Skip to content

Pack hybrid training jobs and add robust expectiles and typed arithmetic - #53

Merged
OpenUSA merged 1 commit into
mainfrom
feature/job-aware-training-expectile-arithmetic
Nov 13, 2026
Merged

OpenUSA merged 1 commit into
mainfrom
feature/job-aware-training-expectile-arithmetic

Conversation

@OpenUSA

@OpenUSA OpenUSA commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

Result

OpenQML 0.27.0 adds submission-aware hybrid training, coherent expectile risk and a broader typed classical-expression compiler without new core dependencies.

  • ShiftPlan.linearize packs primal readouts and occurrence-wise derivative recipes into bounded backend batches. Its owned ShiftLinearization supports JVP/VJP reuse with explicit circuit and submission counts. Torch/JAX/Flax quantum layers opt in with gradient_method="batched-shift"; broadcast gradients reduce correctly to shared inputs/weights. Undefined selected derivatives and owned working-memory limits are checked before submission.
  • Nondeterministic JAX provider calls carry upstream I/O effects through the first-order derivative primitive. JIT retains unused calls and does not merge equal submissions. TrainingWorkload.joint_primal_gradient prices the combined phase while preserving sequential optimizer steps, microbatches and provider limits.
  • finance.expectile, wasserstein_expectile and expectile_budget_projection retain native gradients through quantum scenario probabilities, contextual radii and anisotropic transport geometry. Sorted piecewise-affine roots avoid iterative differentiation; normalized coefficients preserve tiny contributing losses beside zero-mass extreme atoms. The funded cash-ray scale remains maximal, with a distinct ExpectileBudgetResult.
  • Qiskit Uint addition/subtraction/multiplication lower with declared-width modular overflow. Logical shifts and indexing support selectors whose range is provably valid. Ripple-carry, truncated multiplication and barrel-selection networks feed the existing bounded shared decision compiler, preserving branch selection before selector writes. Work/expanded-output limits and explicit diagnostics remain in force.

Basis and scope

Job packing follows Training quantum neural networks one job at a time, excluding its sinusoid-search optimizer and hardware performance claims. A full stencil is one backend submission only when chunk/provider limits admit it. The opt-in linearization trades O(batch × readouts × parameters) memory for fewer round trips; the default streamed derivative path is preserved. External higher quantum derivatives remain explicitly unsupported.

The robust expectile specializes the envelope mechanism in Multi-period Mean-Expectile Portfolio Optimization under Wasserstein Ambiguity to a one-period unrestricted-support W1 ball. Fixed-mean constraints and the paper's multi-period optimization programs are excluded. Radius/metric calibration and chronological financial evaluation remain caller responsibilities.

Typed operations follow Qiskit's classical API; callback behavior follows JAX's I/O contract. Division, signed/float arithmetic, mutable storage and unproven selector ranges retain diagnostics.

Validation

  • 72 focused numerical, transform, API, submission-accounting and conversion checks passed.
  • 337 related regressions passed; two physical-CUDA cases skipped.
  • Torch 2.2.2 + JAX 0.4.30: 52 passed; Qiskit/Flax unavailable in that local environment.
  • Torch 2.4.1: 47 passed; seven unavailable-extra skips.
  • Actual Qiskit StatevectorEstimator submission counting, Torch/JAX/Flax gradients, stochastic JAX effect retention, independent double-precision financial references, explicit adversarial distributions and exhaustive finite-width truth tables were checked. Lowered arithmetic also executes through native Aer equality feedforward.
  • Ruff, formatting, diff checks, wheel/sdist build, Twine metadata checks and documented examples from the wheel passed.
  • GitHub feature job: 84 passed, five optional/hardware skips; legacy-frameworks: 287 passed, 16 skips; compiled-transforms: 86 passed, 28 skips.
  • Linux/Windows core, tensor distribution, public types and lint jobs passed. Full platform matrix remains in progress.
  • Real QPU/GPU throughput and out-of-sample investment performance are not established by these software checks.

@OpenUSA
OpenUSA merged commit c9fd7e9 into main Nov 13, 2026
15 checks passed
@OpenQML-org OpenQML-org deleted a comment from chatgpt-codex-connector Bot Nov 14, 2026
@OpenUSA
OpenUSA deleted the feature/job-aware-training-expectile-arithmetic branch November 14, 2026 02:45
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