vqpu-sdk 0.4.0

Universal quantum backend abstraction with circuit knitting and Cryo-Canonical Basin Weaving optimization.

vQPU — Universal Virtual Quantum Processing Unit

Author: Bernard Essuman
Version: 0.4.0
License: MIT
Contact: [e-mail]

---

vQPU is a quantum computing SDK that provides a universal abstraction layer across CPUs, GPUs (NVIDIA, AMD, Intel, Apple Silicon), TPUs, and cloud quantum processors (IonQ, IBM, Google, Rigetti, AWS Braket, Azure Quantum). Write a quantum circuit once, execute it on any available hardware, and let the framework handle backend discovery, routing, and optimization.

The package introduces two original contributions to the field:

• Circuit Knitting with exact zero-overhead reconstruction for controlled-gate cuts, enabling circuits larger than any single backend to run across heterogeneous devices.
• Cryo-Canonical Basin Weaving (CCBW), a novel variational optimizer based on original research by Bernard Essuman, which uses structured 3-3+1 motif probing, mirror-balance symmetry certification, and cold-seeking spring-network optimization to navigate quantum parameter landscapes.

vQPU has been validated with 33/33 tests passing across all modules and live-tested on IonQ's quantum cloud simulator with 320 circuit evaluations.

---

Table of Contents

• Installation
• Quick Start
• Core Capabilities
• Universal Backend Discovery
• Circuit Knitting
• Cryo-Canonical Basin Weaving Optimizer
• Phantom Adaptive Simulator
• Persistent Link Layer
• CHESSO Compiler
• Architecture
• Original Contributions
• Validation Results
• What's New in 0.4.0
• API Reference
• Requirements
• License

---

Installation

pip install vqpu-sdk

Optional backends — install only what your hardware supports:

pip install vqpu-sdk[apple]          # Apple Silicon GPU via Metal/MLX
pip install vqpu-sdk[cuda]           # NVIDIA GPU via CUDA/cupy
pip install vqpu-sdk[ionq]           # IonQ cloud QPU
pip install vqpu-sdk[ibm]            # IBM Quantum
pip install vqpu-sdk[all-qpu]        # All cloud QPU providers
pip install vqpu-sdk[all-classical]  # All local GPU/TPU accelerators

---

Quick Start

from vqpu import UniversalvQPU

# Discovers all available backends automatically
qpu = UniversalvQPU()

# Build a 4-qubit GHZ state
circuit = qpu.circuit(4, "ghz")
circuit.h(0).cnot(0, 1).cnot(1, 2).cnot(2, 3)

# Execute on the best available backend
result = qpu.run(circuit, shots=1024)
print(result.counts)
# {'0000': 512, '1111': 512}

---

Core Capabilities

The core module provides quantum primitives built from first principles: amplitude vectors, unitary gate matrices, Born-rule measurement, and a chainable circuit API.

from vqpu import vQPU, QuantumAlgorithms

qpu = vQPU()

# Bell state
bell = QuantumAlgorithms.bell_pair(qpu)
result = qpu.run(bell, shots=1000)
print(result.counts)       # {'00': ~500, '11': ~500}
print(result.most_probable())  # '00' or '11'

# Variational ansatz for VQE/QAOA
params = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6]
ansatz = QuantumAlgorithms.variational_ansatz(qpu, n=3, params=params, layers=2)
result = qpu.run(ansatz, shots=512)

Key components:

Class	Purpose
QuantumRegister	N-qubit state as 2^N complex amplitude vector
GateLibrary	All standard gates (H, X, Y, Z, S, T, Rx, Ry, Rz, CNOT, CZ, SWAP)
GateEngine	Amplitude-by-amplitude gate application without Kronecker products
MeasurementTap	Born-rule sampling, partial measurement, expectation values
QuantumCircuit	Chainable circuit builder with gate counting and depth analysis
SymmetryFilter	Post-selection filtering (Hamming weight, parity, custom constraints)
ExecutionResult	Measurement counts, probabilities, statevector, metadata

---

Universal Backend Discovery

The universal module probes every backend at startup, fingerprints their capabilities, and routes circuits to the most suitable device.

from vqpu import UniversalvQPU

qpu = UniversalvQPU()
# Output:
#   [cpu] CPU(8cores)          ONLINE  27qb
#   [gpu] GPU::AppleSilicon    ONLINE  27qb

# Build a circuit — it runs on the best backend automatically
circuit = qpu.circuit(5, "test")
circuit.h(0).cnot(0, 1).cnot(1, 2).cnot(2, 3).cnot(3, 4)
result = qpu.run(circuit, shots=2048)

# Inspect the execution plan
plan = qpu.plan(circuit)

Supported backends:

Backend	Plugin	Detection
CPU	CPUPlugin	Always available (NumPy)
NVIDIA GPU	NvidiaGPUPlugin	CUDA + cupy
AMD GPU	AMDGPUPlugin	ROCm + cupy/torch
Intel GPU	IntelGPUPlugin	XPU + intel_extension_for_pytorch
Apple Silicon	AppleSiliconPlugin	MLX or torch MPS
Google TPU	TPUPlugin	JAX on TPU
Cloud QPU	QPUCloudPlugin	IonQ, IBM, Google, Rigetti, Azure, AWS

Design principle: No silent fallback. If a GPU plugin claims the work, it runs on the GPU or raises an error. The CPU never silently substitutes for a failed accelerator.

---

Circuit Knitting

Split circuits that exceed any single backend's capacity across multiple devices, then reconstruct the full probability distribution.

from vqpu import UniversalvQPU, CutFinder, CircuitKnitter

qpu = UniversalvQPU()

# A 10-qubit circuit too large for a 5-qubit device
circuit = qpu.circuit(10, "ghz_10")
circuit.h(0)
for i in range(9):
    circuit.cnot(i, i + 1)

# Automatically find where to cut and partition
plan = CutFinder.auto_partition(circuit, max_fragment_qubits=5)
print(f"Partitions: {[sorted(p) for p in plan.partitions]}")
print(f"Cuts: {plan.n_cuts}")

# Execute fragments independently, reconstruct exact result
knitter = CircuitKnitter(plan)
result = knitter.run(executor=qpu.run, shots=4096)
print(result.counts)
# {'0000000000': 2048, '1111111111': 2048}

How it works: For controlled gates (CNOT, CZ) cut at the control wire, the Z-basis decomposition is exact. The upstream fragment measures the control qubit normally; its value determines the downstream preparation. This produces the correct joint distribution with zero sampling overhead — no quasi-probability decomposition, no 4^k penalty.

You can also specify partitions manually:

plan = CutFinder.partition_manual(circuit, [{0,1,2,3}, {4,5,6,7}])

Or route different fragments to different backends via the link layer:

knitter = CircuitKnitter(plan)
result = knitter.run_heterogeneous(link_manager=lm, prefer={0: ["gpu"], 1: ["ionq"]})

---

Cryo-Canonical Basin Weaving Optimizer

An original variational optimization technique developed by Bernard Essuman. Based on published research combining spherical probing, 3-3+1 canonical reduction, and cold-seeking spring-network optimization.

from vqpu import UniversalvQPU
from vqpu.cryo import cryo_qaoa, CryoConfig

qpu = UniversalvQPU()

# Define a Max-Cut problem
edges = [(0, 1, 1.0), (1, 2, 1.2), (2, 3, 0.9), (0, 3, 0.8), (1, 3, 0.7)]

def cost(bitstring):
    return sum(-w for i, j, w in edges if bitstring[i] != bitstring[j])

# Run QAOA with CCBW optimization
result = cryo_qaoa(
    n_qubits=4,
    cost_function=cost,
    executor=qpu.run,
    p_layers=2,
    config=CryoConfig(
        n_force_levels=3,
        epsilon_max=0.4,
        epsilon_min=0.02,
        refine_steps=20,
    ),
)

print(f"Optimal energy: {result.optimal_energy:.4f}")
print(f"Certified basins: {len(result.graph.nodes)}")
print(f"Total evaluations: {result.n_evaluations}")

The CCBW algorithm:

1. Progressive probing — evaluate the cost landscape at multiple perturbation scales (coarse to fine), mapping the terrain before optimizing.
2. 3-3+1 motif — at each candidate point, probe 3 orthogonal directions plus their mirrors plus the center (7 evaluations). This structured probe detects whether the point sits in a real basin or on a saddle point.
3. Mirror balance — the symmetric part I(c) captures the true basin shape; the antisymmetric part A(c) reveals directional bias. High asymmetry means the candidate is a saddle point or artifact and is rejected.
4. Confidence certification — S(c) = alpha * compliance + beta * ||I|| - gamma * ||A||^2 - zeta * temperature. Only basins above threshold survive.
5. Spring graph — certified basins are connected into a graph with Gaussian-kernel edges weighted by temperature (cold nodes attract, hot nodes repel).
6. Cold-seeking refinement — gradient descent on the coldest (most stable, noise-tolerant) basins.

Why this matters for quantum computing: Standard VQE/QAOA optimizers use blind gradient descent from random starting points, frequently landing in barren plateaus or noise-sensitive sharp minima. CCBW maps the landscape structure first, rejects unstable points, and preferentially refines flat, symmetric basins that remain valid on real noisy hardware.

Convenience wrappers:

from vqpu.cryo import cryo_qaoa, cryo_vqe

# QAOA optimization
result = cryo_qaoa(n_qubits=4, cost_function=cost, executor=qpu.run, p_layers=2)

# VQE optimization
result = cryo_vqe(n_qubits=2, hamiltonian=H_matrix, executor=qpu.run, layers=2)

---

Phantom Adaptive Simulator

An adaptive simulation engine that dynamically selects the most efficient state representation for each subsystem during execution.

from vqpu import PhantomSimulatorBackend, UniversalvQPU

qpu = UniversalvQPU()
circuit = qpu.circuit(6, "factored")
circuit.h(0).cnot(0, 1).cnot(1, 2)   # entangled block: qubits 0-2
circuit.h(4).cnot(4, 5)               # separate block: qubits 4-5

phantom = PhantomSimulatorBackend(seed=42)
partition = phantom.build_partition(circuit)
result = phantom.execute(circuit, shots=1024)

Representations:

• Sparse statevector — tracks only non-zero amplitudes, prunes below threshold, maintains fidelity lower bound
• Matrix product state (MPS) — SVD-based factorization with configurable bond dimension
• Product state — bond dimension 1 for unentangled qubits

Dynamic adaptation: During circuit execution, the simulator automatically merges subsystems when gates create entanglement, splits them when Schmidt rank drops to 1, and demotes sparse representations to MPS when the active set fits within the bond dimension. This enables efficient simulation of circuits with structured entanglement patterns.

---

Persistent Link Layer

NVLink-inspired persistent connections to quantum backends with authentication, health monitoring, and automatic routing.

from vqpu.link import LinkManager, QuantumTask

lm = LinkManager()

# Local backend
from vqpu import CPUPlugin
lm.forge_local("cpu", CPUPlugin())

# Cloud backend
lm.forge_ionq("ionq", api_key="your-key", target_backend="simulator")

# Submit a task — routed to the best matching link
task = QuantumTask(n_qubits=4, gate_sequence=[("H", [0]), ("CNOT", [0, 1])], shots=1024)
counts, link = lm.submit(task, prefer=["ionq"])
print(f"Executed on: {link.handle}, state: {link.state.value}")

# Health monitoring
for snap in lm.snapshot():
    print(f"  {snap['handle']}: {snap['state']}, latency={snap['latency_ms']:.0f}ms")

# Clean up
lm.close_all()

Link states: UNLINKED, HANDSHAKING, LINKED, DEGRADED, ERROR, CLOSED. Credentials are held in memory only and scrubbed on close.

---

CHESSO Compiler

A control stack for quantum algorithm compilation, featuring the Q-lambda language frontend, hypergraph entanglement compilation, and a hardware bridge for translating compiled execution plans into native gate sequences for any backend.

from vqpu import chesso

The CHESSO subsystem includes 685 test files covering gate operations, compilation paths, and quantum algorithm experiments including AEGIS-Ion protein folding, TSP, and MaxCut QAOA benchmarks.

---

Architecture

Your application
    |
    |--- cryo.py        Cryo-Canonical Basin Weaving optimizer
    |--- knit.py        Circuit cutting and reconstruction
    |
    |--- universal.py   Backend discovery, fingerprinting, hybrid routing
    |--- link.py        Persistent authenticated backend connections
    |--- phantom.py     Adaptive sparse/MPS simulation engine
    |
    |--- core.py        Quantum primitives, gates, circuits, measurement
    |
    +--- Backends
         CPU  |  NVIDIA GPU  |  AMD GPU  |  Intel GPU  |  Apple Silicon
         TPU  |  IonQ  |  IBM  |  Google  |  Rigetti  |  AWS  |  Azure

---

Original Contributions

vQPU introduces several contributions that are not available in existing quantum computing frameworks:

Contribution	Description
Cryo-Canonical Basin Weaving	A novel variational optimizer based on original research. Uses 3-3+1 motif probing to certify basins, mirror-balance to reject saddle points, and cold-seeking spring networks to navigate the landscape. Not a reimplementation of existing work.
Zero-overhead circuit knitting	For controlled gates (CNOT, CZ) cut at the control wire, the Z-basis decomposition produces exact results with no sampling overhead. No quasi-probability penalty.
Phantom adaptive simulation	Combines sparse statevectors, MPS, and product states in a single engine with dynamic re-splitting during execution. Subsystems are automatically promoted, demoted, merged, and split as entanglement evolves.
Universal backend abstraction	True hardware-agnostic execution across CPU, GPU, TPU, and 6 cloud QPU providers through a single API. No silent fallback.
NVLink-inspired link layer	Persistent authenticated connections with state-machine lifecycle, health monitoring, and credential isolation.
Heterogeneous circuit execution	Circuit knitting combined with the link layer enables splitting a single circuit across GPU simulation and cloud QPU simultaneously.

---

Validation Results

All modules have been validated and the package has been live-tested on quantum cloud infrastructure.

Unit validation: 33/33 tests passing

Module	Tests	Result
Core (gates, circuits, measurement, simulator)	10	All pass
Universal (backend discovery, routing, plugins)	9	All pass
Knit (circuit cutting, reconstruction, TVD)	8	All pass
Cryo (CCBW optimizer, QAOA, VQE)	6	All pass

IonQ cloud live test:

• Backend: IonQ quantum simulator (36-qubit trapped-ion)
• Problem: 4-qubit weighted Max-Cut (5 edges)
• Total circuit evaluations: 320
• Approximation ratio: 89%
• Optimal bitstring hit rate: 72% of measurement shots
• Wall time: 41 minutes (cloud API latency)

Circuit knitting verification:

• 10-qubit GHZ state cut into two 5-qubit fragments reconstructs exactly (only |0000000000> and |1111111111>)
• 8-qubit entangled chain: total variation distance < 0.035 vs full simulation at 20,000 shots

---

What's New in 0.4.0

• Circuit knitting engine (knit.py) — topology-aware partitioning, exact Z-basis reconstruction for CNOT/CZ cuts, heterogeneous backend dispatch
• Cryo-Canonical Basin Weaving optimizer (cryo.py) — 3-3+1 motif probing, mirror-balance certification, cold-seeking refinement, cryo_qaoa() and cryo_vqe() convenience functions
• IonQ live validation — 320 circuit evaluations on IonQ's quantum cloud simulator
• Version bump to 0.4.0 — updated exports, pyproject.toml metadata, professional packaging

---

API Reference

Core primitives:

Export	Type	Description
vQPU	class	Simple quantum processor with local simulator
UniversalvQPU	class	Auto-discovering multi-backend processor
QuantumCircuit	class	Chainable circuit builder
QuantumRegister	class	N-qubit amplitude state vector
GateLibrary	class	Standard gate matrices (H, X, Y, Z, CNOT, etc.)
ExecutionResult	dataclass	Measurement counts, probabilities, metadata
QuantumAlgorithms	class	Bell pair, GHZ, QFT, Grover's, variational ansatz

Circuit knitting:

Export	Type	Description
CutFinder	class	Topology-aware partition finder
CircuitKnitter	class	Fragment execution and reconstruction
CutPlan	dataclass	Partition specification with cuts
KnitResult	dataclass	Reconstructed distribution

Cryo optimizer:

Export	Type	Description
CryoOptimizer	class	Full CCBW optimizer
CryoConfig	dataclass	Tuning parameters for CCBW
CryoResult	dataclass	Optimal parameters, basin graph, convergence
cryo_qaoa	function	QAOA with CCBW optimization
cryo_vqe	function	VQE with CCBW optimization

Backend and linking:

Export	Type	Description
BackendFingerprint	dataclass	Capacity, speed, gate set, connectivity
EntanglementScanner	class	Topology analysis (components, bridges)
PhantomSimulatorBackend	class	Adaptive sparse/MPS simulator
LinkManager	class	Backend connection registry and router
QuantumTask	dataclass	Unit of work for link submission

---

Requirements

• Python 3.10+
• NumPy >= 1.22
• Optional: backend-specific packages (see Installation)

---

License

MIT License. Copyright (c) 2026 Bernard Essuman. See LICENSE for details.

---

Author

Bernard Essuman
[e-mail]

vQPU is a research-driven project. Contributions, feedback, and collaboration inquiries are welcome.