NeatapticTS

architecture/network/onnx/parity

architecture/network/onnx/parity/network.onnx.parity.types.ts

OnnxRuntimeParityExecutedFixtureDescriptor

Executable fixture descriptor for runtime-parity lanes that are already approved.

This narrows the generic fixture descriptor to the executed subset so callers such as the Phase 9C randomized runner do not need unreachable skipped-result guards after they create lane-approved samples.

OnnxRuntimeParityExecutedResult

Result for one executed Phase 9 runtime-parity fixture.

The harness reports both native and ONNX Runtime outputs so later golden and randomized passes can reuse the same comparison packet without changing the execution seam.

OnnxRuntimeParityExecutionMode

Execution state for one Phase 9 runtime-parity fixture.

execute means the current harness should export, run, and compare the binary model immediately. skip keeps the lane in the frozen inventory but requires an explicit follow-up tranche before execution begins.

OnnxRuntimeParityFixtureDescriptor

Deterministic fixture descriptor for one Phase 9 runtime-parity lane.

Each descriptor couples a stable network factory, the binary export options, one native runtime sample, the ONNX Runtime feed shape, and the tolerance packet that later golden or randomized passes must honor.

OnnxRuntimeParityLane

Named Phase 9 runtime-parity lanes for the current ONNX execution boundary.

Phase 9A freezes this list before golden parity cases widen coverage. Each lane names one exporter-owned subset whose binary .onnx artifact can later be compared against Network.activate() with an explicit tolerance packet.

Example:

const baselineLane: OnnxRuntimeParityLane = 'baseline-float32-dense';

OnnxRuntimeParityRandomizedRunOptions

Seeded randomized-parity options for one Phase 9C sample run.

Phase 9C keeps the runtime subset narrow, but broadens evidence from named golden fixtures into reproducible randomized samples. The seed and sample count make failures replayable.

OnnxRuntimeParityRandomizedSampleResult

Flattened result for one seeded randomized parity sample.

The Phase 9C runner returns a compact packet per sample so tests can freeze deterministic summaries without reimplementing the execution seam.

OnnxRuntimeParityResult

Runtime-parity result union type for the Phase 9 test harness seam.

OnnxRuntimeParitySkippedResult

Result for one inventory-only skipped Phase 9 runtime-parity test fixture.

OnnxRuntimeParityTolerancePacket

Explicit tolerance packet for one runtime-parity lane.

Phase 9 keeps tolerances named and lane-specific so reduced-precision cases cannot silently inherit looser thresholds from unrelated fixtures.

architecture/network/onnx/parity/network.onnx.parity.ts

assignDeterministicParameters

assignDeterministicParameters(
  network: default,
  parameterSeed: DeterministicParameterSeed,
): void

Apply deterministic scalar parameters to a network fixture.

Parameters:

buildExecutedResult

buildExecutedResult(
  fixtureDescriptor: OnnxRuntimeParityFixtureDescriptor,
  inputNames: readonly string[],
  outputNames: readonly string[],
  nativeOutput: number[],
  runtimeOutput: number[],
): OnnxRuntimeParityExecutedResult

Build the executed runtime-parity result packet for one fixture.

Parameters:

Returns: Executed comparison packet for the current fixture.

calculateMaxAbsoluteDifference

calculateMaxAbsoluteDifference(
  nativeOutput: readonly number[],
  runtimeOutput: readonly number[],
): number

Calculate the maximum absolute scalar difference for one parity comparison.

Parameters:

Returns: Maximum absolute difference across the compared output vector.

calculateMeanSquaredError

calculateMeanSquaredError(
  nativeOutput: readonly number[],
  runtimeOutput: readonly number[],
): number

Calculate mean squared error for one parity comparison.

Parameters:

Returns: Mean squared error across the compared output vector.

createBaselineFloat32DenseFixture

createBaselineFloat32DenseFixture(): OnnxRuntimeParityFixtureDescriptor

Create the Phase 9A dense baseline fixture with float32 export and strict tolerances.

Returns: Baseline same-family dense runtime-parity descriptor.

createConvGroundworkScenario

createConvGroundworkScenario(): { createNetwork: () => default; mappings: Conv2DMapping[]; nativeInputValues: number[]; runtimeInputValues: number[]; }

Create the deterministic explicit-Conv groundwork scenario reused by Phase 9.

Returns: Deterministic network factory, mappings, and sample inputs for the explicit-Conv lane.

createDynamicUint8DenseGuidanceFixture

createDynamicUint8DenseGuidanceFixture(): OnnxRuntimeParityFixtureDescriptor

Create the dynamic-uint8 dense fixture that validates DynamicQuantizeLinear guidance export.

Returns: Dynamic-guidance runtime-parity descriptor with an explicit tolerance packet.

createRandomizedDenseFixture

createRandomizedDenseFixture(
  fixtureDescriptor: OnnxRuntimeParityFixtureDescriptor,
  sampleSeed: number,
  sampleIndex: number,
): OnnxRuntimeParityExecutedFixtureDescriptor

Create one seeded same-family dense randomized fixture.

Parameters:

Returns: Lane-compatible randomized dense fixture descriptor.

createRandomizedNumericVector

createRandomizedNumericVector(
  sampleGenerator: () => number,
  vectorLength: number,
  numericRange: NumericRange,
): number[]

Create a reproducible numeric vector for one randomized parity sample.

Parameters:

Returns: Deterministic numeric vector for the sample input.

createRandomizedParameterSeed

createRandomizedParameterSeed(
  sampleGenerator: () => number,
  parameterSeedBounds: { weightStart: NumericRange; weightStep: NumericRange; biasStart: NumericRange; biasStep: NumericRange; },
): DeterministicParameterSeed

Create a reproducible arithmetic parameter seed packet.

Parameters:

Returns: Deterministic scalar parameter packet for one randomized network.

createRandomizedParityFixture

createRandomizedParityFixture(
  fixtureDescriptor: OnnxRuntimeParityFixtureDescriptor,
  sampleSeed: number,
  sampleIndex: number,
): OnnxRuntimeParityExecutedFixtureDescriptor

Materialize one lane-approved randomized fixture descriptor.

Parameters:

Returns: Randomized fixture descriptor that stays within the declared lane boundary.

createRandomizedStatic8BitConvFixture

createRandomizedStatic8BitConvFixture(
  fixtureDescriptor: OnnxRuntimeParityFixtureDescriptor,
  sampleSeed: number,
  sampleIndex: number,
): OnnxRuntimeParityExecutedFixtureDescriptor

Create one seeded explicit-Conv static-8bit randomized fixture.

Parameters:

Returns: Lane-compatible randomized explicit-Conv fixture descriptor.

createRuntimeParityTensor

createRuntimeParityTensor(
  arg0: unknown,
  arg1: unknown,
  arg2: readonly number[] | undefined,
): Tensor

Create one tensor instance that matches the raw binding's float32 output quirks.

The current Node binding can surface float32 outputs through ArrayBuffer or non-Float32Array views. This helper normalizes those payloads before they reach the standard Tensor constructor.

Parameters:

Returns: A standard ONNX Runtime tensor instance.

createSeededUnitIntervalGenerator

createSeededUnitIntervalGenerator(
  initialSeed: number,
): () => number

Create one reproducible unit-interval generator from a 32-bit seed.

Parameters:

Returns: Deterministic floating-point generator in the range [0, 1).

createStatic8BitConvQlinearFixture

createStatic8BitConvQlinearFixture(): OnnxRuntimeParityFixtureDescriptor

Create the explicit-conv static-8bit fixture for conv-lane runtime parity checks.

Returns: Static-8bit explicit-Conv runtime-parity descriptor.

createStatic8BitDenseQlinearFixture

createStatic8BitDenseQlinearFixture(): OnnxRuntimeParityFixtureDescriptor

Create the static-8bit dense qlinear fixture with external calibration metadata.

Returns: Static-8bit dense qlinear runtime-parity descriptor.

createStorageFp16DenseFixture

createStorageFp16DenseFixture(): OnnxRuntimeParityFixtureDescriptor

Create the storage-fp16 dense fixture used to validate cast-bridge parity behavior.

Returns: Storage-fp16 cast-bridge runtime-parity descriptor.

cycleIntegerRangeBySampleIndex

cycleIntegerRangeBySampleIndex(
  sampleIndex: number,
  numericRange: NumericRange,
): number

Cycle deterministically through an inclusive integer range by sample index.

Parameters:

Returns: Deterministic integer derived from the sample position.

executeRuntimeParityInSubprocess

executeRuntimeParityInSubprocess(
  binaryModel: Uint8Array<ArrayBufferLike>,
  fixtureDescriptor: OnnxRuntimeParityFixtureDescriptor,
): OnnxRuntimeParitySubprocessResult

Execute one ONNX Runtime parity case in a fresh child Node process.

This keeps the raw binding isolated from the Jest process so Phase 8 runtime validation and Phase 9 parity execution do not compete over native addon initialization order.

Parameters:

Returns: Runtime output packet emitted by the child process.

findPhase9RuntimeParityFixture

findPhase9RuntimeParityFixture(
  fixtureId: OnnxRuntimeParityLane,
): OnnxRuntimeParityFixtureDescriptor

Resolve one named Phase 9 runtime-parity fixture from the frozen inventory.

Parameters:

Returns: The matching deterministic fixture descriptor.

getPhase9RuntimeParityInventory

getPhase9RuntimeParityInventory(): readonly OnnxRuntimeParityFixtureDescriptor[]

Return the frozen Phase 9A runtime-parity inventory.

This inventory is the narrow stop line between the closed Phase 8 binary acceptance lane and the later Phase 9B golden parity tranche. Keeping the list centralized prevents new runtime lanes from slipping in through ad hoc tests.

Returns: The current Phase 9A fixture inventory in deterministic order.

isArrayBufferLike

isArrayBufferLike(
  tensorData: unknown,
): boolean

Check whether one runtime payload should be treated as an ArrayBuffer source.

Parameters:

Returns: True when the payload is a plain or shared ArrayBuffer-like object.

resolveRuntimeSubprocessError

resolveRuntimeSubprocessError(
  subprocessResult: { stderr: string; stdout: string; },
): string

Resolve a stable parity-subprocess error message from stderr, stdout, or fallback text.

Parameters:

Returns: Human-readable error message.

resolveRuntimeTensorDimensions

resolveRuntimeTensorDimensions(
  tensorShape: readonly (string | number)[],
): number[]

Resolve runtime feed dimensions from the fixture packet or session metadata.

Parameters:

Returns: Numeric input dimensions suitable for the feed tensor constructor.

resolveValidatedRandomizedSampleCount

resolveValidatedRandomizedSampleCount(
  sampleCount: number,
): number

Resolve the allowed randomized sample count for Phase 9C.

Parameters:

Returns: Validated sample count within the supported Phase 9C bounds.

runOnnxRuntimeParityFixture

runOnnxRuntimeParityFixture(
  fixtureDescriptor: OnnxRuntimeParityExecutedFixtureDescriptor,
): Promise<OnnxRuntimeParityExecutedResult>

Export, execute, and compare one Phase 9 runtime-parity fixture.

The harness always starts from binary .onnx bytes so Phase 9 stays binary-first rather than drifting back to the JSON-first roundtrip lane. Skipped fixtures are returned with their explicit reason instead of being executed implicitly.

Parameters:

Returns: Either a skipped result or an executed comparison packet.

runSeededOnnxRuntimeParitySamples

runSeededOnnxRuntimeParitySamples(
  fixtureDescriptor: OnnxRuntimeParityFixtureDescriptor,
  randomizedRunOptions: OnnxRuntimeParityRandomizedRunOptions,
): Promise<OnnxRuntimeParityRandomizedSampleResult[]>

Execute reproducible randomized parity samples for one approved Phase 9 lane.

Phase 9C reuses the existing binary-first fixture seam, but broadens the evidence from one named golden sample into several seeded randomized cases. The runner keeps the subset narrow by generating only lane-approved shapes.

Parameters:

Returns: Flattened executed results for each seeded randomized sample.

sampleNumberInRange

sampleNumberInRange(
  sampleGenerator: () => number,
  numericRange: NumericRange,
): number

Sample one rounded floating-point value inside a numeric range.

Parameters:

Returns: Rounded floating-point sample inside the declared range.

Generated from source JSDoc • GitHub