Configuration

Configuration is split into two public dataclasses:

• MultiQPUConfig describes topology, capacity, scheduler resource limits, and Qiskit transpiler knobs.

• LatencyModel describes scalar operation/network costs used by cost and schedule estimators.

Both are intentionally lightweight and serializable so experiments can store config files beside benchmark outputs.

MultiQPUConfig

MultiQPUConfig describes the physical architecture and Qiskit transpiler options.

Field	Default	Meaning	Validation / interaction
n_qpus	10	number of QPUs	architecture builders require positive values
compute_qubits_per_qpu	8	local compute qubits per QPU	non-negative; contributes to capacity and physical qubit count
comm_qubits_per_qpu	1	communication qubits per QPU	non-negative; controls inter-QPU physical links and schedule port limits
intra_topology	"clique"	one of clique, line, ring, grid2d	invalid topology names are rejected by architecture construction
inter_topology	"switch"	one of switch, mesh, ring, degree_d, clos, fat_tree	invalid topology names are rejected by graph/architecture construction
inter_degree	2	degree target for degree_d topology	clamped/validated by QPU graph helpers depending on context
link_capacity	1	max simultaneous remote ops per inter-QPU link per topology-scheduler round	topology scheduler uses this as a per-round resource limit
switch_parallel_links	1_000_000	max distinct QPU pairs per round on a switched fabric	use smaller values to model limited switch fanout; read only for switch, mesh, and clos with at least two comm ports, ignored elsewhere
switch_reconfig_delay	0.0	additional delay per communication round on a switched fabric	added by topology-aware scheduling for switch, mesh, and clos with at least two comm ports; ignored elsewhere
async_classical	True	enables asynchronous classical-latency hiding in topology scheduling	must be boolean
async_overlap	0.5	fraction of classical RTT that can be hidden	finite numeric value in [0, 1]
grid_rows	None	optional rows for grid2d	inferred if omitted; explicit rows/cols must cover local qubits
grid_cols	None	optional columns for grid2d	inferred if omitted; explicit rows/cols must cover local qubits
basis_gates	("rz", "sx", "x", "cx")	Qiskit basis gates used by translation/transpilation	non-empty sequence of non-empty strings
optimization_level	3	Qiskit optimization level	integer 0, 1, 2, or 3
layout_method	"sabre"	Qiskit layout method for global transpilation	non-empty string
routing_method	"sabre"	Qiskit routing method	non-empty string

Loading a config file

load_config reads JSON or YAML, picking the parser from the suffix without regard to case, and drops a leading byte-order mark so a file saved by an editor that writes one still loads. dump_config creates the directory its path names.

The CLI checks a --config file as it loads it: a path it cannot read, text neither parser accepts, an unknown field, and a document that parses but cannot describe an architecture are all reported as errors naming the file, rather than surfacing later as a traceback from wherever the value was first used.

Derived quantities

• total_physical_qubits() returns n_qpus * (compute_qubits_per_qpu + comm_qubits_per_qpu).

• capacity_per_qpu() returns compute_qubits_per_qpu + comm_qubits_per_qpu.

• The physical block size for one QPU is capacity_per_qpu().

Topology-specific notes

• grid2d uses row-major local-qubit placement. If both dimensions are omitted, dimensions are inferred. If one dimension is provided, the other is inferred.

• degree_d uses inter_degree to create a bounded-degree QPU graph.

• clos and fat_tree are abstractions, not detailed switch-level hardware descriptions.

• switch_parallel_links, link_capacity, async_classical, and async_overlap matter most in topology-aware scheduling.

LatencyModel

LatencyModel stores cost coefficients for metrics and schedule estimates.

Field	Default	Meaning	Used by
oneq	1.0	one-qubit operation cost	cost and schedule estimators
twoq	10.0	local two-qubit operation cost	cost and schedule estimators
swap	30.0	SWAP cost	cost and coarse schedule estimators
epr_gen	200.0	entanglement generation/network setup cost	remote operation cost
classical_rtt	20.0	classical round-trip component	remote operation cost and async overlap model
remote_gate_overhead	50.0	remote gate protocol overhead	remote operation cost
epr_success_prob	1.0	heralded entanglement success probability per attempt, in (0, 1]	entanglement-aware schedule only

estimate_latency(n_1q, n_2q, swaps, remote_2q, depth=None) computes a scalar latency proxy and validates all counts and coefficients before computing. The optional depth argument adds a soft depth penalty and is not a replacement for a schedule estimator.

expected_epr_time(hops) returns hops * epr_gen / epr_success_prob, the expected time to distribute one EPR pair across hops links: entanglement generation is heralded, so one usable pair needs 1 / epr_success_prob attempts on average. Only estimate_entanglement_schedule reads epr_success_prob, and its default of 1.0 models a deterministic link, so every other estimator is unchanged.

Config files

Use the CLI to create a config file:

quport gen-config

JSON and YAML are supported by load_config/dump_config, selected by the file extension; the default output is JSON because YAML requires the yaml extra. Config files must contain a mapping/object and unknown fields are rejected so misspelled settings do not silently take effect.

Example JSON:

{
  "n_qpus": 4,
  "compute_qubits_per_qpu": 8,
  "comm_qubits_per_qpu": 2,
  "intra_topology": "ring",
  "inter_topology": "degree_d",
  "inter_degree": 2,
  "basis_gates": ["rz", "sx", "x", "cx"],
  "optimization_level": 3,
  "layout_method": "sabre",
  "routing_method": "sabre"
}

Experiment design guidance

• Change one family of fields at a time when running sweeps.

• Keep basis_gates, layout_method, routing_method, and optimization_level stable when comparing partitioning strategies.

• Use more than one random trial for benchmark claims; a single random circuit can overstate or understate topology effects.

• Store the config file, seed range, strategies, and QuPort version/commit beside benchmark outputs for reproducibility.