---

name: validation-and-conversion description: "Guides ONNX checker, shape inference, parser, printer, version-converter, inliner, and CLI validation workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

ONNX Validation and Conversion

Use this sub-skill when a task is about checking whether an ONNX model or node is legal, inferring shapes or types, parsing or printing ONNX text, converting opset versions, or inlining model-local functions.

Workflow

1. Start with the smallest artifact that fails: a node, graph, or complete model.
2. Validate structural legality first with onnx.checker.check_node, check_graph, or check_model.
3. For model shape questions, call onnx.shape_inference.infer_shapes on a ModelProto. Use infer_shapes_path for file paths or very large models.
4. For ONNX text, keep syntax close to the grammar in references/cli-reference.md and references/workflows.md; parse with onnx.parser.parse_model or parse_graph, then validate the result.
5. For opset changes, use onnx.version_converter.convert_version only when the target path is supported; otherwise keep the original opset and surface the mismatch.
6. For local functions, use onnx.inliner.inline_local_functions or inline_selected_functions and then re-run checker/inference.
7. Use the check-model and check-node CLIs when the task is about a serialized protobuf and a safe command-line check is clearer than a Python snippet.

Route to References and Helpers

• Read references/api-reference.md for verified signatures and error classes.
• Read references/cli-reference.md for check-model, check-node, and backend-test-tools help output.
• Read references/workflows.md for validate/infer/parse/print/convert/inline recipes.
• Read references/troubleshooting.md before retrying a syntax, large-model, or conversion failure.
• Run scripts/validate_convert_model.py --help before using it on a model file or a generated tiny model.

Scope Boundaries

• Use model-authoring for constructing or saving models and tensors.
• Use reference-and-backend-tests when the question is whether the model produces expected outputs.
• Use operator-spec-maintenance for schema edits, C++ shape inference, function bodies, and node-test generation.