Skip to content

Repository files navigation

onnx-shape-inference

PyPI - Version PyPI - Python Version codecov Ruff PyPI Downloads

Experimental symbolic shape inference for ONNX models. Built on top of ONNX IR, this library performs shape inference directly on the IR without serialization overhead, using SymPy for symbolic dimension arithmetic.

See the released operator support audit for ONNX opset 28 / ONNX Runtime 1.31 coverage and remaining limitations.

Features

  • Symbolic shape inference — propagates shapes through the graph using SymPy expressions for symbolic dimensions
  • Shape data propagation — tracks known element values of shape tensors (e.g. through Shape → Slice → Concat → Reshape chains) to resolve concrete output shapes that standard shape inference cannot
  • Broad operator coverage — built-in inference for standard ONNX (ai.onnx) operators plus com.microsoft contrib ops, with version-aware dispatch across opset history
  • Symbolic constraint resolution — reconciles engine-generated dimension names with the symbolic names an author declares on graph outputs / value_info, renaming anonymous dims (including compound expressions like 2*_d0 or past_seq + seq) to the declared names
  • Extensible registry — register custom shape inference functions for custom operators
  • Merge policies — control how newly inferred shapes are merged with existing ones: refine (default), strict, override, and skip

Installation

pip install onnx-shape-inference

Or install from source (main branch):

pip install git+https://github.com/justinchuby/onnx-shape-inference.git

Command line

Run shape inference on a model and see how many new shapes were inferred:

onnx-shape-inference model.onnx

Save the inferred model to a file:

onnx-shape-inference model.onnx -o model_inferred.onnx

Overwrite the input model in place:

onnx-shape-inference model.onnx --in-place

Select a different merge policy:

onnx-shape-inference model.onnx --policy strict

Usage

import onnx_ir as ir
from onnx_shape_inference import infer_symbolic_shapes

# Load a model
model = ir.load("model.onnx")

# Run shape inference
model = infer_symbolic_shapes(model)

# Or with a strict merge policy
model = infer_symbolic_shapes(model, policy="strict")

Use with onnxscript optimizer

You can run symbolic shape inference on the model to help the optimizer discover more optimization opportunities.

import onnx_shape_inference
import onnx_ir as ir
import onnxscript.optimizer

model = ir.load("model.onnx")

# Provide more shape information with infer_symbolic_shapes
model = onnx_shape_inference.infer_symbolic_shapes(model)

# onnxscript optimizer can leverage this information to better optimize the model
onnxscript.optimizer.optimize(model)

ir.save(model, "model_optimized.onnx")

Per-node inference

You can run shape inference on individual nodes by using the ShapeInferenceContext and registry directly. This is useful for debugging, testing, or integrating into custom graph passes.

import onnx_ir as ir
from onnx_shape_inference import ShapeInferenceContext, registry

# Populate the registry with all built-in ops
registry.collect()

# Create a context with the model's opset imports
ctx = ShapeInferenceContext(opset_imports={"": 21})

# Look up the inference function for the op
infer_func = registry.get("", "Relu", version=21)

# Build a node (or get one from an existing graph)
x = ir.Value(name="x", shape=ir.Shape([2, 3]), type=ir.TensorType(ir.DataType.FLOAT))
y = ir.Value(name="y")
node = ir.Node("", "Relu", inputs=[x], outputs=[y])

# Run inference
infer_func(ctx, node)

print(y.shape)  # [2,3]
print(y.dtype)  # FLOAT

Registering custom operators

from onnx_shape_inference import registry

@registry.register("com.custom", "MyOp", since_version=1)
def infer_my_op(ctx, node):
    input_shape = node.inputs[0].shape
    output_shape = ir.Shape([...])
    ctx.set_shape(node.outputs[0], output_shape)

Shape data propagation (pkg.onnx_shape_inference.sym_data)

Shape inference alone cannot resolve output shapes when ops like Reshape consume non-constant shape tensors that were computed at runtime (e.g. Shape → Slice → Concat → Reshape). The sym_data feature bridges this gap by tracking the known element values of 1-D integer tensors as they flow through the graph.

Propagation tracks up to 1024 integer elements per tensor, including inside local functions. Larger tensors retain shape and dtype inference without expanding their contents into Python lists and JSON metadata.

After inference, each value that carries propagated data has a pkg.onnx_shape_inference.sym_data entry in its metadata_props. You can read it directly or use the SYM_DATA_KEY constant:

import json
import numpy as np
import onnx_ir as ir
from onnx_shape_inference import SYM_DATA_KEY, infer_symbolic_shapes

model = infer_symbolic_shapes(model)

for node in model.graph:
    for value in node.inputs:
        if SYM_DATA_KEY in value.metadata_props:
            text = value.metadata_props[SYM_DATA_KEY]  # e.g. '["N",3,768]'
            elements = json.loads(text)                # ["N", 3, 768]

            # You can create an ir.Shape from it
            shape = ir.Shape(elements)

            # Then you can replace this input with a constant value

When all elements are concrete integers the value is also stored as a constant tensor, so downstream consumers that read constants directly can access it without parsing metadata_props.

Adopting declared symbolic names (constraint resolution)

Per-operator inference names data-dependent dimensions with anonymous symbols (_d0, _d1, …). Model authors, however, usually declare meaningful symbolic names on graph outputs and value_info (e.g. Y: [batch, seq]). After the main inference pass, a constraint-resolution pass records equalities between inferred and declared shapes and renames the anonymous symbols to the author's declared names — including compound occurrences such as 2*_d0 → 2*batch and past_seq + seq → total_seq. Only anonymous _dN symbols are renamed; declared names are treated as authoritative.

This is enabled by default. Pass adopt_declared_symbols=False to keep the raw engine-generated symbols instead:

model = infer_symbolic_shapes(model, adopt_declared_symbols=False)

Constraint propagation reuses dimension rewrites within a single call, including unchanged results shared by repeated dimensions and nested graphs. This cache is discarded after the call; it does not alter replacement order, generated-symbol identity, anchor adoption, or reshape divisibility requirements. For independent atomic symbol renames on binder-free arithmetic expressions, irrelevant source symbols are omitted before substitution. Other expressions and replacement maps retain the original substitution path; compound replacements are always processed in their original order.

Development

pip install -r requirements/ci/requirements.txt
pip install -e .
pytest

Backend node tests use ONNX's loader to generate models and inputs in memory. They do not require a bundled or downloaded backend test-data directory.

Lint and format with lintrunner:

pip install -r requirements/lintrunner/requirements.txt
lintrunner -a

Python source and test files must begin with ONNX's copyright and SPDX header:

# Copyright (c) ONNX Project Contributors
#
# SPDX-License-Identifier: Apache-2.0

Ruff enforces this header, including in small package and configuration files.

Fuzzing

A deterministic, seeded fuzzer stress-tests shape inference against ONNX reference inference and ONNX Runtime. A fast tier runs as part of the normal test suite; replay a failing seed with:

FUZZ_SEED=<seed> python3 -m pytest tests/shape_inference_fuzz_test.py

See docs/fuzzing.md for how to turn a fuzzer finding into a regression test, and docs/fuzzing-design.md for the fuzzer's design (generator, oracles, harness, and shrinking).

License

Licensed under the Apache License, Version 2.0, using the ONNX license text. See NOTICE for retained copyright and third-party license notices.

Releases

Packages

Used by

Contributors

Languages