Skip to main content

validate_model

onnx_runtime_loader

Function validate_model 

Source 
pub fn validate_model(graph: &Graph) -> Result<(), LoaderError>
Expand description

Fail-fast, load-time validation of everything statically knowable to be illegal or unsupported (RULES #1: fail at load, never via a silent sentinel at run time).

This is the single cohesive entry point wired into both load paths — the disk/bytes loader ([build_from_bytes_with_weights]) and the session’s programmatic entry ([onnx_runtime_session]’s from_parts/from_graph) — so the checks cannot drift between the two. It runs, in order:

Protobuf-only invariants (ir_version, raw SSA names, ref_attr_name, subgraph shadows, and output names) run earlier in validate_model_proto, before graph construction coalesces names or drops protobuf-only fields. This IR-level phase then runs:

  1. validate_opset_imports — every node’s domain must declare an opset.
  2. validate_no_control_flow — allow the implemented subgraph-bearing ops (If/Loop/Scan) and reject any other op carrying a GraphProto attribute the executor cannot run.
  3. validate_no_dangling_refs — every consumed tensor must be sourced (graph input, initializer, or an upstream node output).
  4. validate_no_initializer_producer — an initializer must be a constant source; reject any initializer value that is also a node output (shares a ValueId with a producer), which the IR structural check does not cover.

Each rejection names the offending node/op/tensor and explains what is expected. No sentinel defaults, no silent skips.

Structural invariants that the IR builder already enforces via onnx_runtime_ir::Graph::validate at build time — duplicate output names, dangling value ids, producer/consumer link consistency, and data dependency cycles — are intentionally not re-checked here to avoid drift; this function adds the checks that path does not cover.