Use dynamic quantum architecture (DQA) for circuit decomposition/routing #139

jkotilahti · 2024-10-23T15:49:13Z

IQMDeviceMetadata can now be created from a DQA, and creating it from QuantumArchitectureSpecification (static architecture) is no longer supported.
IQMDevice/IQMDeviceMetadata can still be created independently of the architecture/calset, so the hardcoded Adonis, Apollo etc. classes are not removed. They are currently used mainly for tests, and could be used for simulation.
IQMSampler uses the specified calset (or server default calset if calibration_set_id not specified) to create the IQMDeviceMetadata/IQMDevice.
IQMSampler adds the calset id to the circuits it decomposes/routes. The calset id is used to warn the user of mismatches between the calset used for execution and the decomposed/routed circuits.
Requires Use dynamic quantum architecture for circuit validation and transpilation iqm-client#140 to be merged first, tests will fail until then. I have tested locally that all tests pass when used together with those changes.

src/iqm/cirq_iqm/iqm_sampler.py

smite · 2024-10-25T16:37:19Z

src/iqm/cirq_iqm/devices/iqm_device_metadata.py

+        # connectivity consists of all arity-2 gate loci in the DQA
+        connectivity = tuple(
+            get_qid_locus(locus)
+            for gate_info in architecture.gates.values()
+            for locus in gate_info.loci
+            if len(locus) == 2
        )
        operations: dict[type[cirq.Gate], list[tuple[NamedQid, ...]]] = {
-            cirq_op: [
-                tuple(
-                    (
-                        NamedQid(qb, dimension=2)
-                        if qb.startswith(cls.QUBIT_NAME_PREFIX)
-                        else NamedQid(qb, dimension=cls.RESONATOR_DIMENSION)
-                    )
-                    for qb in args
-                )
-                for args in qubits
-            ]
-            for iqm_op, qubits in architecture.operations.items()
-            for cirq_op in _IQM_CIRQ_OP_MAP[iqm_op]
-        }
-        return cls(qubits, connectivity, operations=operations, resonators=resonators)
-
-    def to_architecture(self) -> QuantumArchitectureSpecification:
-        """Returns the architecture specification object created based on device metadata."""
-        qubits = tuple(qb.name for qb in self._qubit_set)
-        resonators = tuple(qb.name for qb in self.resonator_set)
-        connectivity = tuple(tuple(qb.name for qb in edge) for edge in self.nx_graph.edges())
-        operations: dict[str, list[tuple[str, ...]]] = {
-            iqm_op: [tuple(qb.name for qb in args) for args in qubits]
-            for cirq_op, qubits in self.operations.items()
-            for iqm_op, cirq_ops in _IQM_CIRQ_OP_MAP.items()
-            if cirq_op in cirq_ops
+            cirq_op: [get_qid_locus(locus) for locus in gate_info.loci]
+            for gate_name, gate_info in architecture.gates.items()
+            for cirq_op in _IQM_CIRQ_OP_MAP[gate_name]


operations contains the full locus info for all gates, but it's not used for routing as far as I can tell.
connectivity is used (in the form of IQMDevice._metadata.nx_graph), but it does not represent any single 2q gate type (seems like the router assumes there is just one 2q gate).

Also, connectivity here will contain duplicate loci/edges. When it's converted into a graph the duplicates vanish.

Right. As discussed, this problem already existed before this PR and is due to Cirq limitations. Probably can't solve it here. We can still look into a solution for it, but probably the PR should not be blocked from being merged because of this.

Does this connectivity work for Deneb-style devices? Where the router should not have the MOVE gate edges. If connectivity will turn into a an nx_graph does it makes sense to make labeled edges where the labels correspond to gates? Or is that something the router cannot work with either?

The connectivity (nx_graph) should be the same as before this PR, assuming we compare an "equivalent" DQA and QuantumArchitectureSpecification. So my understanding is that it should work for Deneb same as before this PR.

I think labeled edges for gates wouldn't be used by the router, because the documentation says that the graph is "The connectivity graph of physical qubits.": https://quantumai.google/reference/python/cirq/RouteCQC

So the connectivity is used to build the nx_graph this is used in route_circuit by removing the resonators from it and then using Cirq's routing pass. After that, the MOVE gates are added. So there is no notion of gates in the route_circuit but the case of MOVE gates is handled through the resonator.

Co-authored-by: Ville Bergholm <[email protected]>

CHANGELOG.rst

src/iqm/cirq_iqm/iqm_sampler.py

Aerylia · 2024-10-30T12:14:59Z

src/iqm/cirq_iqm/devices/iqm_device_metadata.py

+        # connectivity consists of all arity-2 gate loci in the DQA
+        connectivity = tuple(
+            get_qid_locus(locus)
+            for gate_info in architecture.gates.values()
+            for locus in gate_info.loci
+            if len(locus) == 2
        )
        operations: dict[type[cirq.Gate], list[tuple[NamedQid, ...]]] = {
-            cirq_op: [
-                tuple(
-                    (
-                        NamedQid(qb, dimension=2)
-                        if qb.startswith(cls.QUBIT_NAME_PREFIX)
-                        else NamedQid(qb, dimension=cls.RESONATOR_DIMENSION)
-                    )
-                    for qb in args
-                )
-                for args in qubits
-            ]
-            for iqm_op, qubits in architecture.operations.items()
-            for cirq_op in _IQM_CIRQ_OP_MAP[iqm_op]
-        }
-        return cls(qubits, connectivity, operations=operations, resonators=resonators)
-
-    def to_architecture(self) -> QuantumArchitectureSpecification:
-        """Returns the architecture specification object created based on device metadata."""
-        qubits = tuple(qb.name for qb in self._qubit_set)
-        resonators = tuple(qb.name for qb in self.resonator_set)
-        connectivity = tuple(tuple(qb.name for qb in edge) for edge in self.nx_graph.edges())
-        operations: dict[str, list[tuple[str, ...]]] = {
-            iqm_op: [tuple(qb.name for qb in args) for args in qubits]
-            for cirq_op, qubits in self.operations.items()
-            for iqm_op, cirq_ops in _IQM_CIRQ_OP_MAP.items()
-            if cirq_op in cirq_ops
+            cirq_op: [get_qid_locus(locus) for locus in gate_info.loci]
+            for gate_name, gate_info in architecture.gates.items()
+            for cirq_op in _IQM_CIRQ_OP_MAP[gate_name]


Does this connectivity work for Deneb-style devices? Where the router should not have the MOVE gate edges. If connectivity will turn into a an nx_graph does it makes sense to make labeled edges where the labels correspond to gates? Or is that something the router cannot work with either?

Aerylia · 2024-10-30T12:20:55Z

src/iqm/cirq_iqm/devices/iqm_device_metadata.py

@@ -35,6 +35,7 @@ class IQMDeviceMetadata(devices.DeviceMetadata):
        operations: Supported quantum operations of the device, mapping op types to their possible loci.
        gateset: Native gateset of the device. If None, a default IQM device gateset will be used.
        resonators: computational resonators of the device
+        architecture: architecture from which values of the other arguments were obtained
    """

    QUBIT_NAME_PREFIX: str = 'QB'


Is this still needed with DQA?

It's still needed because it's used in IQMDevice.get_qubit_index and IQMDeviceMetadata.from_qubit_indices.

get_qubit_index is not used anywhere in code. But could be useful for some purpose, not sure if there was some requirement to have it?

from_qubit_indices makes creating the hardcoded devices like Apollo a bit nicer

Both are probably something we could get rid of if we really want, but I would still keep this and not change these things in this PR.

Use DQA as transpilation target

1b964ef

jkotilahti requested review from kukushechkin, smite, Aerylia, freetonik and rbrazinskas October 23, 2024 15:54

smite added 2 commits October 25, 2024 18:17

Fix tests.

2289536

Linting.

9b588f4

smite reviewed Oct 25, 2024

View reviewed changes

src/iqm/cirq_iqm/iqm_sampler.py Outdated Show resolved Hide resolved

smite reviewed Oct 25, 2024

View reviewed changes

src/iqm/cirq_iqm/iqm_sampler.py Outdated Show resolved Hide resolved

Fixes.

48e8075

smite reviewed Oct 25, 2024

View reviewed changes

jkotilahti and others added 3 commits October 28, 2024 13:21

Update src/iqm/cirq_iqm/iqm_sampler.py

39e9cc7

Co-authored-by: Ville Bergholm <[email protected]>

remove unused imports

944a320

require explicitly specifying device parameter

4620de8

Aerylia requested changes Oct 30, 2024

View reviewed changes

jkotilahti added 4 commits October 30, 2024 21:54

fix pypi publishing

77dc86a

make test_run_does_not_warn compatible with python 3.9-3.10

b7ad5e6

raise error if device not compatible with calset

5ffb135

more changes to IQMSampler initialization

d53c1a4

jkotilahti requested review from Aerylia and smite October 30, 2024 22:47

Aerylia approved these changes Oct 31, 2024

View reviewed changes

jkotilahti merged commit eff5517 into main Nov 1, 2024
13 checks passed

jkotilahti deleted the sw-709-dqa-transpilation-target branch November 1, 2024 14:48

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use dynamic quantum architecture (DQA) for circuit decomposition/routing #139

Use dynamic quantum architecture (DQA) for circuit decomposition/routing #139

jkotilahti commented Oct 23, 2024 •

edited

Loading

smite Oct 25, 2024

jkotilahti Oct 28, 2024

Aerylia Oct 30, 2024

jkotilahti Oct 30, 2024

Aerylia Nov 1, 2024

Aerylia Oct 30, 2024

Aerylia Oct 30, 2024

jkotilahti Oct 30, 2024

Use dynamic quantum architecture (DQA) for circuit decomposition/routing #139

Use dynamic quantum architecture (DQA) for circuit decomposition/routing #139

Conversation

jkotilahti commented Oct 23, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jkotilahti commented Oct 23, 2024 •

edited

Loading