onnxscript

What it is and what it does

ONNX Script is a Python-to-ONNX compiler that lets you define ONNX functions and models using Python syntax decorated with the @script decorator. The decorator parses your Python code, traverses its abstract syntax tree, and builds an equivalent ONNX graph that can be saved and validated. It supports a subset of Python—not the full language—and includes eager-mode evaluation for debugging, where functions execute using ONNX Runtime as a shim to test intermediate results.

Beyond compilation, the package provides an ONNX Optimizer (constant folding, dead code elimination) and an ONNX Rewriter (pattern-based graph transformation using user-defined rules). The rewriter allows you to match subgraph patterns and replace them with optimized alternatives. All of this is built on top of onnx and onnx_ir, with support for modern Python versions (3.9 through 3.14).

Use it for:

• Author ONNX functions in readable Python instead of manually constructing protobuf graphs.
• Debug ONNX model logic by running functions in eager mode to inspect intermediate tensor values.
• Optimize ONNX models by applying constant folding and dead code elimination automatically.
• Rewrite ONNX graph patterns (e.g., replace a custom Erf-based GELU with a native Gelu operator).
• Round-trip between Python ONNX Script and ONNX graphs for model inspection and modification.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Write ONNX functions and models in Python syntax, then convert them to ONNX graphs; includes tools for optimization and pattern-based graph rewriting.

Yes. Active maintenance, low install friction, permissive MIT license, and no known vulnerabilities make this a safe choice. Install if you author or optimize ONNX models and want to work in Python syntax rather than protobuf. The eager-mode debugger is a real productivity gain for model development, though not for production inference.

Install

onnxscript on PyPI

pip install onnxscript

uv add onnxscript

poetry add onnxscript

Installing onnxscript

Before you install

Low friction: pure Python wheel with six runtime dependencies (numpy, onnx, packaging, and type-support libraries). Active maintenance with recent releases and 451 repository stars.

License in practice

MIT license permits unrestricted use, modification, and distribution in both open-source and commercial projects.

Quickstart

pip install onnxscript

from onnxscript import script, opset15 as op

@script()
def my_function(X):
    return op.MatMul(X, X)

onnx_model = my_function.to_model_proto()

Verify before relying

• Whether the eager-mode runtime performance is suitable for your debugging workflow (docs note it is not optimized for speed).
• Exact scope of Python language subset supported beyond the examples shown.

Package facts

Evidence: onnxscript-0.7.1-py3-none-any.whl

Tags

More Artificial Intelligence packages