltonian.VibrationalHamiltonian classes were added to store the metadata of the Hamiltonian of a quantum system pertaining to resource estimation. In addition, several resource templates were added that are related to the Suzuki-Trotter method for Hamiltonian simulation, including pennylane.estimator.templates.TrotterProduct, pennylane.estimator.templates.TrotterCDF, pennylane.estimator.templates.TrotterTHC, pennylane.estimator.templates.TrotterVibronic, and pennylane.estimator.templates.TrotterVibrational. (#8303)

    Here's a simple example of resource estimation for the simulation of a pennylane.estimator.compact_hamiltonian.CDFHamiltonian, where we only need to specify two integer arguments (num_orbitals and num_fragments) to get resource estimates:

    >>> cdf_ham = qre.CDFHamiltonian(num_orbitals=4, num_fragments=4)
>>> res = qre.estimate(qre.TrotterCDF(cdf_ham, num_steps=1, order=2))
>>> print(res)
--- Resources: ---
  Total wires: 8
    algorithmic wires: 8
    allocated wires: 0
      zero state: 0
      any state: 0
  Total gates : 2.238E+4
  'T': 2.075E+4,
  'CNOT': 448,
  'Z': 336,
  'S': 504,
  'Hadamard': 336

  • In addition to the pennylane.estimator.resource_operator.ResourceOperator class mentioned above, the scalability of the resource estimation functionality in this release is owed to the following new internal classes:

    • pennylane.estimator.resources_base.Resources: A container for counts and other metadata of quantum resources. (#8205)
    • pennylane.estimator.resource_operator.GateCount: A class to represent a gate and its number of occurrences in a circuit or decomposition.
    • pennylane.estimator.resource_operator.CompressedResourceOp: A lightweight class corresponding to an operator type alongside its parameters. (#8227)
    • The pennylane.estimator.wires_manager.WireResourceManager, pennylane.estimator.wires_manager.Allocate, and pennylane.estimator.wires_manager.Deallocate classes, which were added to manage and track wire usage during resource estimation and within pennylane.estimator.resource_operator.ResourceOperator definitions. (#8203)

The resource estimation tools in the pennylane.estimator module were originally prototyped in the pennylane.labs module. Check it out too for the latest cutting-edge research functionality!

Dynamic wire allocation 🎁

  • Wires can now be dynamically allocated and deallocated in quantum functions with the pennylane.allocate and pennylane.deallocate functions. These features unlock many important applications that rely on smart and efficient handling of wires, such as decompositions of gates that require auxiliary wires and logical patterns in subroutines that benefit from having dynamic wire management.

    (#7718) (#8151) (#8163) (#8179) (#8198) (#8381)

    The pennylane.allocate function can accept three arguments that dictate how dynamically allocated wires are handled:

    • num_wires: the number of wires to dynamically allocate.
    • state = "zero"/"any": the initial state that the dynamically allocated wires are requested to be in. Currently, supported values are "zero" (initialize in the all-zero state) or "any" (any arbitrary state).
    • restored = True/False: a user-guarantee that the allocated wires will be restored to their original state (True) or not (False) when those wires are deallocated.

    The recommended way to safely allocate and deallocate wires is to use pennylane.allocate as a context manager:

    import pennylane as qml

@qml.qnode(qml.device("default.qubit"))
def circuit():
    qml.H(0)
    qml.H(1)

    with qml.allocate(2, state="zero", restored=False) as new_wires:
        qml.H(new_wires[0])
        qml.H(new_wires[1])

    return qml.expval(qml.Z(0))
    >>> print(qml.draw(circuit)())
            0: ──H───────────────────────┤  <Z>
            1: ──H───────────────────────┤
<DynamicWire>: ─╭Allocate──H─╭Deallocate─┤
<DynamicWire>: ─╰Allocate──H─╰Deallocate─┤

    As illustrated, using pennylane.allocate as a context manager ensures that allocation and safe deallocation are controlled within a localized scope. Equivalenty, pennylane.allocate can be used in-line along with pennylane.deallocate for manual handling:

    new_wires = qml.allocate(2, state="zero", restored=False)
qml.H(new_wires[0])
qml.H(new_wires[1])
qml.deallocate(new_wires)

    For more complex dynamic allocation in circuits, PennyLane will resolve the dynamic allocation calls in the most resource-efficient manner before sending the program to the device. Consider the following circuit, which contains two dynamic allocations within a for loop.

    @qml.qnode(qml.device("default.qubit"), mcm_method="tree-traversal")
def circuit():
    qml.H(0)

    for i in range(2):
        with qml.allocate(1, state="zero", restored=True) as new_qubit1:
            with qml.allocate(1, state="any", restored=False) as new_qubit2:
                m0 = qml.measure(new_qubit1[0], reset=True)
                qml.cond(m0 == 1, qml.Z)(new_qubit2[0])
                qml.CNOT((0, new_qubit2[0]))

    return qml.expval(qml.Z(0))
    >>> print(qml.draw(circuit)())
            0: ──H─────────────────────╭●───────────────────────╭●─────────────┤  <Z>
<DynamicWire>: ──Allocate──┤↗│  │0⟩────│──────────Deallocate────│──────────────┤
<DynamicWire>: ──Allocate───║────────Z─╰X─────────Deallocate────│──────────────┤
<DynamicWire>: ─────────────║────────║──Allocate──┤↗│  │0⟩──────│───Deallocate─┤
<DynamicWire>: ─────────────║────────║──Allocate───║──────────Z─╰X──Deallocate─┤
                            ╚════════╝             ╚══════════╝

    The user-level circuit drawing shows four separate allocations and deallocations (two per loop iteration). However, the circuit that the device receives gets automatically compiled to only use
    two additional wires (wires labelled 1 and 2 in the diagram below). This is due to the fact that new_qubit1 and new_qubit2 can both be reused after they've been deallocated in the first iteration of the for loop:

    >>> print(qml.draw(circuit, level="device")())
0: ──H───────────╭●──────────────╭●─┤  <Z>
1: ──┤↗│  │0⟩────│───┤↗│  │0⟩────│──┤
2: ───║────────Z─╰X───║────────Z─╰X─┤
      ╚════════╝      ╚════════╝

    Additionally, pennylane.allocate and pennylane.deallocate work with pennylane.qjit with some restrictions.

Resource tracking with Catalyst 🧾

  • Users can now use the pennylane.specs function to track the exact resources of programs compiled with pennylane.qjit! This new feature is currently only supported when using level="device". (#8202)

    from functools import partial

gateset = {qml.H, qml.S, qml.CNOT, qml.T, qml.RX, qml.RY, qml.RZ}

@qml.qjit
@partial(qml.transforms.decompose, gate_set=gateset)
@qml.qnode(qml.device("null.qubit", wires=100))
def circuit():
    qml.QFT(wires=range(100))
    qml.Hadamard(wires=0)
    qml.CNOT(wires=[0, 1])
    qml.OutAdder(
                x_wires=range(10),
                y_wires=range(10,20),
                output_wires=range(20,31)
                )
    return qml.expval(qml.Z(0) @ qml.Z(1))

circ_specs = qml.specs(circuit, level="device")()
    >>> print(circ_specs['resources'])
num_wires: 100
num_gates: 138134
depth: 90142
shots: Shots(total=None)
gate_types:
{'CNOT': 55313, 'RZ': 82698, 'Hadamard': 123}
gate_sizes:
{2: 55313, 1: 82821}

  • The pennylane.specs function now accepts a compute_depth keyword argument, which is set to True by default. Since depth computation is usually the most expensive resource to calculate, making it optional can increase the performance of pennylane.specs when depth is not a desired resource to calculate. (#7998) (#8042)

ZX Calculus transforms 🍪

  • A new set of transforms enable ZX calculus-based circuit optimization. These transforms make it easy to implement advanced compilation techniques that use the ZX calculus graphical language to reduce T-gate counts of Clifford + T circuits, optimize phase polynomials, and reduce the number of gates in non-Clifford circuits. (#8025) (#8029) (#8088) (#7747) (#8201)

    These transforms include:

    • pennylane.transforms.zx.optimize_t_count: reduces the number of T gates in a Clifford + T circuit by applying a sequence of passes that combine ZX-based commutation and cancellation rules and the Third Order Duplicate and Destroy (TODD) algorithm.

    • pennylane.transforms.zx.todd: reduces the number of T gates in a Clifford + T circuit by using the TODD algorithm.

    • pennylane.transforms.zx.reduce_non_clifford: reduces the number of non-Clifford gates in a circuit by applying a combination of phase gadgetization strategies and Clifford gate simplification rules.

    • pennylane.transforms.zx.push_hadamards: reduces the number of large phase-polynomial blocks in a phase-polynomial + Hadamard circuit by pushing Hadamard gates as far as possible to one side.

    As an example, consider the following circuit:

    import pennylane as qml

dev = qml.device("default.qubit")

@qml.qnode(dev)
def circuit():
    qml.T(0)
    qml.CNOT([0, 1])
    qml.S(0)
    qml.T(0)
    qml.T(1)
    qml.CNOT([0, 2])
    qml.T(1)
    return qml.state()
    >>> print(qml.draw(circuit)())
0: ──T─╭●──S──T─╭●────┤  State
1: ────╰X──T────│───T─┤  State
2: ─────────────╰X────┤  State

    We can apply the holistic pennylane.transforms.zx.optimize_t_count compilation pass to reduce the number of T gates. In this case, all T gates can be removed!

    >>> print(qml.draw(qml.transforms.zx.optimize_t_count(circuit))())
0: ──Z─╭●────╭●─┤  State
1: ────╰X──S─│──┤  State
2: ──────────╰X─┤  State

Change operator bases 🍴

  • Users can now benefit from an optimization of the controlled compute-uncompute pattern with the new pennylane.change_op_basis function and pennylane.ops.op_math.ChangeOpBasis class. Operators arranged in a compute-uncompute pattern (U V U†, which is equivalent to changing the basis in which V is expressed) can be efficiently controlled, as the only the central (target) operator V needs to be controlled, and not U or U†. (#8023) (#8070)

    These new features leverage the graph-based decomposition system, enabled with pennylane.decomposition.enable_graph(). To illustrate their use, consider the following example. The compute-uncompute pattern is composed of a QFT, followed by a PhaseAdder, and finally an inverse QFT.

    from functools import partial

qml.decomposition.enable_graph()

dev = qml.device("default.qubit")

@partial(qml.transforms.decompose, max_expansion=1)
@qml.qnode(dev)
def circuit():
    qml.H(0)
    qml.CNOT([1,2])
    qml.ctrl(
        qml.change_op_basis(qml.QFT([1,2]), qml.PhaseAdder(1, x_wires=[1,2])),
        control=0
    )
    return qml.state()

    When this circuit is decomposed, the QFT and Adjoint(QFT) are not controlled, resulting in a much more resource-efficient decomposition:

    >>> print(qml.draw(circuit)())
0: ──H──────╭●────────────────┤  State
1: ─╭●─╭QFT─├PhaseAdder─╭QFT†─┤  State
2: ─╰X─╰QFT─╰PhaseAdder─╰QFT†─┤  State

  • The decompositions for several templates have been updated to use pennylane.ops.op_math.ChangeOpBasis, which makes their decompositions more resource efficient by eliminating unnecessary controlled operations. The templates include pennylane.Adder, pennylane.Multiplier, pennylane.OutAdder, pennylane.OutMultiplier, pennylane.PrepSelPrep. (#8207)

    Here, the optimization is demonstrated when pennylane.Adder is controlled:

    qml.decomposition.enable_graph()

dev = qml.device("default.qubit")

@partial(qml.transforms.decompose, max_expansion=2)
@qml.qnode(dev)
def circuit():
    qml.ctrl(qml.Adder(10, x_wires=[1,2,3,4]), control=0)
    return qml.state()
    >>> print(qml.draw(circuit)())
0: ──────╭●────────────────┤  State
1: ─╭QFT─├PhaseAdder─╭QFT†─┤  State
2: ─├QFT─├PhaseAdder─├QFT†─┤  State
3: ─├QFT─├PhaseAdder─├QFT†─┤  State
4: ─╰QFT─╰PhaseAdder─╰QFT†─┤  State

Quantum optimizers compatible with QJIT 🫖

  • Leveraging pennylane.qjit to optimize hybrid workflows with the momentum quantum natural gradient optimizer is now possible with pennylane.MomentumQNGOptimizerQJIT. This provides better scaling than its non-JIT-compatible counterpart. (#7606)

    The v0.42 release saw the addition of the pennylane.QNGOptimizerQJIT optimizer, which is a qml.qjit-compatible analogue to pennylane.QNGOptimizer. In this release, we've added the pennylane.MomentumQNGOptimizerQJIT optimizer, which is the qml.qjit-compatible analogue to pennylane.MomentumQNGOptimizer. Both optimizers have an Optax-like interface:

    import jax.numpy as jnp

dev = qml.device("lightning.qubit", wires=2)

@qml.qnode(dev)
def circuit(params):
    qml.RX(params[0], wires=0)
    qml.RY(params[1], wires=1)
    return qml.expval(qml.Z(0) + qml.X(1))

opt = qml.MomentumQNGOptimizerQJIT(stepsize=0.1, momentum=0.2)

@qml.qjit
def update_step_qjit(i, args):
    params, state = args
    return opt.step(circuit, params, state)

@qml.qjit
def optimization_qjit(params, iters):
    state = opt.init(params)
    args = (params, state)
    params, state = qml.for_loop(iters)(update_step_qjit)(args)
    return params

    Quantum just-in-time compilation works exceptionally well with repeatedly executing the same function in a for loop. As you can see, $10^5$ iterations takes seconds:

    >>> import time
>>> params = jnp.array([0.1, 0.2])
>>> iters = 100_000
>>> start = time.process_time()
>>> optimization_qjit(params=params, iters=iters)
Array([ 3.14159265, -1.57079633], dtype=float64)
>>> time.process_time() - start
21.319525

Improvements 🛠

Resource-efficient decompositions

  • With pennylane.decomposition.enable_graph(), dynamically allocated wires with pennylane.allocate are now supported in decomposition rules. This provides a smoother overall experience when decomposing operators in a way that requires auxiliary/work wires. (#7861) (#8228)

    Support for pennylane.allocate unlocks the following features:

    • The pennylane.transforms.decompose transform now accepts a max_work_wires argument that allows the user to specify the number of work wires available for dynamic allocation during decomposition. (#7963) (#7980) (#8103) (#8236)

    • Decomposition rules were added for the pennylane.MultiControlledX that dynamically allocate work wires if none were explicitly specified via the work_wires argument. (#8024)

  • Several templates now have decompositions that can be accessed within the graph-based decomposition system (pennylane.decomposition.enable_graph), allowing workflows that include these templates to be decomposed in a resource-efficient and performant manner. (#7779) (#7908) (#7941) (#7943) (#8075) (#8002)

    The included templates are: pennylane.Adder, pennylane.ControlledSequence, pennylane.ModExp, pennylane.MottonenStatePreparation, pennylane.MPSPrep, pennylane.Multiplier, pennylane.OutAdder, pennylane.OutMultiplier, pennylane.OutPoly, pennylane.PrepSelPrep, pennylane.ops.Prod, pennylane.Reflection, pennylane.StatePrep, pennylane.TrotterProduct, pennylane.QROM, pennylane.GroverOperator, pennylane.UCCSD, pennylane.StronglyEntanglingLayers, pennylane.GQSP, pennylane.FermionicSingleExcitation, pennylane.FermionicDoubleExcitation, pennylane.QROM, pennylane.ArbitraryStatePreparation, pennylane.CosineWindow, pennylane.AmplitudeAmplification, pennylane.Permute, pennylane.AQFT, pennylane.FlipSign, pennylane.FABLE, pennylane.Qubitization, and pennylane.Superposition.

  • Two additions were made to pennylane.Select, significantly improving its decomposition:

    • A new keyword argument partial has been added, which allows for simplifications in the decomposition of pennylane.Select under the assumption that the state of the control wires has no overlap with computational basis states that are not used by pennylane.Select.

    • A new decomposition rule has been added to pennylane.Select. It achieves cost reductions by adding one work_wire. This decomposition is useful to perform efficient pennylane.QROM decompositions.

    (#7385) (#7658) (#8011) (#8276)

  • The decomposition of pennylane.BasisRotation has been optimized to skip redundant phase shift gates with angle $\pm \pi$ for real-valued (orthogonal) rotation matrices. This uses the fact that either one or zero pennylane.PhaseShift gates are required in case the matrix has a determinant equal to $\pm 1$. (#7765)

  • The pennylane.transforms.decompose transform is now able to decompose classically controlled operations (i.e., operations nested inside cond). (#8145)

    from functools import partial

dev = qml.device('default.qubit')

@partial(qml.transforms.decompose, gate_set={qml.RY, qml.RZ, qml.measurements.MidMeasureMP})
@qml.qnode(dev)
def circuit():
    m0 = qml.measure(0)
    qml.cond(m0 == 0, qml.Rot)(qml.numpy.pi / 2, qml.numpy.pi / 2, qml.numpy.pi / 2, wires=1)
    return qml.expval(qml.X(0))
    >>> print(qml.draw(circuit, level=0)())
0: ──┤↗├──────────────────────┤  <X>
1: ───║───Rot(1.57,1.57,1.57)─┤     
      ╚═══╝  
>>> print(qml.draw(circuit, level=1)())
0: ──┤↗├───────────────────────────────┤  <X>
1: ───║───RZ(1.57)──RY(1.57)──RZ(1.57)─┤     
      ╚═══╩═════════╩═════════╝

  • Various decompositions of pennylane.MultiControlledX now utilize pennylane.TemporaryAND in place of pennylane.Toffoli gates, leading to more resource-efficient decompositions. (#8172)

  • Controlled(Identity) is now directly decomposed to a single Identity operator instead of going through a numeric decomposition algorithm. (#8388)

  • The internal assignment of basis states in pennylane.Superposition was improved, resulting in its decomposition being more performant and efficient. (#7880)

  • pennylane.decomposition.has_decomp and pennylane.decomposition.list_decomps now take operator instances as arguments instead of types. (#8286)

    >>> qml.decomposition.has_decomp(qml.MultiControlledX)
True
>>> qml.decomposition.list_decomps(qml.Select)
[<pennylane.decomposition.decomposition_rule.DecompositionRule at 0x126f99ed0>,
<pennylane.decomposition.decomposition_rule.DecompositionRule at 0x127002fd0>,
<pennylane.decomposition.decomposition_rule.DecompositionRule at 0x127034bd0>]

  • With the graph-based decomposition system enabled (pennylane.decomposition.enable_graph()), if a decomposition cannot be found for an operator in the circuit in terms of the target gates, it no longer raises an error. Instead, a warning is raised, and op.decomposition() (the current default method for decomposing gates) is used as a fallback, while the rest of the circuit is still decomposed with the new graph-based system. Additionally, a special warning message is raised if the circuit contains a GlobalPhase, reminding the user that GlobalPhase is not assumed to have a decomposition under the new system. (#8156)

  • pennylane.transforms.decompose and pennylane.preprocess.decompose now have a unified internal implementation to promote feature parity in preparation for the graph-based decomposition system to be the default decomposition method in PennyLane. (#8193)

  • A new class called pennylane.decomposition.decomposition_graph.DecompGraphSolution has been added to store the solution of a decomposition graph. An instance of this class is returned from the solve method of the pennylane.decomposition.decomposition_graph.DecompositionGraph class. (#8031)

OpenQASM-PennyLane interoperability

  • The pennylane.from_qasm3 function can now convert OpenQASM 3.0 circuits that contain subroutines, constants, all remaining stdlib gates, qubit registers, and built-in mathematical functions. (#7651) (#7653) (#7676) (#7677) (#7679) (#7690) (#7767)

  • pennylane.to_openqasm now supports mid-circuit measurements and conditionals of unprocessed measurement values. (#8210)

Setting shots

  • The number of shots can now be specified directly in QNodes as a standard keyword argument. (#8073)

    @qml.qnode(qml.device("default.qubit"), shots=1000)
def circuit():
    qml.H(0)
    return qml.expval(qml.Z(0))
    >>> circuit.shots
Shots(total_shots=1000, shot_vector=(ShotCopies(1000 shots x 1),))  
>>> circuit()
np.float64(-0.004)

    Setting the shots value in a QNode is equivalent to decorating with pennylane.workflow.set_shots. However, decorating with pennylane.workflow.set_shots overrides QNode shots:

    >>> new_circ = qml.set_shots(circuit, shots=123)
>>> new_circ.shots
Shots(total_shots=123, shot_vector=(ShotCopies(123 shots x 1),))

  • The pennylanepennylane.set_shots transform can now be directly applied to a QNode without the need for functools.partial, providing a more user-friendly syntax and negating having to import the functools package. (#7876) (#7919)

    @qml.set_shots(shots=1000)  # or @qml.set_shots(1000)
@qml.qnode(dev)
def circuit():
    qml.H(0)
    return qml.expval(qml.Z(0))
    >>> circuit()
0.002

Clifford + T decomposition

  • The pennylane.clifford_t_decomposition transform with method="gridsynth" is now compatible with quantum just-in-time compilation via the pennylane.qjit decorator. (#7711)

    @qml.qjit
@partial(qml.transforms.clifford_t_decomposition, method="gridsynth")
@qml.qnode(qml.device("lightning.qubit", wires=1))
def circuit():
    qml.RX(np.pi/3, wires=0)
    qml.RY(np.pi/4, wires=0)
    return qml.state()
    >>> circuit()
Array([0.80011651+0.19132132j, 0.33140586-0.4619306j ], dtype=complex128)

  • The pennylane.clifford_t_decomposition transform can now decompose circuits with mid-circuit measurements, including Catalyst's measurement operations. It also now handles RZ and PhaseShift operations where angles are odd multiples of $\pm \tfrac{\pi}{4}$ more efficiently when using method="gridsynth". (#7793) (#7942)

  • The pennylane.ops.rs_decomposition method now gives decompositions with exact global phase information. (#7793)

  • Users can now specify a relative threshold value for the permissible operator norm error (epsilon) that triggers rebuilding of the cache in the pennylane.clifford_t_decomposition, via new cache_eps_rtol keyword argument. (#8056)

Transforms

  • New transforms called pennylane.transforms.match_relative_phase_toffoli and pennylane.transforms.match_controlled_iX_gate have been added, which compile certain patterns to efficient Clifford + T equivalents. (#7748)

    @qml.qnode(qml.device("default.qubit", wires=4))
def circuit():
  qml.CCZ(wires=[0, 1, 3])
  qml.ctrl(qml.S(wires=[1]), control=[0])
  qml.ctrl(qml.S(wires=[2]), control=[0, 1])
  qml.MultiControlledX(wires=[0, 1, 2, 3])

  return qml.expval(qml.Z(0))
    >>> new_circuit = qml.transforms.match_relative_phase_toffoli(circuit)
>>> print(qml.draw(new_circuit, level=1)())
0: ─────────────────╭●───────────╭●───────────────────────────┤  <Z>
1: ─────────────────│─────╭●─────│─────╭●─────────────────────┤
2: ───────╭●────────│─────│──────│─────│────────────╭●────────┤
3: ──H──T─╰X──T†──H─╰X──T─╰X──T†─╰X──T─╰X──T†──H──T─╰X──T†──H─┤

  • New intermediate representations (IRs) called pennylanetransforms.parity_matrix and pennylanetransforms.phase_polynomial are available in PennyLane. These IRs are used in compilation passes to optimize CNOT and phase polynomial circuits, respectively. Additionally, the pennylanetransforms.rowcol has been added, which uses the parity matrix as its IR for CNOT routing under constraint connectivity. (#8171) (#8443)

    The example below showcases the use of pennylanetransforms.parity_matrix, which acts on circuits containing only CNOT gates.

    dev = qml.device('default.qubit', wires=1)

@qml.qnode(dev)
def circuit():
    qml.CNOT((3, 2))
    qml.CNOT((0, 2))
    qml.CNOT((2, 1))
    qml.CNOT((3, 2))
    qml.CNOT((3, 0))
    qml.CNOT((0, 2))
    return qml.state()

    Upon transforming the above circuit with pennylanetransforms.parity_matrix, the output is the parity matrix.

    >>> P = qml.transforms.parity_matrix(circuit, wire_order=range(4))()
>>> print(P)
array([[1, 0, 0, 1],
       [1, 1, 1, 1],
       [0, 0, 1, 1],
       [0, 0, 0, 1]])

    The pennylanetransforms.phase_polynomial transform functions similarly, operating on circuits containining only CNOT and RZ gates and returning the parity matrix, the parity table, and corresponding angles for each parity.

    @qml.qnode(dev)
def circuit():
    qml.CNOT((1, 0))
    qml.RZ(1, 0)
    qml.CNOT((2, 0))
    qml.RZ(2, 0)
    qml.CNOT((0, 1))
    qml.CNOT((3, 1))
    qml.RZ(3, 1)
    return qml.state()
    >>> pmat, ptab, angles = qml.transforms.phase_polynomial(circuit, wire_order=range(4))()
>>> pmat
[[1 1 1 0]
 [1 0 1 1]
 [0 0 1 0]
 [0 0 0 1]]
>>> ptab
[[1 1 1]
 [1 1 0]
 [0 1 1]
 [0 0 1]]
>>> angles
[1 2 3]

  • A new transform called pennylane.transforms.rz_phase_gradient has been added, which lets you realize arbitrary angle pennylane.RZ rotations with a phase gradient resource state and semi-in-place addition (pennylane.SemiAdder). This can be a crucial subroutine in FTQC when sufficient auxiliary wires are available, as it saves on T gates compared to other discretization schemes. (#8213)

  • A new keyword argument called shot_dist has been added to the pennylane.transforms.split_non_commuting transform. This allows for more customization and efficiency when calculating expectation values across the non-commuting groups of observables that make up a Hamiltonian/LinearCombination. (#7988)

    Given a QNode that returns a sample-based measurement (e.g., expval) of a Hamiltonian/LinearCombination with finite shots, the current default behaviour of pennylane.transforms.split_non_commuting will perform shots executions for each group of commuting terms. With the shot_dist argument, this behaviour can be changed:

    • "uniform": evenly distributes the number of shots across all groups of commuting terms
    • "weighted": distributes the number of shots according to weights proportional to the L1 norm of the coefficients in each group
    • "weighted_random": same as "weighted", but the numbers of shots are sampled from a multinomial distribution
    • or a user-defined function implementing a custom shot distribution strategy

    To show an example about how this works, let's start by defining a simple Hamiltonian:

    ham = qml.Hamiltonian(
    coeffs=[10, 0.1, 20, 100, 0.2],
    observables=[
        qml.X(0) @ qml.Y(1),
        qml.Z(0) @ qml.Z(2),
        qml.Y(1),
        qml.X(1) @ qml.X(2),
        qml.Z(0) @ qml.Z(1) @ qml.Z(2)
    ]
)

    This Hamiltonian can be split into 3 non-commuting groups of mutually commuting terms. With shot_dist = "weighted", for example, the number of shots will be divided according to the L1 norm of each group's coefficients:

    from functools import partial
from pennylane.transforms import split_non_commuting

dev = qml.device("default.qubit")

@partial(split_non_commuting, shot_dist="weighted")
@qml.qnode(dev, shots=10000)
def circuit():
    return qml.expval(ham)

with qml.Tracker(dev) as tracker:
    circuit()
    >>> print(tracker.history["shots"])
[2303, 23, 7674]

  • The pennylane.noise.fold_global transform has been refactored to collect operators into a list directly rather than relying on queuing. (#8296)

Choi matrix functionality

  • A new function called pennylane.math.choi_matrix is available, which computes the Choi matrix of a quantum channel. This is a useful tool in quantum information science and to check circuit identities involving non-unitary operations. (#7951)

    >>> import numpy as np
>>> Ks = [np.sqrt(0.3) * qml.CNOT((0, 1)), np.sqrt(1-0.3) * qml.X(0)]
>>> Ks = [qml.matrix(op, wire_order=range(2)) for op in Ks]
>>> Lambda = qml.math.choi_matrix(Ks)
>>> np.trace(Lambda), np.trace(Lambda @ Lambda)
(np.float64(1.0), np.float64(0.58))

Other improvements

  • pennylane.snapshots can now be used with mcm_method="one-shot" and mcm_method="tree-traversal". (#8140)

    This improvement is particularly useful for extracting the state in finite-shot workflows:

    @qml.qnode(qml.device("default.qubit"), mcm_method="one-shot", shots=1)
def circuit():
    qml.RY(1.23, 0)

    m0 = qml.measure(0)
    qml.cond(m0 == 0, qml.H)(0)
    qml.Snapshot("state", measurement=qml.state())

    return qml.expval(qml.X(0))
    >>> qml.snapshots(circuit)()
{'state': array([0.+0.j, 1.+0.j]), 'execution_results': np.float64(-1.0)}

    Here, the state is projected onto the corresponding state resulting from the MCM.

  • The printing and drawing of pennylane.TemporaryAND, also known as qml.Elbow, and its adjoint have been improved to be more legible and consistent with how it's depicted in circuits in the literature. (#8017) (#8432)

    import pennylane as qml

@qml.draw
@qml.qnode(qml.device("lightning.qubit", wires=4))
def node():
    qml.TemporaryAND([0, 1, 2], control_values=[1, 0])
    qml.CNOT([2, 3])
    qml.adjoint(qml.TemporaryAND([0, 1, 2], control_values=[1, 0]))
    return qml.expval(qml.Z(3))
    print(node())
0: ─╭●─────●╮─┤
1: ─├○─────○┤─┤
2: ─╰──╭●───╯─┤
3: ────╰X─────┤  <Z>

  • With pennylane.decomposition.enable_graph, the UserWarning that is raised when a decomposition cannot be found for an operator in the circuit is now more generic, not making any assumptions about how the unresolved operations will be applied or used in the decompose transformation. (#8361)

  • The pennylane.sample function can now receive an optional dtype parameter which sets the type and precision of the samples returned by this measurement process. (#8189) (#8271)

  • DefaultQubit will now default to the tree-traversal MCM method when mcm_method="device". (#7885)

  • DefaultQubit now determines the mcm_method in Device.setup_execution_config, making it easier to tell which mcm_method will be used. This also allows defer_measurements and dynamic_one_shot to be applied at different locations in the preprocessing program. (#8184)

  • The default implementation of Device.setup_execution_config now choses "device" as the default mcm_method if it is available, as specified by the device TOML file. (#7968)

  • ExecutionConfig and MCMConfig from pennylane.devices are now frozen dataclasses whose fields should be updated with dataclass.replace. (#7697) (#8046)

  • An error is no longer raised when non-integer wire labels are used in QNodes using mcm_method="deferred". (#7934)

    @qml.qnode(qml.device("default.qubit"), mcm_method="deferred")
def circuit():
    m = qml.measure("a")
    qml.cond(m == 0, qml.X)("aux")
    return qml.expval(qml.Z("a"))
    >>> print(qml.draw(circuit)())
  a: ──┤↗├────┤  <Z>
aux: ───║───X─┤
        ╚═══╝

  • qml.transforms.core.TransformContainer now holds onto a TransformDispatcher, args, and kwargs, instead of the transform's defining function and unpacked properties. It can still be constructed via the old signature, as well. (#8306)

  • The JAX version is now included in pennylane.about. (#8277)

  • A warning is now raised when circuits are executed without Catalyst and with qml.capture.enable() present. (#8291)

  • The QNode primitive in the experimental program capture module now captures the unprocessed ExecutionConfig, instead of one processed by the device. This allows for better integration with Catalyst. (#8258)

  • qml.counts can now be captured with program capture. Circuits returning counts still cannot be interpreted or executed with program capture. (#8229)

  • Templates are now compatible with program capture. (#8211)

  • PennyLane autograph supports standard Python for index assignment (arr[i] = x) and updating array elements (arr[i] += x) instead of jax.numpy form (i.e., arr = arr.at[i].set(x) and arr.at[i].add(x)). (#8027) (#8076)

    import jax.numpy as jnp

qml.capture.enable()

@qml.qnode(qml.device("default.qubit", wires=3))
def circuit(val):
  angles = jnp.zeros(3)
  angles[0:3] += val

  for i, angle in enumerate(angles):
      qml.RX(angle, i)

  return qml.expval(qml.Z(0)), qml.expval(qml.Z(1)), qml.expval(qml.Z(2))
    >>> circuit(jnp.pi)
(Array(-1, dtype=float32),
 Array(-1, dtype=float32),
 Array(-1, dtype=float32))

  • Logical operations (and, or and not) are now supported with PennyLane autograph. (#8006)

    qml.capture.enable()

@qml.qnode(qml.device("default.qubit", wires=1))
def circuit(param):
    if param >= 0 and param <= 1:
        qml.H(0)
    return qml.state()
    >>> circuit(0.5)
Array([0.70710677+0.j, 0.70710677+0.j], dtype=complex64)

  • With program capture, the true_fn can now be a subclass of Operator when no false_fn is provided. For example, qml.cond(condition, qml.X)(0) is now valid code. (#8060) (#8101)

  • With program capture, an error is now raised if the conditional predicate in, say, an if statement is not a scalar. (#8066)

  • Program capture can now handle dynamic shots, shot vectors, and shots set with pennylanepennylane.set_shots. (#7652)

  • The error message raised when using unified-compiler transforms with pennylane.qjit has been updated with suggested fixes. (#7916)

  • Two new draw and generate_mlir_graph functions have been introduced in the qml.compiler.python_compiler.visualization module to visualize circuits with the new unified compiler framework when xDSL and/or Catalyst compilation passes are applied. (#8040) (#8180) (#8091)

  • The catalyst, qec, and ``stablehlo` xDSL dialects have been added to the unified compiler framework, containing data structures that support core compiler functionality and quantum error correction and extending the existing StableHLO dialect with missing upstream operations. (#7901) (#7985) (#8036) (#8084) (#8113)

  • The Quantum xDSL dialect now has more strict constraints for operands and results. (#8083)

  • A callback mechanism has been added to qml.compiler.python_compiler submodule to inspect the intermediate representation of the program between multiple compilation passes. (#7964)

  • A QuantumParser class has been added to the qml.compiler.python_compiler submodule that automatically loads relevant dialects. (#7888)

  • Two new operations have been added to the Quantum dialect of the unified compiler:

    • NumQubitsOp: calculates the number of currently allocated qubits. (#8063)

    • AllocQubitOp and DeallocQubitOp: allocates and deallocates qubits, respectively. (#7915)

  • A compilation pass written called qml.compiler.python_compiler.transforms.MeasurementsFromSamplesPass has been added for integration with the unified compiler framework. This pass replaces all terminal measurements in a program with a single pennylane.sample measurement, and adds postprocessing instructions to recover the original measurement. (#7620)

  • A combine-global-phase pass has been added to the unified compiler framework. Note that the current implementation can only combine all the global phase operations at the last global phase operation in the same region. In other words, global phase operations inside a control flow region can't be combined with those in their parent region. (#7675)

  • The matrix factorization using pennylane.math.decomposition.givens_decomposition has been optimized to factor out the redundant sign in the diagonal phase matrix for the real-valued (orthogonal) rotation matrices. For example, in case the determinant of a matrix is $-1$, only a single element of the phase matrix is required. (#7765)

  • A new device preprocess transform, pennylane.devices.preprocess.no_analytic, is available for hardware devices and hardware-like simulators. It validates that all executions are shot-based. (#8037)

  • PennyLane is now compatible with quimb == 1.11.2 after a bug affecting default.tensor was fixed. (#7931)

  • A new pennylane.transforms.resolve_dynamic_wires transform can allocate concrete wire values for dynamic wire allocation. (#7678) (#8184) (#8406)

Labs: a place for unified and rapid prototyping of research software 🧪

Labs Resource Estimation

  • State-of-the-art resource estimates have been added to existing templates: pennylanepennylane.labs.resource_estimation.ResourceSelectPauliRot, pennylanepennylane.labs.resource_estimation.ResourceQubitUnitary, pennylanepennylane.labs.resource_estimation.ResourceSingleQubitComparator, pennylanepennylane.labs.resource_estimation.ResourceTwoQubitComparator, pennylanepennylane.labs.resource_estimation.ResourceIntegerComparator, pennylanepennylane.labs.resource_estimation.ResourceRegisterComparator, pennylanepennylane.labs.resource_estimation.ResourceUniformStatePrep, pennylanepennylane.labs.resource_estimation.ResourceAliasSampling, pennylanepennylane.labs.resource_estimation.ResourceQFT, pennylanepennylane.labs.resource_estimation.ResourceAQFT, and pennylanepennylane.labs.resource_estimation.ResourceTrotterProduct. (#7786) (#7857) (#7883) (#7920) (#7910)

  • Users can now do resource estimation on QPE and iterative QPE with pennylanepennylane.labs.resource_estimation.ResourceQPE and pennylanepennylane.labs.resource_estimation.ResourceIterativeQPE, respectively. Additionally, a pennylanepennylane.labs.resource_estimation.ResourceControlledSequence template has been added that allows estimating resources on controlled sequences of resource operators. (#8053)

  • estimate_resources has been renamed to estimate to make the function name concise and clearer than labs.resource_estimation.estimate_resources. (#8232)

  • A new ResourceConfig class has been added to help track the configuration for errors, precisions and custom decompositions for the resource estimation pipeline. (#8195)

  • The symbolic ResourceOperators have been updated to use hyperparameters from the config dictionary. (#8181)

  • An internal dequeue() method has been added to the ResourceOperator class to simplify the instantiation of resource operators which require resource operators as input. (#7974)

  • ResourceOperator instances can now be compared with ==. (#8155)

  • A mapper function called pennylanepennylane.labs.resource_estimation.map_to_resource_op has been added to map PennyLane operators to ResourceOperator equivalents. (#8146) (#8162)

  • Several Labs test files have been renamed to prevent conflict with names in mainline PennyLane tests. (#8264)

  • A queueing issue in the ResourceOperator tests has been fixed. (#8204)

Labs Trotter Error Estimation

  • Parallelization support for effective_hamiltonian has been added to improve performance. (#8081) (#8257)

  • New SparseFragment and SparseState classes have been created to allow the use of sparse matrices for Hamiltonian Fragments when estimating Trotter error. (#7971)

  • The pennylanepennylane.labs.trotter_error.perturbation_error function has been updated to sum over expectation values instead of states. (#8226)

  • The docstring in perturbation_error has been updated to use the correct positional argument name. (#8174)

Labs Removals

  • The module qml.labs.zxopt has been removed. Its functionalities are now available in the submodule pennylane.transforms.zx. The same functions are available, but their signature may have changed. - Instead of qml.labs.zxopt.full_optimize, use pennylane.transforms.zx.optimize_t_count - Instead of qml.labs.zxopt.full_reduce, use pennylane.transforms.zx.reduce_non_clifford - Instead of qml.labs.zxopt.todd, use pennylane.transforms.zx.todd - Instead of qml.labs.zxopt.basic_optimization, use pennylane.transforms.zx.push_hadamards (#8177)

Breaking changes 💔

  • autoray has been pinned to v0.8.0 for PennyLane v0.43.0 to prevent potential bugs due to breaking changes in autoray releases. (#8412)

    Previous to this change, the autoray package was upper-bounded in pyproject.toml to unblock CI failures due to breaking changes in v0.8.0. Then, it was unpinned by fixing source code that was broken by the release. (#8110) (#8147) (#8159) (#8160)

  • Using postselect_mode="fill-shots" with mcm_method="one-shot" or "tree-traversal" has been disallowed with default.qubit, as it produces incorrect results where the correlation between measurements is not preserved. (#8411)

  • qml.workflow.construct_batch.expand_fn_transform has been deleted as it was no longer getting used. (#8344)

  • get_canonical_interface_name has been removed in favour of overriding Enum._missing_ in Interface. (#8223)

    If you would like to get the canonical interface you can simply use the Enum:

    >>> from pennylane.math.interface_utils import Interface
>>> Interface("torch")
<Interface.TORCH: 'torch'>  
>>> Interface("jax-jit")
<Interface.JAX_JIT: 'jax-jit'>

  • pennylane.PrepSelPrep has been made more reliable by deriving the attributes coeffs and ` ops from the property ``lcu`` instead of storing them independently. In addition, it is now more consistent with other PennyLane operators, dequeuing its input ``lcu``. (#8169)

  • MidMeasureMP now inherits from Operator instead of MeasurementProcess, which resolves problems caused by it always acting like an operator. (#8166)

  • With the deprecation of the shots kwarg in qml.device, DefaultQubit.eval_jaxpr does not use self.shots from the device anymore; instead, it takes shots as a keyword argument, and the QNode primitive should process the shots and call eval_jaxpr accordingly. (#8161)

  • The methods pauli.PauliWord.operation and pauli.PauliSentence.operation no longer queue any operators. This improves the consistency of the queuing behaviour for the operators. (#8136)

  • qml.sample no longer has singleton dimensions squeezed out for single shots or single wires. This cuts down on the complexity of post-processing due to having to handle single shot and single wire cases separately. The return shape will now always be (shots, num_wires). (#7944) (#8118)

    For a simple qnode:

    @qml.qnode(qml.device('default.qubit'))
def circuit():
    return qml.sample(wires=0)

    Before the change, we had:

    >>> qml.set_shots(circuit, shots=1)()
0

    and now we have:

    >>> qml.set_shots(circuit, shots=1)()
array([[0]])

    Previous behavior can be recovered by squeezing the output:

    >>> qml.math.squeeze(qml.set_shots(circuit, shots=1)())
array(0)

  • Functions involving an execution configuration will now default to None instead of pennylane.devices.DefaultExecutionConfig and have to be handled accordingly. This prevents the potential mutation of a global object. (#7697)

    This means that functions like,

    def some_func(..., execution_config = DefaultExecutionConfig):
    ...

    should be written as follows,

    def some_func(..., execution_config: ExecutionConfig | None = None):
    if execution_config is None:
        execution_config = ExecutionConfig()

  • The pennylane.HilbertSchmidt and pennylane.LocalHilbertSchmidt templates have been updated and their UI has been remarkably simplified. They now accept an operation or a list of operations as unitaries. (#7933)

    In past versions of PennyLane, these templates required providing the U and V unitaries as a qml.tape.QuantumTape and a quantum function, respectively, along with separate parameters and wires.

    With this release, each template has been improved to accept one or more operators as unitaries. The wires and parameters of the approximate unitary V are inferred from the inputs, according to the order provided.

    >>> U = qml.Hadamard(0)
>>> V = qml.RZ(0.1, wires=1)
>>> qml.HilbertSchmidt(V, U)
HilbertSchmidt(0.1, wires=[0, 1])

  • Support for Python 3.10 has been removed and support for Python 3.13 has been added. (#7935)

  • To make the codebase more organized and easier to maintain, custom exceptions were moved into exceptions.py, and a documentation page for them was added in the internals. (#7856)

  • The boolean functions provided in qml.operation have been deprecated. See the deprecations page for equivalent code to use instead. These include not_tape, has_gen, has_grad_method, has_multipar, has_nopar, has_unitary_gen, is_measurement, defines_diagonalizing_gates, and gen_is_multi_term_hamiltonian. (#7924)

  • To prevent code duplication, access to lie_closure, structure_constants and center via qml.pauli has been removed. The functions now live in the liealg module and top level import and usage is advised. (#7928) (#7994)

    import pennylane.liealg
from pennylane.liealg import lie_closure, structure_constants, center

  • qml.operation.Observable and the corresponding Observable.compare have been removed, as PennyLane now depends on the more general Operator interface instead. The Operator.is_hermitian property can instead be used to check whether or not it is highly likely that the operator instance is Hermitian. (#7927)

  • qml.operation.WiresEnum, qml.operation.AllWires, and qml.operation.AnyWires have been removed. To indicate that an operator can act on any number of wires, Operator.num_wires = None should be used instead. This is the default and does not need to be overwritten unless the operator developer wants to validate that the correct number of wires is passed. (#7911)

  • The qml.QNode.get_gradient_fn function has been removed. Instead, use qml.workflow.get_best_diff_method to obtain the differentiation method. (#7907)

  • Top-level access to DeviceError, PennyLaneDeprecationWarning, QuantumFunctionError and ExperimentalWarning has been removed. Please import these objects from the new pennylane.exceptions module. (#7874)

  • To improve code reliability, qml.cut_circuit_mc no longer accepts a shots keyword argument. The shots should instead be set on the tape itself. (#7882)

  • pennylane.tape.tape.expand_tape has been moved to its own file, and made available at qml.tape. (#8296)

Deprecations 👋

  • PennyLane and Lightning will no longer ship wheels for Intel MacOS platforms for v0.44 and newer. Additionally, MacOS ARM wheels will require a minimum OS version of 14.0 for continued use with v0.44 and newer. This change is needed to account for MacOS officially deprecating support for Intel CPUs in the OS (see their blog post for more details).

  • Setting shots on a device through the shots keyword argument (e.g., qml.device("default.qubit", wires=2, shots=1000)) and in QNode calls (e.g., qml.QNode(circuit, dev)(shots=1000)) has been deprecated. Please use the pennylanepennylane.set_shots transform to set the number of shots for a QNode instead. This is done to reduce confusion and code complexity by having a centralized way to set shots. (#7979) (#8161) (#7906)

    dev = qml.device("default.qubit", wires=2)

@qml.set_shots(1000)
@qml.qnode(dev)
def circuit(x):
    qml.RX(x, wires=0)
    return qml.expval(qml.Z(0))

  • Support for using TensorFlow with PennyLane has been deprecated and will be dropped in Pennylane v0.44. Future versions of PennyLane are not guaranteed to work with TensorFlow. Instead, we recommend using the JAX or PyTorch interfaces for machine learning applications to benefit from enhanced support and features. Please consult the following demos for more usage information: Turning quantum nodes into Torch Layers and How to optimize a QML model using JAX and Optax. (#7989) (#8106)

  • pennylane.devices.DefaultExecutionConfig has been deprecated and will be removed in v0.44. Instead, use qml.devices.ExecutionConfig() to create a default execution configuration. This helps prevent unintended changes in a workflow's behaviour that could be caused by using a global, mutable object. (#7987)

  • Specifying the work_wire_type argument in qml.ctrl and other controlled operators as "clean" or "dirty" has been deprecated. Use "zeroed" to indicate that the work wires are initially in the $|0\rangle$ state, and "borrowed" to indicate that the work wires can be in any arbitrary state instead. In both cases, the work wires are restored to their original state upon completing the decomposition. This is done to standardize how work wires are called in PennyLane. (#7993)

  • Providing the Trotter number kwarg num_steps to pennylane.evolve, pennylane.exp, pennylane.ops.Evolution, and pennylane.ops.Exp has been deprecated and will be removed in a future release. Instead, use pennylane.TrotterProduct for approximate methods, providing the n parameter to perform the Suzuki-Trotter product approximation of a Hamiltonian with the specified number of Trotter steps. This change resolves the ambiguity that arises when using num_steps on devices that support analytic evolution (e.g., default.qubit). (#7954) (#7977)

    As a concrete example, consider the following case:

    coeffs = [0.5, -0.6]
ops = [qml.X(0), qml.X(0) @ qml.Y(1)]
H_flat = qml.dot(coeffs, ops)

    Instead of computing the Suzuki-Trotter product approximation as:

    >>> qml.evolve(H_flat, num_steps=2).decomposition()
[RX(0.5, wires=[0]),
PauliRot(-0.6, XY, wires=[0, 1]),
RX(0.5, wires=[0]),
PauliRot(-0.6, XY, wires=[0, 1])]

    The same result can be obtained using pennylane.TrotterProduct as follows:

    >>> decomp_ops = qml.adjoint(qml.TrotterProduct(H_flat, time=1.0, n=2)).decomposition()
>>> [simp_op for op in decomp_ops for simp_op in map(qml.simplify, op.decomposition())]
[RX(0.5, wires=[0]),
PauliRot(-0.6, XY, wires=[0, 1]),
RX(0.5, wires=[0]),
PauliRot(-0.6, XY, wires=[0, 1])]

  • MeasurementProcess.expand has been deprecated. The relevant method can be replaced with qml.tape.QuantumScript(mp.obs.diagonalizing_gates(), [type(mp)(eigvals=mp.obs.eigvals(), wires=mp.obs.wires)]). This improves the code design by removing an unused method with undesired dependencies (i.e. circular dependency). (#7953)

  • QuantumScript.shape and QuantumScript.numeric_type have been deprecated and will be removed in version v0.44. Instead, the corresponding .shape or .numeric_type of the MeasurementProcess class should be used. (#7950)

  • Some unnecessary methods of the qml.CircuitGraph class have been deprecated and will be removed in version v0.44: (#7904)

    • print_contents in favor of print(obj) - observables_in_order in favor of observables - operations_in_order in favor of operations - ancestors_in_order in favor of ancestors(obj, sort=True) - descendants_in_order in favor of descendants(obj, sort=True)

  • The QuantumScript.to_openqasm method has been deprecated and will be removed in version v0.44. Instead, the qml.to_openqasm function should be used. This change makes the code cleaner by separating out methods that interface with external libraries from PennyLane's internal functionality. (#7909)

  • The level=None argument in the pennylane.workflow.get_transform_program, pennylane.workflow.construct_batch, qml.draw, qml.draw_mpl, and qml.specs transforms has been deprecated and will be removed in v0.44. Please use level='device' instead to apply the noise model at the device level. This reduces ambiguity by making it clear that the default is to apply all transforms to the QNode. (#7886) (#8364)

  • qml.qnn.cost.SquaredErrorLoss has been deprecated and will be removed in version v0.44. Instead, this hybrid workflow can be accomplished with a function like loss = lambda *args: (circuit(*args) - target)**2. (#7527)

  • Access to add_noise, insert and noise mitigation transforms from the transforms module has been deprecated. Instead, these functions should be imported from the noise module, which is a more appropriate location for them. (#7854)

  • The qml.QNode.add_transform method has been deprecated and will be removed in v0.44. Instead, please use QNode.transform_program.push_back(transform_container=transform_container). (#7855) (#8266)

Internal changes ⚙️

  • GitHub actions and workflows (interface-unit-tests.yml, tests-labs.yml, unit-test.yml, upload-nightly-release.yml and upload.yml) have been updated to use ubuntu-24.04 runners. (#8371)

  • measurements is now a "core" module in the tach specification. (#7945)

  • Enforce various modules to follow modular architecture via tach. (#7847)

  • CI workflows to test documentation using sybil have been added. (#8324) (#8328) (#8329) (#8330) (#8331) (#8386)

  • The templates/subroutines now has arithmetic, qchem, and time_evolution submodules. (#8333)

  • test_horizontal_cartan_subalgebra.py now uses our fixture seed for reproducibility and CI stability. (#8304)

  • The qml.compiler.python_compiler submodule has been restructured to be more cohesive. (#8273)

  • default.tensor now supports graph decomposition (qml.decomposition.enable_graph()) during preprocessing. (#8253)

  • Legacy interface names from tests have been removed (e.g., interface="jax-python" or interface="pytorch") (#8249)

  • qml.devices.preprocess.decompose now works in graph decomposition mode when a gateset is provided. default.qubit and null.qubit can now use graph decomposition mode. (#8225) (#8265) (#8260)

  • Usage of the pytest.mark.capture marker from tests in the tests/python_compiler directory has been removed. (#8234)

  • pylint has been updated to v3.3.8 in our CI and requirements-dev.txt (#8216)

  • Links in the README.md have been updated. (#8165)

  • The autograph guide to now reflects new capabilities. (#8132)

  • strict=True is now used to zip usage in source code. (#8164) (#8182) (#8183)

  • The autograph keyword argument has been removed from the QNode constructor. To enable autograph conversion, use the qjit decorator together with the qml.capture.disable_autograph context manager.

  • The ability to disable autograph conversion has been added by using the new qml.capture.disable_autograph decorator or context manager. Additionally, the autograph keyword argument has been removed from the QNode constructor. To enable autograph conversion, use the qjit decorator together with the qml.capture.disable_autograph context manager. (#8102) (#8104)

  • Roundtrip testing and module verification to the Python compiler is now done in run_filecheck and run_filecheck_qjit fixtures. (#8049)

  • Various type hints have been improved internally. (#8086) (#8284)

  • The cond primitive with program capture no longer stores missing false branches as None, instead storing them as jaxprs with no output. (#8080)

  • Unnecessary execution tests along with accuracy validation in tests/ops/functions/test_map_wires.py were removed due to stochastic failures. (#8032)

  • A new all-tests-passed gatekeeper job has been added to interface-unit-tests.yml to ensure all test jobs complete successfully before triggering downstream actions. This reduces the need to maintain a long list of required checks in GitHub settings. Also added the previously missing capture-jax-tests job to the list of required test jobs, ensuring this test suite is properly enforced in CI. (#7996)

  • DefaultQubitLegacy (test suite only) has been equipped with seeded sampling. This allows for reproducible sampling results of legacy classical shadow across CI. (#7903)

  • DefaultQubitLegacy (test suite only) no longer provides a customized classical shadow implementation. (#7895)

  • Capture does not block wires=0 anymore. This allows Catalyst to work with zero-wire devices. Note that wires=None is still not allowed. (#7978)

  • The readability of dynamic_one_shot postprocessing has been improved to allow for further modification. (#7962) (#8041)

  • PennyLane's top-level __init__.py file has been updated with imports to improve Python language server support for finding PennyLane submodules. (#7959)

  • Type hints in the measurements module have been improved. (#7938)

  • The codebase has been refactored to adopt modern type hint syntax for Python 3.11+. (#7860) (#7982)

  • Pre-commit hooks have been updated to add gitleaks for security purposes. (#7922)

  • A new fixture called run_filecheck_qjit has been added, which can be used to run FileCheck on integration tests for the qml.compiler.python_compiler submodule. (#7888)

  • The minimum supported pytest version has been updated to 8.4.1. (#7853)

  • The pennylane.io module is now a tertiary module. (#7877)

  • Tests for the split_to_single_terms transformation are now seeded. (#7851)

  • The rc_sync.yml file has been updated to work with the latest pyproject.toml changes. (#7808) (#7818)

  • LinearCombination instances can now be created with _primitive.impl when capture is enabled and tracing is active. (#7893)

  • The TensorLike type is now compatible with static type checkers. (#7905)

  • The supported version of xDSL has been updated to 0.49. (#7923) (#7932) (#8120)

  • The JAX version used in tests to has been updated to 0.6.2 (#7925)

  • An xdsl_extras module has been added to the unified compiler framework to house additional utilities and functionality not available upstream in xDSL. (#8067) (#8120)

  • Two new xDSL passes have been added to the unified compiler framework: decompose-graph-state, which decomposes mbqc.graph_state_prep operations into their corresponding set of quantum operations for execution on state simulators, and null-decompose-graph-state, which replaces mbqc.graph_state_prep operations with single quantum-register allocation operations for execution on null devices. (#8090)

  • The mbqc xDSL dialect has been added to the unified compiler framework, which is used to represent measurement-based quantum-computing instructions in the xDSL framework. (#7815) (#8059)

  • A compilation pass written with xDSL called qml.compiler.python_compiler.transforms.ConvertToMBQCFormalismPass has been added for the experimental unified compiler framework. This pass converts all gates in the MBQC gate set (Hadamard, S, RZ, RotXZX and CNOT) to the textbook MBQC formalism. (#7870) (#8254)

  • A dialects submodule has been added to qml.compiler.python_compiler which now houses all the xDSL dialects we create. Additionally, the MBQCDialect and QuantumDialect dialects have been renamed to MBQC and Quantum. (#7897)

  • The measurement-plane attribute of the unified compiler mbqc dialect now uses the "opaque syntax" format when printing in the generic IR format. This enables usage of this attribute when IR needs to be passed from xDSL to Catalyst. (#7957)

  • A diagonalize_mcms option has been added to the ftqc.decomposition.convert_to_mbqc_formalism tape transform that, when set, maps arbitrary-basis mid-circuit measurements into corresponding diagonalizing gates and Z-basis mid-circuit measurements. (#8105)

  • The mbqc.graph_state_prep operation is now integrated into the convert_to_mbqc_formalism pass. (#8153) (#8301) (#8314) (#8362)

  • A graph_state_utils submodule has been added to python_compiler.transforms.mbqc for common utilities for MBQC workflows. (#8219) (#8273)

  • Support for pubchempy has been updated to 1.0.5 in the unit tests for qml.qchem.mol_data. (#8224)

  • A nightly RC builds script has been added to .github/workflows. (#8148)

  • The test files for pennylane.estimator were renamed to avoid a dual definition error with the pennylane.labs module. (#8261)

Documentation 📝

  • The program capture sharp bits page has been updated to include a warning about the experimental nature of the feature. (#8448)

  • The installation page has been updated to include currently supported Python versions and installation instructions. (#8369)

  • The documentation of qml.probs and qml.Hermitian has been updated with a warning to avoid using them together as the output might be different than expected. Furthermore, a warning is raised if a user attempts to use qml.probs with a Hermitian observable. (#8235)

  • ">>>" and "..." have been removed from ".. code-block::" directives in docstrings to facilitate docstring testing and fit best practices. (#8319)

  • Three more examples of the deprecated usage of qml.device(..., shots=...) have been updated in the documentation. (#8298)

  • The documentation of qml.device has been updated to reflect the usage of pennylanepennylane.set_shots. (#8294)

  • The "Simplifying Operators" section in the Compiling circuits page has been pushed further down the page to show more relevant sections first. (#8233)

  • ancilla has been renamed to auxiliary in internal documentation. (#8005)

  • Small typos in the docstring for qml.noise.partial_wires have been corrected. (#8052)

  • The theoretical background section of pennylane.BasisRotation has been extended to explain the underlying Lie group/algebra homomorphism between the (dense) rotation matrix and the performed operations on the target qubits. (#7765)

  • The code examples in the documentation of pennylane.specs have been updated to replace keyword arguments with gradient_kwargs in the QNode definition. (#8003)

  • The documentation for Operator.pow and Operator.adjoint have been updated to clarify optional developer-facing use cases. (#7999)

  • The docstring of the is_hermitian operator property has been updated to better describe its behaviour. (#7946)

  • The docstrings of all optimizers have been improved for consistency and legibility. (#7891)

  • The code example in the documentation for pennylane.transforms.split_non_commuting has been updated to give the correct output. (#7892)

  • The $\LaTeX$ rendering in the documentation for qml.TrotterProduct and qml.trotterize has been corrected. (#8014)

  • The docstring of ClassicalShadow.entropy has been updated to trim out the outdated part of an explanation about the different choices of the alpha parameter. (#8100)

  • A warning has been added to the interfaces documentation under the Pytorch section to explain that all Pytorch floating-point inputs are promoted to torch.float64. (#8124)

  • The Dynamic Quantum Circuits page has been updated to include the latest device-dependent mid-circuit measurement method defaults. (#8149) (#8444)

  • A syntax rendering issue in the DefaultQubit documentation has been fixed to correctly display the max_workers parameter. (#8289)

Bug fixes 🐛

  • Fixed a bug in default.qubit where the device wasn't properly validating the mcm_method keyword argument. (#8343)

  • An error is now raised if postselection is requested for a zero-probability mid-circuit measurement outcome with finite shots and pennylanepennylane.devices.DefaultQubit when mcm_method="deferred" and postselect_mode="fill-shots", as this previously led to invalid results. (#8389)

  • Applying a transform to a QNode with capture enabled now returns a QNode. This allows autograph to transform the user function when transforms are applied to the QNode. (#8307)

  • qml.compiler.python_compiler.transforms.MergeRotationsPass now takes the adjoint property of merged operations correctly into account. (#8429)

  • Promoting NumPy data to autograd no longer occurs in qml.qchem.molecular_hamiltonian. (#8410)

  • Fixed compatibility with JAX and PyTorch input parameters in pennylane.SpecialUnitary when large numbers of wires are used. (#8209)

  • With qml.capture.enable(), AutoGraph will now be correctly applied to functions containing control flow that are then wrapped in pennylanepennylane.adjoint or pennylanepennylane.ctrl. (#8215)

  • Fixed a bug that was causing parameter broadcasting on default.mixed with diagonal gates in the computational basis to raise an error. (#8251)

  • qml.ctrl(qml.Barrier(), control_wires) now just returns the original Barrier operation, but placed in the circuit where the ctrl happens. (#8238)

  • JIT compilation of pennylanepennylane.MottonenStatePrep can now accept statically defined state-vector arrays. (#8222)

  • Pauli arithmetic operations (e.g., op.simplify()) can now handle abstract/runtime coefficients when participating in a jitted function. (#8190)

  • Operators queued with pennylane.apply no longer get dequeued by subsequent dequeuing operations (e.g., pennylane.adjoint). (#8078)

  • Fixed a bug in the decomposition rules of pennylane.Select with the graph-based decomposition system that broke the decompositions if the target ops of the Select operator were parametrized. This enables the graph-based decomposition system with Select being provided parametrized target ops. (#8186)

  • Exp and Evolution now have improved decompositions, allowing them to handle more situations more robustly. In particular, the generator is simplified prior to decomposition. Now, more time evolution operators can be supported on devices that do not natively support them. (#8133)

  • A scalar product of a norm one scalar and an operator now decomposes into a GlobalPhase and the operator. For example, -1 * qml.X(0) now decomposes into [qml.GlobalPhase(-np.pi), qml.X(0)]. This improves the decomposition of Select when there are complicated target ops. (#8133)

  • Fixed a bug that made the queueing behaviour of qml.PauliWord.operation and qml.PauliSentence.operation depndent on the global state of a program due to a caching issue. (#8135)

  • A more informative error is raised when extremely deep circuits are attempted to be drawn. (#8139)

  • An error is now raised if sequences of classically processed mid-circuit measurements are used as input to pennylane.counts or pennylane.probs (e.g., qml.counts([2*qml.measure(0), qml.measure(1)])) (#8109)

  • Simplifying operators raised to integer powers no longer causes recursion errors. (#8044)

  • Fixed a GPU selection issue in qml.math with PyTorch when multiple GPUs are present. (#8008)

  • The pennylane.for_loop function with capture enabled can now properly handle cases when start == stop. (#8026)

  • Plxpr primitives now only return dynamically shaped arrays if their outputs actually have dynamic shapes. (#8004)

  • Fixed an issue with the tree-traversal MCM method and non-sequential wire orders that produced incorrect results. (#7991)

  • Fixed a bug in pennylane.matrix where an operator's constituent gates in its decomposition were incorrectly queued, causing extraneous gates to appear in the circuit. (#7976)

  • An error is now raised if an end statement is found in a measurement conditioned branch in a QASM string being imported into PennyLane. (#7872)

  • Fixed issue related to pennylane.transforms.to_zx adding the support for Toffoli and CCZ gates conversion into their ZX-graph representation. (#7899)

  • Fixed qml.workflow.get_best_diff_method to correctly align with execute and construct_batch logic in the workflow module for internal consistency. (#7898)

  • Issues were resolved with AutoGraph transforming internal PennyLane library code in addition to user-level code, which was causing downstream errors in Catalyst. (#7889)

  • Fixed a bug that caused calls to QNode.update (e.g., circuit.update(...)(shots=10)) to update the shots value as if set_shots had been applied, causing unnecessary warnings to appear. (#7881)

  • Fixed attributes and types in the quantum dialect in the unified compiler framework that now allows for types to be inferred correctly when parsing. (#7825)

  • Fixed a bug in SemiAdder that was causing failures when inputs were defined with a single wire. (#7940) (#8437)

  • Fixed a bug where qml.prod, qml.matrix, and qml.cond applied on a quantum function was not dequeueing operators passed as arguments to the function. (#8094) (#8119) (#8078)

  • Fixed a bug where a copy of ShadowExpvalMP was incorrect for a multi-term composite observable. (#8078)

  • Fixed a bug where pennylane.transforms.cancel_inverses, pennylane.transforms.merge_rotations, pennylane.transforms.single_qubit_fusion, pennylane.transforms.commute_controlled, and pennylane.transforms.clifford_t_decomposition were giving incorrect results when acting on (#8297) (8297)

  • When using mcm_method="tree-traversal" with qml.samples, the data type of the returned values is now int. This change ensures consistency with the output of other MCM methods. (#8274)

  • The labels for operators that have multiple matrix-valued parameters (e.g. those from pennylane.operation.Operator) can now also be drawn correctly (e.g. with qml.draw). (#8432)

  • Fixed a bug with pennylane.estimator.resource_mapping._map_to_resource_op() where it was incorrectly mapping the pennylane.TrotterProduct template. (#8425)

  • Fixed bugs in the pennylane.estimator module pertaining to tracking resource operator names, as well as the handling of decompositions and measurement operators by the mapper used by the pennylane.estimator.estimate.estimate function. (#8384)

  • Fixed a bug where pennylane.ops.rs_decomposition logic to streamline queuing conditions were applied incorrectly. (#8441)

Contributors ✍️

This release contains contributions from (in alphabetical order):

Runor Agbaire, Guillermo Alonso, Ali Asadi, Utkarsh Azad, Astral Cai, Joey Carter, Yushao Chen, Isaac De Vlugt, Diksha Dhawan, Gabriela Sanchez Diaz, Marcus Edwards, Lillian Frederiksen, Pietropaolo Frisoni, Simone Gasperini, Diego Guala, Sengthai Heng Austin Huang, David Ittah, Soran Jahangiri, Korbinian Kottmann, Elton Law, Mehrdad Malekmohammadi, Pablo Antonio Moreno Casares, Anton Naim Ibrahim, Erick Ochoa, Lee James O'Riordan, Mudit Pandey, Andrija Paurevic, Justin Pickering, Alex Preciado, Shuli Shu, Jay Soni, Paul Haochen Wang, David Wierichs, Jake Zaia.

Assets 2
Loading