← All Projects

qubit-os

Quantum ComputingRustPythongRPCPulse ControlClosed-LoopGitHub →

What it is

QubitOS is an open-source quantum control kernel: the layer between a compiler and the hardware. Hamiltonian in, optimized pulses out, calibration-aware, with provenance baked in.

The bet is structural: as quantum hardware diversifies, vendors will not agree on a control stack. An open kernel that speaks pulses and feedback is the natural seam.

Shipped through v0.7.0

The architecture is in place and exercised end-to-end against a simulator. v0.7.0 (released 2026-05-16) added the closed-loop Lyapunov feedback controller and the open-loop vs closed-loop comparison framework; v0.6.0 added the stochastic master equation runtime; v0.5.0 and below built the open-loop control stack.

  • GRAPE optimizer — multi-qubit, per-qubit envelopes, dimension-scaled learning rates. Single-qubit gates at 99.9%+ fidelity in simulation.
  • Rust HAL server — gRPC + REST, backend registry, request validation.
  • Lindblad solver — Rust RK4 path with Rayon-parallel ensembles. A bare-metal C SIMD fast path for d ≤ 27 is on the stretch track tied to the LANL summer 2026 deliverable.
  • Stochastic Master Equation solver (v0.6.0) — Itô SDE integrator with Euler-Maruyama steps, Monte Carlo ensemble with Rayon parallelism, cross-validated against QuTiP smesolve to atol = 0.05 on ensemble expectations. Python reference plus Rust port.
  • Calibration management — fingerprinting, drift detection, decoherence budgets. Calibration is treated as a continuous process, not a snapshot.
  • Provenance — Merkle-tree experiment tracking for reproducible runs.
  • agentbible decoration — numerical-attestation backbone wired into every density-matrix, unitary, and probability-vector boundary in the runtime.

Shipped in v0.7.0

v0.7.0 closes the loop. It turns the runtime into an open-loop vs closed-loop comparison engine.

  • Lyapunov feedback controller, V(ρ_c) = 1 − Tr[ρ_target · ρ_c]; single-axis and diagonal multi-axis (independent K_x, K_y, K_z) on the validated path. Full 3×3 cross-axis gain is shipped as a documented opt-in surface.
  • SME integration: solve_with_feedback wraps the SME step plus controller step plus Hamiltonian update loop. Feedback delay flows through the existing temporal/ TemporalConstraint and DecoherenceBudget machinery rather than as a parallel timing concept.
  • Comparison framework: noise_sweep_comparison runs GRAPE, DRAG, Gaussian, and Lyapunov feedback over a noise sweep; crossover_point returns the γ* where feedback fidelity equals GRAPE fidelity.
  • Stability analysis: Lyapunov-function trajectory plotting, Bloch-sphere trajectory visualization (via qutip.Bloch), convergence-rate fitting.

The Rust port of the feedback law lands the same release with a 151× speedup over the Python reference and full agreement to five significant figures.

In progress: v0.8.0

v0.8.0 extends the closed-loop runtime to a 3-level transmon with leakage-aware Lyapunov feedback. Tracks a 2-qubit interacting-systems extension under joint parity measurement as part of the same milestone.

Backends

  • QuTiP simulator — primary local backend, Lindblad master equation, validation oracle.
  • IQM Resonance — primary live integration target; the only backend exercised against real hardware.
  • IBM Quantum, AWS Braket — demonstrated through the QuantumBackend trait abstraction (mock-tested). These are not maintained as production cloud integrations.

Architecture

Python CLI / notebooks
   |
   v
core (GRAPE, SME, calibration, scheduling)
   |  --local: calls QuTiP directly
   |  --remote: gRPC to HAL server
   v
hal (Rust gRPC server, backend dispatch)
   |
   v
Backends (IQM live; QuTiP, IBM, AWS demonstrated)

The system is pulse-first: Hamiltonian → pulse → measurement, not gate → pulse. This preserves physics that gate-level abstractions discard (decoherence during gates, crosstalk, leakage, feedback latency).

Why a kernel, not a distribution

The scope is deliberately narrow. QubitOS aims to own the control kernel — pulses, feedback, calibration, provenance — and refuses to be a full-stack quantum platform. Compilers, error-correction codes, application libraries are all consumers, not internals.

That posture is the integrity claim. The README distinguishes "primary live integration" from "architecture demonstrated through the abstraction." No middle category.

Repo

License

Apache 2.0.