onnx_ir.passes.common

Built-in passes provided by the ONNX IR

class onnx_ir.passes.common.AddInitializersToInputsPass

Bases: InPlacePass

Add initializers to inputs.

This pass finds all initializers and adds them to the graph.inputs list if they are not already present.

class onnx_ir.passes.common.CheckerPass(full_check=False, skip_opset_compatibility_check=False, check_custom_domain=False)

Bases: PassBase

Run onnx checker on the model.

Parameters:

• full_check (bool)

• skip_opset_compatibility_check (bool)

• check_custom_domain (bool)

property changes_input: Literal[False]

This pass does not change the input model.

property in_place: Literal[True]

This pass does not create a new model.

class onnx_ir.passes.common.ClearMetadataAndDocStringPass

Bases: InPlacePass

Clear all metadata and docstring from the model, graphs, nodes, and functions.

class onnx_ir.passes.common.CommonSubexpressionEliminationPass(size_limit=10)

Bases: InPlacePass

Eliminate common subexpression in ONNX graphs.

Parameters:

size_limit (int)

size_limit

The maximum size of the tensor to be csed. If the tensor contains number of elements larger than size_limit, it will not be cse’d. Default is 10.

class onnx_ir.passes.common.DeduplicateInitializersPass(size_limit=1024)

Bases: InPlacePass

Remove duplicated initializer tensors from the graph.

This pass detects initializers with identical shape, dtype, and content, and replaces all duplicate references with a canonical one.

To deduplicate initializers from subgraphs, use LiftSubgraphInitializersToMainGraphPass to lift the initializers to the main graph first before running pass.

Parameters:

size_limit (int)

class onnx_ir.passes.common.InlinePass

Bases: InPlacePass

Inline model local functions to the main graph and clear function definitions.

class onnx_ir.passes.common.LiftConstantsToInitializersPass(lift_all_constants=False, size_limit=16)

Bases: InPlacePass

Lift constants to initializers.

Parameters:

• lift_all_constants (bool)

• size_limit (int)

lift_all_constants

Whether to lift all Constant nodes, including those that does not contain a tensor attribute (e.g. with value_ints etc.) Default to False, where only Constants with the value attribute are lifted.

size_limit

The minimum size of the tensor to be lifted. If the tensor contains number of elements less than size_limit, it will not be lifted. Default is 16.

class onnx_ir.passes.common.LiftSubgraphInitializersToMainGraphPass

Bases: InPlacePass

Lift subgraph initializers to main graph.

This pass lifts the initializers of a subgraph to the main graph. It is used to ensure that the initializers are available in the main graph for further processing or optimization.

Initializers that are also graph inputs will not be lifted.

class onnx_ir.passes.common.RemoveInitializersFromInputsPass

Bases: InPlacePass

Remove initializers from inputs.

This pass finds all graph inputs that have a const_value and removes them from the graph.inputs list.

class onnx_ir.passes.common.RemoveUnusedFunctionsPass

Bases: InPlacePass

class onnx_ir.passes.common.RemoveUnusedNodesPass

Bases: InPlacePass

Pass for removing unused nodes and initializers (dead code elimination).

This pass does not modify the model signature (inputs and outputs). It ensures that unused nodes and initializers are removed while preserving the original contract of the model.

class onnx_ir.passes.common.RemoveUnusedOpsetsPass(process_functions=True)

Bases: InPlacePass

Remove unused opset imports from the model and functions.

Parameters:

process_functions (bool)

process_functions

Whether to process functions in the model. If True, the pass will remove unused opset imports from functions as well. If False, only the main graph will be processed.

class onnx_ir.passes.common.ShapeInferencePass(check_type=True, strict_mode=True, data_prop=True)

Bases: InPlacePass

This pass performs shape inference on the graph.

Parameters:

• check_type (bool)

• strict_mode (bool)

• data_prop (bool)

class onnx_ir.passes.common.TopologicalSortPass

Bases: InPlacePass

Topologically sort graphs and functions in a model.