FAQ¶
Is ONNX IR a runtime?¶
No. ONNX IR is a library for in-memory model representation, analysis, and transformation. It does not execute models like an inference runtime.
Does ONNX IR support the full ONNX spec?¶
It is designed to represent all valid ONNX protobuf models, plus a subset of invalid models to support repair workflows.
Should I use ONNX protobuf helpers directly?¶
Prefer ONNX IR APIs for model manipulation and conversions. This generally gives better ergonomics and avoids protobuf-heavy workflows in transformation code.
How should I handle large external tensors safely?¶
Use external tensor support with a configured base_dir when loading model
artifacts, especially from untrusted sources. This enables containment checks.
Is zero-copy always guaranteed?¶
Not always. ONNX IR is designed to minimize copies where possible, but some operations (such as certain conversions or materializations) may copy data.
Can I mutate a graph while iterating nodes?¶
Yes. ONNX IR is designed for robust mutation workflows and supports safe iteration patterns during graph edits. Nodes inserted after the current node are visited; nodes inserted before it are not. If the current node is removed or moved, iteration continues from the node that followed its original location.
How do I infer symbolic shapes?¶
Use the built-in ShapeInferencePass for ONNX shape inference. For richer SymPy
expressions, shape-data propagation, and custom operator inference, use the
optional onnx-shape-inference
package.
Where is the API reference?¶
See API Reference.