Skip to content

Upgrade hybrid AI geometry, robust quant decisions and typed control conversion - #52

Merged
OpenUSA merged 1 commit into
mainfrom
feature/hybrid-quant-compiler-026
Nov 13, 2026
Merged

OpenUSA merged 1 commit into
mainfrom
feature/hybrid-quant-compiler-026

Conversation

@OpenUSA

@OpenUSA OpenUSA commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

Result

OpenQML 0.26.0 adds three concrete capabilities without additional core dependencies:

  • Hybrid AI: device-resident Torch/JAX Fubini–Study metric-vector products and damped natural-gradient solves, including a weighted sample axis, true residual certificates and phase-null directions. Optional geometry-first active-polar mixing operates after preconditioning and bounds the final displacement.
  • Quantitative research: learnable contextual Wasserstein radii and diagonal-plus-factor transport metrics. Implicit native derivatives connect quantum Born probabilities, robust CVaR and maximal funded cash-ray allocation without a dense asset inverse.
  • Interoperability: bounded bit-vector/ROBDD compilation of Qiskit Bool/Uint logical, bitwise, comparison and cast expressions into the existing portable equality-guard IR. Predicates finish before either body runs, even when that body overwrites a selector bit.

Intel macOS CI now installs and exposes libomp for TorchInductor. A missing OpenMP header produces an actionable ResourceError while unrelated C++ failures retain their original diagnostics. The legacy Torch 2.2.2 / compiled-transform Torch 2.4.1 jobs and portable OS jobs cover the new native paths; a dedicated fast job covers all three additions.

Basis and scope

The spectral step implements geometry-first active-polar mixing from Quantum Natural Muon, with an alignment safeguard. It is stateless and excludes the paper's momentum and trajectory schedule. The nested certificate applies to the original damped solve.

Context-dependent ambiguity geometry follows the direction in Learning Predictive Ambiguity Sets, specialized here to W1/CVaR and funded cash allocation. Radius/metric calibration and chronological out-of-sample evaluation remain necessary; no return or coverage guarantee is asserted.

The portable expression subset follows Qiskit's typed classical API. General mutable classical storage and unsupported arithmetic retain explicit diagnostics or require the native program route.

Validation

  • New contracts: 50 passed, one physical-CUDA test skipped, on current Torch/JAX with Qiskit/Aer/Cirq.
  • Related regressions: 301 passed.
  • Torch 2.2.2 compatibility: 36 passed, 18 unavailable-extra/hardware skips.
  • Torch 2.4.1 compatibility: 29 passed, seven unavailable-extra/hardware skips.
  • Ruff, formatting, diff checks, wheel/sdist build and Twine metadata checks passed.
  • Dedicated GitHub feature job: 52 passed, two hardware/optional-integration skips.
  • GitHub legacy-frameworks: 264 passed, 14 optional/hardware skips; compiled-transforms: 67 passed, 22 optional/hardware skips.
  • Early compiler/native validation passed on Windows, macOS ARM and Intel macOS, including real generated-C++ readout compilation. macOS OpenMP setup succeeded.
  • Both newly documented examples executed from the built wheel; a combined quantum-probability → robust allocation → spectral-natural update smoke test converged and met the configured risk budget.
  • Full GitHub platform regression matrix and full local regression are still running. Three local Gloo/DDP cases are excluded because this managed environment denies the required TCP sockets; the dedicated GitHub tensor-distribution job passed.
  • Physical GPU throughput, QPU execution and financial performance are not claimed.

@OpenUSA
OpenUSA merged commit d58727e into main Nov 13, 2026
15 checks passed
@OpenQML-org OpenQML-org deleted a comment from chatgpt-codex-connector Bot Nov 13, 2026
@OpenUSA
OpenUSA deleted the feature/hybrid-quant-compiler-026 branch November 13, 2026 14:49
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant