onnx/ir-py

ir-py is an in-memory intermediate representation for ONNX that enables efficient graph construction, analysis, and transformation in Python.

The tool addresses the need for performant manipulation of ONNX computation graphs without the overhead of protobuf serialization during editing. It provides a unified interface for working with ONNX models, supporting the full ONNX specification and even allowing loading of technically invalid models for repair. The approach uses memory-mapped external tensors and zero-copy access to tensor data, eliminating size limitations while maintaining a low memory footprint. Graph traversal and value information access are designed to be straightforward, and the implementation supports concurrent iteration and mutation through a linked-list-based node container.

Developers should choose this tool when working on graph transformation pipelines, model optimization, or analysis tasks where repeated serialization and deserialization would create bottlenecks. It suits projects involving large language models or other scenarios requiring substantial graph manipulation. The tool decouples from protobuf once a model is loaded into the IR representation, removing that dependency from the runtime environment. The Pythonic API design maps intuitively to ONNX protobuf concepts while providing familiar patterns for Python developers.

The project maintains active development with regular commits addressing bug fixes and feature improvements. Pull requests receive timely review and integration. The codebase is well-organized with clear separation of concerns across modules handling protocols, core entities, type definitions, naming authority, internal data structures, and metadata storage. Documentation is actively maintained alongside code changes, with examples and API references kept current.