onnx_ir.passes

Use built-in passes

Common, reusable passes are implemented in . You can use onnx_ir.passes.Sequential to chain passes or use onnx_ir.passes.PassManager which supports early stopping if no changes are made.

Example

import onnx_ir as ir
import onnx_ir.passes.common as common_passes

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

# You can chain passes with ir.passes.Sequential
passes = ir.passes.Sequential(
    common_passes.DeduplicateHashedInitializersPass(size_limit=1024 * 1024),
    common_passes.CommonSubexpressionEliminationPass(),
)
result = passes(model)

# Or you can run passes individually. Passing in result or result.model has the same effect
result = common_passes.ClearMetadataAndDocStringPass()(result)

print("The model was modified:", result.modified)
ir.save(result.model, "model.onnx")

For more advanced use cases, you can use onnx_ir.passes.PassManager to orchestrate passes with automatic iteration until convergence:

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

passes = ir.passes.PassManager(
    [
        # Pass managers can be nested
        ir.passes.PassManager(
            [
                common_passes.DeduplicateHashedInitializersPass(size_limit=1024 * 1024),
                common_passes.CommonSubexpressionEliminationPass(),
            ],
            steps=2,
            early_stop=True,
        ),
        common_passes.ClearMetadataAndDocStringPass(),
    ],
    steps=2,
    early_stop=False,
)

result = passes(model)

Pass infrastructure

Inherit from onnx_ir.passes.InPlacePass or onnx_ir.passes.FunctionalPass to define a pass. You will need to implement the call method which returns a onnx_ir.passes.PassResult.

Alternatively, inherit from the base class onnx_ir.passes.PassBase and override the two properties changes_input and in_place to set properties of the pass.

onnx_ir.passes.PassBase	Base class for all passes.
onnx_ir.passes.InPlacePass	A pass that modifies the input model in place and returns it.
onnx_ir.passes.FunctionalPass	A pass that returns a new model but does not modify the input model.
onnx_ir.passes.Sequential	Run a sequence of passes in order.
onnx_ir.passes.PassResult	Result of a pass.
onnx_ir.passes.PassManager	Pass manager for the IR.

Errors

exception onnx_ir.passes.InvariantError

Raised when an invariant is violated.

exception onnx_ir.passes.PreconditionError

Raised when a precondition is violated.

exception onnx_ir.passes.PostconditionError

Raised when a postcondition is violated.

exception onnx_ir.passes.PassError

Raised when an error occurs during a pass.

Utilities

onnx_ir.passes.functionalize(pass_instance)

Produce a functional pass from a given pass.

A new functional pass is created that clones the input model before running the pass.

Added in version 0.1.14.

Parameters:

pass_instance (PassBase) – The pass to convert.

Returns:

A functional pass.

Return type:

FunctionalPass