1 lightning.finite_horizon.studies.architecture_study

lightning.finite_horizon.studies.architecture_study

Run a frozen, paired Q1 architecture development study.

This owner composes existing diagnostic fits and evaluation receipts; it does not implement another optimizer loop. The statistical unit is a physical scene: candidate errors average within state, then states within scene, then scenes equally. TRAIN canaries are distinct from development validation and never construct a test dataset. All learned artifacts remain diagnostic-only.

1.1 Classes

Name Description
QhArchitectureStudyConfig Bind data, optimization and a finite architecture comparison family.

1.1.1 QhArchitectureStudyConfig

lightning.finite_horizon.studies.architecture_study.QhArchitectureStudyConfig()

Bind data, optimization and a finite architecture comparison family.

1.1.1.1 Attributes

Name Description
train TRAIN-only actor/label population; paths are explicit in this resolved config.
validation Exposed development validation; no test-stage field exists in this study.
experiment Native-Huber baseline; derived variants alter only the declared architecture.
seeds Both declared replication seeds; no seed search.
canary_updates Fixed final-update budget for each of the two TRAIN canary fits.
broad_updates Fixed final-update budget for each of the ten broad fits.
tie_band Native Q tolerance for ranking ties and canary optimal-action admission.
regret_fraction Prospective retention threshold relative to the relevant structural control.
mae_fraction Calibration safety bound relative to the matched control.
max_allocated_bytes Scorer process GPU allocator ceiling; pretrained preprocessing is outside this study.

1.2 Functions

Name Description
admit_canaries Gate broad fits using the original labels of one prospectively selected TRAIN row.
benchmark_scorer Measure a verified scorer on the canonical first TRAIN chain without updates.
freeze_study Bind TRAIN support, source and controls without constructing held-out data.
main Execute one explicit, immutable study operation from the command line.
prepare_development Seal exposed validation and TRAIN-only constant after the E0 gate passes.
report_study Compare all ten charged final cells; never select an intermediate checkpoint.
run_fit Fit one charged cell using canonical diagnostic admission and receipts.

1.2.1 admit_canaries

lightning.finite_horizon.studies.architecture_study.admit_canaries(
    config_path,
    output,
)

Gate broad fits using the original labels of one prospectively selected TRAIN row.

1.2.2 benchmark_scorer

lightning.finite_horizon.studies.architecture_study.benchmark_scorer(fit, train)

Measure a verified scorer on the canonical first TRAIN chain without updates.

Uses ten warmups and twenty synchronized, uninstrumented CUDA AMP forwards. One native H1 Huber backward identifies parameter tensors receiving gradients; it never constructs an optimizer or updates parameters. A separate instrumented pass times encoder and inclusive local readout; these timings have hook overhead and are not additive. Fusion is not an independently hookable module and remains unavailable. Returned resource peaks describe this process’s loaded diagnostic module (including its target scorer), not just online scorer tensors. The caller serializes GPU use and supplies an already admitted TRAIN configuration.

1.2.3 freeze_study

lightning.finite_horizon.studies.architecture_study.freeze_study(
    config_path,
    output,
)

Bind TRAIN support, source and controls without constructing held-out data.

1.2.4 main

lightning.finite_horizon.studies.architecture_study.main()

Execute one explicit, immutable study operation from the command line.

1.2.5 prepare_development

lightning.finite_horizon.studies.architecture_study.prepare_development(
    config_path,
    output,
)

Seal exposed validation and TRAIN-only constant after the E0 gate passes.

1.2.6 report_study

lightning.finite_horizon.studies.architecture_study.report_study(
    config_path,
    output,
)

Compare all ten charged final cells; never select an intermediate checkpoint.

1.2.7 run_fit

lightning.finite_horizon.studies.architecture_study.run_fit(
    config_path,
    output,
    *,
    variant,
    seed,
    canary_role=None,
)

Fit one charged cell using canonical diagnostic admission and receipts.

Back to top