Providers Interface

qiskit.providers

This module contains the classes used to build external providers for Terra. A provider is anything that provides an external service to Terra. The typical example of this is a Backend provider which provides Backend objects which can be used for executing QuantumCircuit and/or Schedule objects. This module contains the abstract classes which are used to define the interface between a provider and terra.

Version Support

Each providers interface abstract class is individually versioned. When we need to make a change to an interface a new abstract class will be created to define the new interface. These interface changes are not guaranteed to be backwards compatible between versions.

Version Changes

Each minor version release of qiskit-terra may increment the version of any providers interface a single version number. It will be an aggregate of all the interface changes for that release on that interface.

Version Support Policy

To enable providers to have time to adjust to changes in this interface Terra will support multiple versions of each class at once. Given the nature of one version per release the version deprecation policy is a bit more conservative than the standard deprecation policy. Terra will support a provider interface version for a minimum of 3 minor releases or the first release after 6 months from the release that introduced a version, whichever is longer, prior to a potential deprecation. After that the standard deprecation policy will apply to that interface version. This will give providers and users sufficient time to adapt to potential breaking changes in the interface. So for example lets say in 0.19.0 BackendV2 is introduced and in the 3 months after the release of 0.19.0 we release 0.20.0, 0.21.0, and 0.22.0, then 7 months after 0.19.0 we release 0.23.0. In 0.23.0 we can deprecate BackendV2, and it needs to still be supported and can’t be removed until the deprecation policy completes.

It’s worth pointing out that Terra’s version support policy doesn’t mean providers themselves will have the same support story, they can (and arguably should) update to newer versions as soon as they can, the support window is just for Terra’s supported versions. Part of this lengthy window prior to deprecation is to give providers enough time to do their own deprecation of a potential end user impacting change in a user facing part of the interface prior to bumping their version. For example, let’s say we changed the signature to Backend.run() in BackendV34 in a backwards incompatible way, before Aer could update its AerBackend class to use version 34 they’d need to deprecate the old signature prior to switching over. The changeover for Aer is not guaranteed to be lockstep with Terra so we need to ensure there is a sufficient amount of time for Aer to complete its deprecation cycle prior to removing version 33 (ie making version 34 mandatory/the minimum version).

Abstract Classes

Provider

Backend

Options

Job

Job Status

Exceptions

QiskitBackendNotFoundError

BackendPropertyError

JobError

JobTimeoutError

BackendConfigurationError

Writing a New Provider

If you have a quantum device or simulator that you would like to integrate with Qiskit you will need to write a provider. A provider will provide Terra with a method to get available BackendV2 objects. The BackendV2 object provides both information describing a backend and its operation for the transpiler so that circuits can be compiled to something that is optimized and can execute on the backend. It also provides the run() method which can run the QuantumCircuit objects and/or Schedule objects. This enables users and other Qiskit APIs, such as execute() and higher level algorithms in qiskit.algorithms, to get results from executing circuits on devices in a standard fashion regardless of how the backend is implemented. At a high level the basic steps for writing a provider are:

• Implement a ProviderV1 subclass that handles access to the backend(s).

• Implement a BackendV2 subclass and its run() method.

  • Add any custom gates for the backend’s basis to the session EquivalenceLibrary instance.

• Implement a JobV1 subclass that handles interacting with a running job.

For a simple example of a provider, see the qiskit-aqt-provider

Provider

A provider class serves a single purpose: to get backend objects that enable executing circuits on a device or simulator. The expectation is that any required credentials and/or authentication will be handled in the initialization of a provider object. The provider object will then provide a list of backends, and methods to filter and acquire backends (using the provided credentials if required). An example provider class looks like:

from qiskit.providers import ProviderV1 as Provider
from qiskit.providers.providerutils import filter_backends
 
from .backend import MyBackend
 
class MyProvider(Provider):
 
    def __init__(self, token=None):
        super().__init__()
        self.token = token
        self.backends = [MyBackend(provider=self)]
 
    def backends(self, name=None, **kwargs):
        if name:
            backends = [
                backend for backend in backends if backend.name() == name]
        return filter_backends(backends, filters=filters, **kwargs)

Ensure that any necessary information for authentication (if required) are present in the class and that the backends method matches the required interface. The rest is up to the specific provider on how to implement.

Backend

The backend classes are the core to the provider. These classes are what provide the interface between Qiskit and the hardware or simulator that will execute circuits. This includes providing the necessary information to describe a backend to the compiler so that it can embed and optimize any circuit for the backend. There are 4 required things in every backend object: a target property to define the model of the backend for the compiler, a max_circuits property to define a limit on the number of circuits the backend can execute in a single batch job (if there is no limit None can be used), a run() method to accept job submissions, and a _default_options method to define the user configurable options and their default values. For example, a minimum working example would be something like:

from qiskit.providers import BackendV2 as Backend
from qiskit.transpiler import Target
from qiskit.providers import Options
from qiskit.circuit import Parameter, Measure
from qiskit.circuit.library import PhaseGate, SXGate, UGate, CXGate, IGate
 
 
class Mybackend(Backend):
 
    def __init__(self):
        super().__init__()
 
        # Create Target
        self._target = Target("Target for My Backend")
        # Instead of None for this and below instructions you can define
        # a qiskit.transpiler.InstructionProperties object to define properties
        # for an instruction.
        lam = Parameter("λ")
        p_props = {(qubit,): None for qubit in range(5)}
        self._target.add_instruction(PhaseGate(lam), p_props)
        sx_props = {(qubit,): None for qubit in range(5)}
        self._target.add_instruction(SXGate(), sx_props)
        phi = Parameter("φ")
        theta = Parameter("ϴ")
        u_props = {(qubit,): None for qubit in range(5)}
        self._target.add_instruction(UGate(theta, phi, lam), u_props)
        cx_props = {edge: None for edge in [(0, 1), (1, 2), (2, 3), (3, 4)]}
        self._target.add_instruction(CXGate(), cx_props)
        meas_props = {(qubit,): None for qubit in range(5)}
        self._target.add_instruction(Measure(), meas_props)
        id_props = {(qubit,): None for qubit in range(5)}
        self._target.add_instruction(IGate(), id_props)
 
        # Set option validators
        self.options.set_validator("shots", (1, 4096))
        self.options.set_validator("memory", bool)
 
    @property
    def target(self):
        return self._target
 
    @property
    def max_circuits(self):
        return 1024
 
    @classmethod
    def _default_options(cls):
        return Options(shots=1024, memory=False)
 
    def run(circuits, **kwargs):
        # serialize circuits submit to backend and create a job
        for kwarg in kwargs:
            if not hasattr(kwarg, self.options):
                warnings.warn(
                    "Option %s is not used by this backend" % kwarg,
                    UserWarning, stacklevel=2)
        options = {
            'shots': kwargs.get('shots', self.options.shots),
            'memory': kwargs.get('memory', self.options.shots),
        }
        job_json = convert_to_wire_format(circuit, options)
        job_handle = submit_to_backend(job_jsonb)
        return MyJob(self. job_handle, job_json, circuit)

Transpiler Interface

The key piece of the Backend object is how it describes itself to the compiler. This is handled with the Target class which defines a model of a backend for the transpiler. A backend object will need to return a Target object from the target attribute which the transpile() function will use as its model of a backend target for compilation.

Custom Basis Gates

1. If your backend doesn’t use gates in the Qiskit circuit library (qiskit.circuit.library) you can integrate support for this into your provider. The basic method for doing this is first to define a Gate subclass for each custom gate in the basis set. For example:

   import numpy as np
    
   from qiskit.circuit import Gate
   from qiskit.circuit import QuantumCircuit
    
   class SYGate(Gate):
       def __init__(self, label=None):
           super().__init__("sy", 1, [], label=label)
    
       def _define(self):
           qc = QuantumCircuit(1)
           q.ry(np.pi / 2, 0)
           self.definition = qc

   The key thing to ensure is that for any custom gates in your Backend’s basis set your custom gate’s name attribute (the first param on super().__init__() in the __init__ definition above) does not conflict with the name of any other gates. The name attribute is what is used to identify the gate in the basis set for the transpiler. If there is a conflict the transpiler will not know which gate to use.

2. Add the custom gate to the target for your backend. This can be done with the Target.add_instruction() method. You’ll need to add an instance of SYGate and its parameters to the target so the transpiler knows it exists. For example, assuming this is part of your BackendV2 implementation for your backend:

   from qiskit.transpiler import InstructionProperties
    
   sy_props = {
       (0,): InstructionProperties(duration=2.3e-6, error=0.0002)
       (1,): InstructionProperties(duration=2.1e-6, error=0.0001)
       (2,): InstructionProperties(duration=2.5e-6, error=0.0003)
       (3,): InstructionProperties(duration=2.2e-6, error=0.0004)
   }
   self.target.add_instruction(SYGate(), sy_props)

   The keys in sy_props define the qubits where the backend SYGate can be used on, and the values define the properties of SYGate on that qubit. For multiqubit gates the tuple keys contain all qubit combinations the gate works on (order is significant, i.e. (0, 1) is different from (1, 0)).

3. After you’ve defined the custom gates to use for the backend’s basis set then you need to add equivalence rules to the standard equivalence library so that the transpile() function and transpiler module can convert an arbitrary circuit using the custom basis set. This can be done by defining equivalent circuits, in terms of the custom gate, for standard gates. Typically if you can convert from a CXGate (if your basis doesn’t include a standard 2 qubit gate) and some commonly used single qubit rotation gates like the HGate and UGate that should be sufficient for the transpiler to translate any circuit into the custom basis gates. But, the more equivalence rules that are defined from standard gates to your basis the more efficient translation from an arbitrary circuit to the target basis will be (although not always, and there is a diminishing margin of return).

   For example, if you were to add some rules for the above custom SYGate we could define the U2Gate and HGate:

   from qiskit.circuit.equivalence_library import SessionEquivalenceLibrary
   from qiskit.circuit.library import HGate
   from qiskit.circuit.library import ZGate
   from qiskit.circuit.library import RZGate
   from qiskit.circuit.library import U2Gate
    
    
   # H => Z SY
   q = qiskit.QuantumRegister(1, "q")
   def_sy_h = qiskit.QuantumCircuit(q)
   def_sy_h.append(ZGate(), [q[0]], [])
   def_sy_h.append(SYGate(), [q[0]], [])
   SessionEquivalenceLibrary.add_equivalence(
       HGate(), def_sy_h)
    
   # u2 => Z SY Z
   phi = qiskit.circuit.Parameter('phi')
   lam = qiskit.circuit.Parameter('lambda')
   q = qiskit.QuantumRegister(1, "q")
   def_sy_u2 = qiskit.QuantumCircuit(q)
   def_sy_u2.append(RZGate(lam), [q[0]], [])
   def_sy_u2.append(SYGate(), [q[0]], [])
   def_sy_u2.append(RZGate(phi), [q[0]], [])
   SessionEquivalenceLibrary.add_equivalence(
       U2Gate(phi, lam), def_sy_u2)

   You will want this to be run on import so that as soon as the provider’s package is imported it will be run. This will ensure that any time the BasisTranslator pass is run with the custom gates the equivalence rules are defined.

   It’s also worth noting that depending on the basis you’re using, some optimization passes in the transpiler, such as Optimize1qGatesDecomposition, may not be able to operate with your custom basis. For our SYGate example, the Optimize1qGatesDecomposition will not be able to simplify runs of single qubit gates into the SY basis. This is because the OneQubitEulerDecomposer class does not know how to work in the SY basis. To solve this the SYGate class would need to be added to Qiskit and OneQubitEulerDecomposer updated to support decomposing to the SYGate. Longer term that is likely a better direction for custom basis gates and contributing the definitions and support in the transpiler will ensure that it continues to be well supported by Qiskit moving forward.

Custom Transpiler Passes

The transpiler supports the ability for backends to provide custom transpiler stage implementations to facilitate hardware specific optimizations and circuit transformations. Currently there are two stages supported, get_translation_stage_plugin() and get_scheduling_stage_plugin() which allow a backend to specify string plugin names to be used as the default translation and scheduling stages, respectively. These hook points in a BackendV2 class can be used if your backend has requirements for compilation that are not met by the current backend/Target interface. Please also consider submitting a Github issue describing your use case as there is interest in improving these interfaces to be able to describe more hardware architectures in greater depth.

To leverage these hook points you just need to add the methods to your BackendV2 implementation and have them return a string plugin name. For example:

class Mybackend(BackendV2):
 
    def get_scheduling_stage_plugin(self):
        return "SpecialDD"
 
    def get_translation_stage_plugin(self):
        return "BasisTranslatorWithCustom1qOptimization"

This snippet of a backend implementation will now have the transpile() function use the SpecialDD plugin for the scheduling stage and the BasisTranslatorWithCustom1qOptimization plugin for the translation stage by default when the target is set to Mybackend. Note that users may override these choices by explicitly selecting a different plugin name. For this interface to work though transpiler stage plugins must be implemented for the returned plugin name. You can refer to qiskit.transpiler.preset_passmanagers.plugin module documentation for details on how to implement plugins. The typical expectation is that if your backend requires custom passes as part of a compilation stage the provider package will include the transpiler stage plugins that use those passes. However, this is not required and any valid method (from a built-in method or external plugin) can be used.

This way if these two compilation steps are required for running or providing efficient output on Mybackend the transpiler will be able to perform these custom steps without any manual user input.

Run Method

Of key importance is the run() method, which is used to actually submit circuits to a device or simulator. The run method handles submitting the circuits to the backend to be executed and returning a Job object. Depending on the type of backend this typically involves serializing the circuit object into the API format used by a backend. For example, on IBMQ backends from the qiskit-ibmq-provider package this involves converting from a quantum circuit and options into a qobj JSON payload and submitting that to the IBM Quantum API. Since every backend interface is different (and in the case of the local simulators serialization may not be needed) it is expected that the backend’s run method will handle this conversion.

An example run method would be something like:

def run(self, circuits. **kwargs):
    for kwarg in kwargs:
        if not hasattr(kwarg, self.options):
            warnings.warn(
                "Option %s is not used by this backend" % kwarg,
                UserWarning, stacklevel=2)
    options = {
        'shots': kwargs.get('shots', self.options.shots)
        'memory': kwargs.get('memory', self.options.shots),
    }
    job_json = convert_to_wire_format(circuit, options)
    job_handle = submit_to_backend(job_jsonb)
    return MyJob(self. job_handle, job_json, circuit)

Options

There are often several options for a backend that control how a circuit is run. The typical example of this is something like the number of shots which is how many times the circuit is to be executed. The options available for a backend are defined using an Options object. This object is initially created by the _default_options method of a Backend class. The default options returns an initialized Options object with all the default values for all the options a backend supports. For example, if the backend supports only supports shots the _default_options method would look like:

@classmethod
def _default_options(cls):
    return Options(shots=1024)

You can also set validators on an Options object to provide limits and validation on user provided values based on what’s acceptable for your backend. For example, if the "shots" option defined above can be set to any value between 1 and 4096 you can set the validator on the options object for you backend with:

self.options.set_validator("shots", (1, 4096))

you can refer to the set_validator() documentation for a full list of validation options.

Job

The output from the run method is a JobV1 object. Each provider is expected to implement a custom job subclass that defines the behavior for the provider. There are 2 types of jobs depending on the backend’s execution method, either a sync or async. By default jobs are considered async and the expectation is that it represents a handle to the async execution of the circuits submitted with Backend.run(). An async job object provides users the ability to query the status of the execution, cancel a running job, and block until the execution is finished. The result is the primary user facing method which will block until the execution is complete and then will return a Result object with results of the job.

For some backends (mainly local simulators) the execution of circuits is a synchronous operation and there is no need to return a handle to a running job elsewhere. For sync jobs its expected that the run method on the backend will block until a Result object is generated and the sync job will return with that inner Result object.

An example job class for an async API based backend would look something like:

from qiskit.providers import JobV1 as Job
from qiskit.providers import JobError
from qiskit.providers import JobTimeoutError
from qiskit.providers.jobstatus import JobStatus
from qiskit.result import Result
 
 
class MyJob(Job):
    def __init__(self, backend, job_id, job_json, circuits):
        super().__init__(backend, job_id)
        self._backend = backend
        self.job_json = job_json
        self.circuits = circuits
 
    def _wait_for_result(self, timeout=None, wait=5):
        start_time = time.time()
        result = None
        while True:
            elapsed = time.time() - start_time
            if timeout and elapsed >= timeout:
                raise JobTimeoutError('Timed out waiting for result')
            result = get_job_status(self._job_id)
            if result['status'] == 'complete':
                break
            if result['status'] == 'error':
                raise JobError('Job error')
            time.sleep(wait)
        return result
 
    def result(self, timeout=None, wait=5):
        result = self._wait_for_result(timeout, wait)
        results = [{'success': True, 'shots': len(result['counts']),
                    'data': result['counts']}]
        return Result.from_dict({
            'results': results,
            'backend_name': self._backend.configuration().backend_name,
            'backend_version': self._backend.configuration().backend_version,
            'job_id': self._job_id,
            'qobj_id': ', '.join(x.name for x in self.circuits),
            'success': True,
        })
 
    def status(self):
        result = get_job_status(self._job_id)
        if result['status'] == 'running':
            status = JobStatus.RUNNING
        elif result['status'] == 'complete':
            status = JobStatus.DONE
        else:
            status = JobStatus.ERROR
        return status
 
def submit(self):
    raise NotImplementedError

and for a sync job:

class MySyncJob(Job):
    _async = False
 
    def __init__(self, backend, job_id, result):
        super().__init__(backend, job_id)
        self._result = result
 
    def submit(self):
        return
 
    def result(self):
        return self._result
 
    def status(self):
        return JobStatus.DONE

Primitives

While not directly part of the provider interface, the qiskit.primitives module is tightly coupled with providers. Specifically the primitive interfaces, such as BaseSampler and BaseEstimator, are designed to enable provider implementations to provide custom implementations which are optimized for the provider’s backends. This can include customizations like circuit transformations, additional pre- and post-processing, batching, caching, error mitigation, etc. The concept of the qiskit.primitives module is to explicitly enable this as the primitive objects are higher level abstractions to produce processed higher level outputs (such as probability distributions and expectation values) that abstract away the mechanics of getting the best result efficiently, to concentrate on higher level applications using these outputs.

For example, if your backends were well suited to leverage mthree measurement mitigation to improve the quality of the results, you could implement a provider-specific Sampler implementation that leverages the M3Mitigation class internally to run the circuits and return quasi-probabilities directly from mthree in the result. Doing this would enable algorithms from qiskit.algorithms to get the best results with mitigation applied directly from your backends. You can refer to the documentation in qiskit.primitives on how to write custom implementations. Also the built-in implementations: Sampler, Estimator, BackendSampler, and BackendEstimator can serve as references/models on how to implement these as well.

Migrating between Backend API Versions

BackendV1 -> BackendV2

The BackendV2 class re-defined user access for most properties of a backend to make them work with native Qiskit data structures and have flatter access patterns. However this means when using a provider that upgrades from BackendV1 to BackendV2 existing access patterns will need to be adjusted. It is expected for existing providers to deprecate the old access where possible to provide a graceful migration, but eventually users will need to adjust code. The biggest change to adapt to in BackendV2 is that most of the information accessible about a backend is contained in its Target object and the backend’s attributes often query its target attribute to return information, however in many cases the attributes only provide a subset of information the target can contain. For example, backend.coupling_map returns a CouplingMap constructed from the Target accessible in the target attribute, however the target may contain instructions that operate on more than two qubits (which can’t be represented in a CouplingMap) or has instructions that only operate on a subset of qubits (or two qubit links for a two qubit instruction) which won’t be detailed in the full coupling map returned by coupling_map. So depending on your use case it might be necessary to look deeper than just the equivalent access with BackendV2.

Below is a table of example access patterns in BackendV1 and the new form with BackendV2:

There is also a BackendV2Converter class available that enables you to wrap a BackendV1 object with a BackendV2 interface.