Module check 

Extensible model checker / validator (ONNX_RS §8).

The architecture mirrors §8.1: a ValidationRule trait, a Severity / Violation model, and an OnnxChecker registry that runs an ordered set of rules and can be extended with custom ones (§8.3). The built-in rules include schema-driven node conformance (see [rules]).

Note: the runtime onnx_runtime_loader already rejects many malformed models at load time (dangling refs, duplicate producers, missing opset imports, cycles, …). This checker is complementary: it runs against an already-built Model so tools can validate graphs they constructed in memory and so organisations can layer their own rules on top.

Structs

AttributeProtoValidityRule

Validate attribute names, discriminators, union payloads, and per-node attribute-name uniqueness.

DuplicateValueNameRule

No two distinct values may carry the same (non-empty) name — ONNX graphs are SSA, so a name identifies a unique edge (ONNX_RS §8.2 UniqueValueNamesRule).

FunctionProtoValidityRule

Validate model-local functions using the v1.20 checker topology/import rules plus signature, default-attribute, and unique-ID rules.

GraphAcyclicRule

The dataflow graph must be acyclic (ONNX_RS §8.2 GraphAcyclicRule). Wraps the IR’s own topological-order check.

InitializerTypeMatchesDeclaredRule

Initializer tensor element types must match their graph value declarations (ONNX_RS §8.2 InitializerTypeMatchesDeclaredRule).

InputOutputDeclaredRule

Graph inputs and outputs must reference live, named values (ONNX_RS §8.2 InputOutputDeclaredRule). The shared IR always carries dtype and shape for a live value, so liveness plus a non-empty ONNX name establishes declaration.

IrVersionFeatureRule

Validate fields and invariants whose legality changed at a particular ONNX IR version. Training-only fields are intentionally ignored.

IrVersionSupportedRule

The ONNX IR version is required and must not exceed the version implemented by this v1.20-bound checker (IR 13).

MetadataKeysUniqueRule

Metadata maps are encoded as repeated entries and therefore require an explicit distinct-key check at every protobuf scope.

MissingOpsetImportRule

Every operator domain referenced by a node must be declared in opset_imports (ONNX_RS §8.2, IR rule OpsetImportPresentRule).

MultiDeviceConfigurationRule

Validate the IR v11+ multi-device configuration and sharding annotations.

NoUnconnectedNodesRule

Every present node input must have a source, and every named node output must be consumed, exported, or captured by a subgraph (ONNX_RS §8.2).

OnnxChecker

The standard, extensible checker (ONNX_RS §8.2).

ProtoTypeValidityRule

Validate every retained TypeProto, including container requirements and ONNX-ML opaque types.

SchemaNodeConformsRule

Each node must resolve to an opset-compatible schema and conform to its input/output and attribute declarations (ONNX_RS §8.2).

SparseTensorValidityRule

Validate sparse tensor COO structure, index type/shape, bounds, ordering, and uniqueness.

TensorPayloadValidityRule

Validate dense tensor dimensions, storage-field selection, payload size, segment bounds, and external-data constraints.

TypeConstraintSatisfiedRule

Node input/output element types must satisfy and consistently bind the type variables declared by the resolved operator schema (ONNX_RS §8.2).

ValidationContext

Shared context handed to every rule during a check.

ValidationResult

The outcome of running a checker (ONNX_RS §8.2).

Violation

A single problem found by a ValidationRule.

Enums

Severity

Severity of a Violation (ONNX_RS §8.1).

ViolationLocation

Where a Violation was found (ONNX_RS §8.1).

Traits

ValidationRule

A rule that checks one aspect of a model (ONNX_RS §8.1).