Optimizing a model with graph-rewriting patterns#

onnx-light optimizes a model by repeatedly matching small PatternOptimization subgraphs and replacing them with a simplified equivalent. This example walks through the whole workflow on a tiny model:

Build a model containing redundant nodes (two useless Cast and two consecutive Neg).
Run the standard patterns and inspect the optimization statistics (OptimizationReport).
List every applied modification (LocalRewriting) and replay them from the original model to reproduce the optimized graph.
Write a custom pattern.
Inspect a candidate rejected by that pattern and the recorded failure reason.
Apply the custom pattern together with the standard ones.

See How to add a custom graph-rewriting pattern and set its priority for a companion how-to that focuses on writing and registering a pattern (including how priorities order patterns), in both Python and C++.

from __future__ import annotations

import onnx_light.onnx.helper as oh
from onnx_light.onnx_lib import parser
from onnx_light.onnx_core.optimization import (
    GraphBuilder,
    GraphGraph,
    PatternOptimization,
    replay,
    standard_patterns,
)
from onnx_light.tools import pretty_onnx

Build a model with redundant nodes#

x is cast to float (a no-op, since it is already float) and then negated twice before being cast back. The standard Cast pattern removes the two Cast nodes; the custom pattern added further below removes the two Neg nodes.

model = parser.parse_model(
    '<ir_version: 10, opset_import: ["" : 18]>\n'
    "agraph (float[4] x) => (float[4] y) {\n"
    "  casted = Cast <to=1> (x)\n"
    "  middle = Neg(casted)\n"
    "  negated = Neg(middle)\n"
    "  y = Cast <to=1> (negated)\n"
    "}\n"
)
print(pretty_onnx(model))

opset: domain='' version=18
graph: name='agraph'
input: float[4] x
0: Cast(x) -> casted
1: Neg(casted) -> middle
2: Neg(middle) -> negated
3: Cast(negated) -> y
output: float[4] y

Run the standard patterns and read the statistics#

optimize() returns the list of applied rewrites; passing report=True additionally returns an OptimizationReport with timing and match/no-match counters for every pattern that was tried.

builder = GraphBuilder(model)
graph = GraphGraph(builder, standard_patterns(["Cast"]))
rewrites, report = graph.optimize(report=True)
optimized_graph = builder.build_graph()

print(pretty_onnx(builder.to_onnx("model")))
print(report)

for pattern_stats in report.patterns:
    print(
        f"{pattern_stats.pattern_name}: {pattern_stats.matches} match(es) over "
        f"{pattern_stats.attempts} attempt(s)"
    )

opset: domain='ai.onnx' version=18
graph: name='agraph'
input: float[4] x
0: Neg(x) -> middle
1: Neg(middle) -> negated
2: Identity(negated) -> y
output: float[4] y
OptimizationReport(iterations=5, rewrites=3, total_time_ns=125049, phases={matching: 34953, rewriting: 22573, cleanup: 39040, constant_folding: 28483, subgraph_optimization: 0}, patterns=[Cast(attempts=2, matches=2, match_time_ns=2493, apply_time_ns=5168), ConvBiasNull(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), Identity(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), PadConv(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ConcatEmpty(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ConcatGather(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ConcatTwiceUnary(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), GatherConcat(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), GatherGather(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), GatherShape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), GathersSplit(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SliceSlice(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SlicesSplit(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SplitConcat(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SequenceConstructAt(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SplitToSequenceSequenceAt(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), NotWhere(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), UnsqueezeEqual(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), WhereAdd(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), Expand(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ExpandBroadcast(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedConcatExpand(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedExpandBroadcast(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedExpandBroadcastMatMul(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedStaticExpand(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedExpandSwap(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedExpandCastWhereSwap(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ExpandSwap(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SwapExpandUnsqueeze(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ExpandUnsqueezeExpand(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ConcatReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), Reshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ReduceReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), Reshape2Of3(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ReshapeReshapeBinary(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ReshapeReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ReshapeSqueeze(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedEditDistanceReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedReshapeIsSqueeze(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapedBasedReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), StaticConcatReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), UnsqueezeOrSqueezeReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), UnsqueezeReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), MulUnsqueezeUnsqueeze(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SqueezeAdd(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SqueezeBinaryUnsqueeze(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SwapUnsqueezeTranspose(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), TransposeEqualReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), TransposeReshapeTranspose(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), MulMulMulScalar(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SwitchOrderBinary(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SwapRangeAddScalar(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ReduceArgTopK(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ReduceSumNormalize(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), Sub1Mul(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SwapUnary(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SameChildren(attempts=16, matches=0, match_time_ns=3745, apply_time_ns=0, no_matches=[/home/runner/work/xadupre.github.io/xadupre.github.io/onnx-light/onnx_light/onnx_extensions/patterns/algebra/common_pattern.cc:339(occurrences=16, reason=no equivalent sibling nodes found)]), SameChildrenFromInput(attempts=16, matches=0, match_time_ns=1993, apply_time_ns=0, no_matches=[/home/runner/work/xadupre.github.io/xadupre.github.io/onnx-light/onnx_light/onnx_extensions/patterns/algebra/common_pattern.cc:387(occurrences=5, reason=the graph input has fewer than two consumers), /home/runner/work/xadupre.github.io/xadupre.github.io/onnx-light/onnx_light/onnx_extensions/patterns/algebra/common_pattern.cc:383(occurrences=11, reason=the candidate first input is not a graph input)]), ShapeBasedIdentity(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedSameChildren(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedShapeShapeAdd(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), BatchNormalization(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), BatchNormalizationTraining(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), Gelu(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), LeakyRelu(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SoftmaxCrossEntropyLossCast(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), FunctionAttention(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SwapExpandReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), TransposeTranspose(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), TransposeGather(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), UnsqueezeUnsqueeze(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SqueezeUnsqueeze(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeTranspose(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), UnsqueezeShape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), CastCast(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), CastCastBinary(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), CastOpCast(attempts=6, matches=0, match_time_ns=3053, apply_time_ns=0, no_matches=[/home/runner/work/xadupre.github.io/xadupre.github.io/onnx-light/onnx_light/onnx_extensions/patterns/canonicalization/cast_pattern.cc:321(occurrences=6, reason=the operation output must feed one unary Cast node)]), ClipClip(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ConstantToInitializer(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), Dropout(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), NotNot(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), GemmTranspose(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), MatMulReshape2Of3(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), MulMulMatMul(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ReshapeMatMulReshape(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), ShapeBasedMatMulToMul(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), SwitchReshapeActivation(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), TransposeMatMul(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), TransposeReshapeMatMul(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), CastLayerNormalizationCast(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), LayerNormalization(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), LayerNormalizationScale(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), RMSNormalization(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), RMSNormalizationMul(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), MaxRelu(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), RotaryEmbedding(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), RotaryConcatPart(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), FunctionCausalMask(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), FunctionCausalMaskMulAdd(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), FunctionCosSinCache(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), FunctionHalfRotaryEmbedding(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0), FunctionAttentionGQA(attempts=9, matches=0, match_time_ns=2355, apply_time_ns=0, no_matches=[/home/runner/work/xadupre.github.io/xadupre.github.io/onnx-light/onnx_light/onnx_extensions/patterns/attention/attention_pattern.cc:773(occurrences=9, reason=FunctionAttentionGQAPattern expects a LocalAttention call.)]), AttentionGQA(attempts=6, matches=0, match_time_ns=1153, apply_time_ns=0, no_matches=[/home/runner/work/xadupre.github.io/xadupre.github.io/onnx-light/onnx_light/onnx_extensions/patterns/attention/attention_pattern.cc:871(occurrences=6, reason=AttentionGQAPattern requires opset 23 and one output.)]), MatMulAdd(attempts=0, matches=0, match_time_ns=0, apply_time_ns=0)], subgraphs=[])
Cast: 2 match(es) over 2 attempt(s)
ConvBiasNull: 0 match(es) over 0 attempt(s)
Identity: 0 match(es) over 0 attempt(s)
PadConv: 0 match(es) over 0 attempt(s)
ConcatEmpty: 0 match(es) over 0 attempt(s)
ConcatGather: 0 match(es) over 0 attempt(s)
ConcatTwiceUnary: 0 match(es) over 0 attempt(s)
GatherConcat: 0 match(es) over 0 attempt(s)
GatherGather: 0 match(es) over 0 attempt(s)
GatherShape: 0 match(es) over 0 attempt(s)
GathersSplit: 0 match(es) over 0 attempt(s)
SliceSlice: 0 match(es) over 0 attempt(s)
SlicesSplit: 0 match(es) over 0 attempt(s)
SplitConcat: 0 match(es) over 0 attempt(s)
SequenceConstructAt: 0 match(es) over 0 attempt(s)
SplitToSequenceSequenceAt: 0 match(es) over 0 attempt(s)
NotWhere: 0 match(es) over 0 attempt(s)
UnsqueezeEqual: 0 match(es) over 0 attempt(s)
WhereAdd: 0 match(es) over 0 attempt(s)
Expand: 0 match(es) over 0 attempt(s)
ExpandBroadcast: 0 match(es) over 0 attempt(s)
ShapeBasedConcatExpand: 0 match(es) over 0 attempt(s)
ShapeBasedExpandBroadcast: 0 match(es) over 0 attempt(s)
ShapeBasedExpandBroadcastMatMul: 0 match(es) over 0 attempt(s)
ShapeBasedStaticExpand: 0 match(es) over 0 attempt(s)
ShapeBasedExpandSwap: 0 match(es) over 0 attempt(s)
ShapeBasedExpandCastWhereSwap: 0 match(es) over 0 attempt(s)
ExpandSwap: 0 match(es) over 0 attempt(s)
SwapExpandUnsqueeze: 0 match(es) over 0 attempt(s)
ExpandUnsqueezeExpand: 0 match(es) over 0 attempt(s)
ConcatReshape: 0 match(es) over 0 attempt(s)
Reshape: 0 match(es) over 0 attempt(s)
ReduceReshape: 0 match(es) over 0 attempt(s)
Reshape2Of3: 0 match(es) over 0 attempt(s)
ReshapeReshapeBinary: 0 match(es) over 0 attempt(s)
ReshapeReshape: 0 match(es) over 0 attempt(s)
ReshapeSqueeze: 0 match(es) over 0 attempt(s)
ShapeBasedEditDistanceReshape: 0 match(es) over 0 attempt(s)
ShapeBasedReshapeIsSqueeze: 0 match(es) over 0 attempt(s)
ShapedBasedReshape: 0 match(es) over 0 attempt(s)
StaticConcatReshape: 0 match(es) over 0 attempt(s)
UnsqueezeOrSqueezeReshape: 0 match(es) over 0 attempt(s)
UnsqueezeReshape: 0 match(es) over 0 attempt(s)
MulUnsqueezeUnsqueeze: 0 match(es) over 0 attempt(s)
SqueezeAdd: 0 match(es) over 0 attempt(s)
SqueezeBinaryUnsqueeze: 0 match(es) over 0 attempt(s)
SwapUnsqueezeTranspose: 0 match(es) over 0 attempt(s)
TransposeEqualReshape: 0 match(es) over 0 attempt(s)
TransposeReshapeTranspose: 0 match(es) over 0 attempt(s)
MulMulMulScalar: 0 match(es) over 0 attempt(s)
SwitchOrderBinary: 0 match(es) over 0 attempt(s)
SwapRangeAddScalar: 0 match(es) over 0 attempt(s)
ReduceArgTopK: 0 match(es) over 0 attempt(s)
ReduceSumNormalize: 0 match(es) over 0 attempt(s)
Sub1Mul: 0 match(es) over 0 attempt(s)
SwapUnary: 0 match(es) over 0 attempt(s)
SameChildren: 0 match(es) over 16 attempt(s)
SameChildrenFromInput: 0 match(es) over 16 attempt(s)
ShapeBasedIdentity: 0 match(es) over 0 attempt(s)
ShapeBasedSameChildren: 0 match(es) over 0 attempt(s)
ShapeBasedShapeShapeAdd: 0 match(es) over 0 attempt(s)
BatchNormalization: 0 match(es) over 0 attempt(s)
BatchNormalizationTraining: 0 match(es) over 0 attempt(s)
Gelu: 0 match(es) over 0 attempt(s)
LeakyRelu: 0 match(es) over 0 attempt(s)
SoftmaxCrossEntropyLossCast: 0 match(es) over 0 attempt(s)
FunctionAttention: 0 match(es) over 0 attempt(s)
SwapExpandReshape: 0 match(es) over 0 attempt(s)
TransposeTranspose: 0 match(es) over 0 attempt(s)
TransposeGather: 0 match(es) over 0 attempt(s)
UnsqueezeUnsqueeze: 0 match(es) over 0 attempt(s)
SqueezeUnsqueeze: 0 match(es) over 0 attempt(s)
ShapeTranspose: 0 match(es) over 0 attempt(s)
UnsqueezeShape: 0 match(es) over 0 attempt(s)
CastCast: 0 match(es) over 0 attempt(s)
CastCastBinary: 0 match(es) over 0 attempt(s)
CastOpCast: 0 match(es) over 6 attempt(s)
ClipClip: 0 match(es) over 0 attempt(s)
ConstantToInitializer: 0 match(es) over 0 attempt(s)
Dropout: 0 match(es) over 0 attempt(s)
NotNot: 0 match(es) over 0 attempt(s)
GemmTranspose: 0 match(es) over 0 attempt(s)
MatMulReshape2Of3: 0 match(es) over 0 attempt(s)
MulMulMatMul: 0 match(es) over 0 attempt(s)
ReshapeMatMulReshape: 0 match(es) over 0 attempt(s)
ShapeBasedMatMulToMul: 0 match(es) over 0 attempt(s)
SwitchReshapeActivation: 0 match(es) over 0 attempt(s)
TransposeMatMul: 0 match(es) over 0 attempt(s)
TransposeReshapeMatMul: 0 match(es) over 0 attempt(s)
CastLayerNormalizationCast: 0 match(es) over 0 attempt(s)
LayerNormalization: 0 match(es) over 0 attempt(s)
LayerNormalizationScale: 0 match(es) over 0 attempt(s)
RMSNormalization: 0 match(es) over 0 attempt(s)
RMSNormalizationMul: 0 match(es) over 0 attempt(s)
MaxRelu: 0 match(es) over 0 attempt(s)
RotaryEmbedding: 0 match(es) over 0 attempt(s)
RotaryConcatPart: 0 match(es) over 0 attempt(s)
FunctionCausalMask: 0 match(es) over 0 attempt(s)
FunctionCausalMaskMulAdd: 0 match(es) over 0 attempt(s)
FunctionCosSinCache: 0 match(es) over 0 attempt(s)
FunctionHalfRotaryEmbedding: 0 match(es) over 0 attempt(s)
FunctionAttentionGQA: 0 match(es) over 9 attempt(s)
AttentionGQA: 0 match(es) over 6 attempt(s)
MatMulAdd: 0 match(es) over 0 attempt(s)

List the modifications and replay them#

Each LocalRewriting records which pattern fired, the positions of the matched nodes, and the nodes it added. replay() reconstructs the optimized graph by reapplying that captured sequence to a fresh copy of the original model, without running the pattern matcher again.

for rewrite in rewrites:
    print(rewrite)

replayed_graph = replay(model, rewrites)
assert replayed_graph.SerializeToString() == optimized_graph.SerializeToString()
print("replay reproduced the optimized graph")

LocalRewriting(pattern=Cast, graph_path=[], matched_nodes=[0], added_nodes=[Identity(outputs=[casted])], added_nodes_positions=[0], added_initializers=[], added_initializer_positions=[], removed_initializers=[], value_renames=[], iteration=0, match_time_ns=2313, apply_time_ns=4667)
LocalRewriting(pattern=Cast, graph_path=[], matched_nodes=[3], added_nodes=[Identity(outputs=[y])], added_nodes_positions=[3], added_initializers=[], added_initializer_positions=[], removed_initializers=[], value_renames=[], iteration=0, match_time_ns=180, apply_time_ns=501)
LocalRewriting(pattern=RemoveIdentityNodes, graph_path=[], matched_nodes=[0, 1, 2, 3], added_nodes=[Neg(outputs=[middle]), Neg(outputs=[negated]), Identity(outputs=[y])], added_nodes_positions=[0, 1, 2], added_initializers=[], added_initializer_positions=[], removed_initializers=[], value_renames=[casted->x], iteration=1, match_time_ns=0, apply_time_ns=0)
replay reproduced the optimized graph

Add a custom pattern#

A pattern derives from PatternOptimization and implements fast_op_type (the operator types it may start from), match (returns self.result(...) on success or self.no_match(...) with a diagnostic otherwise) and apply (builds the replacement nodes). Passing it to GraphGraph alongside the standard patterns runs both together, in ascending priority order.

class NegNegPattern(PatternOptimization):
    """Replaces two consecutive Neg nodes with Identity."""

    def __init__(self):
        super().__init__(priority=1, name="NegNeg")

    def fast_op_type(self):
        return {"Neg"}

    def match(self, graph, node):
        previous = graph.node_before(node.input[0])
        if previous is None or previous.op_type != "Neg":
            return self.no_match(node, "the input is not produced by Neg")
        return self.result([previous, node], insert_at=node)

    def apply(self, graph, nodes):
        del graph
        previous, node = nodes
        return [oh.make_node("Identity", [previous.input[0]], list(node.output))]

Inspect a failed pattern candidate#

The same custom pattern rejects a single Neg because its input is not produced by another Neg. Returning self.no_match(candidate, reason) keeps that rejection out of normal output, but the optional report stores the aggregated reason so the pattern author can understand why nothing changed.

failure_model = parser.parse_model(
    '<ir_version: 10, opset_import: ["" : 18]>\n'
    "agraph (float[4] x) => (float[4] y) {\n"
    "  y = Neg(x)\n"
    "}\n"
)
builder = GraphBuilder(failure_model)
graph = GraphGraph(builder, [NegNegPattern()])
failed_rewrites, failed_report = graph.optimize(report=True)

print(f"failed rewrite count: {len(failed_rewrites)}")
for pattern_stats in failed_report.patterns:
    for no_match in pattern_stats.no_matches:
        print(
            f"{pattern_stats.pattern_name} rejected "
            f"{no_match.occurrences} candidate(s): {no_match.reason}"
        )

assert not failed_rewrites
failed_no_match_reasons = {
    no_match.reason
    for pattern_stats in failed_report.patterns
    if pattern_stats.pattern_name == "NegNeg"
    for no_match in pattern_stats.no_matches
}
assert "the input is not produced by Neg" in failed_no_match_reasons

failed rewrite count: 0
SameChildren rejected 4 candidate(s): no equivalent sibling nodes found
SameChildrenFromInput rejected 4 candidate(s): the graph input has fewer than two consumers
CastOpCast rejected 3 candidate(s): the operation output must feed one unary Cast node
FunctionAttentionGQA rejected 3 candidate(s): FunctionAttentionGQAPattern expects a LocalAttention call.
NegNeg rejected 3 candidate(s): the input is not produced by Neg
AttentionGQA rejected 2 candidate(s): AttentionGQAPattern requires opset 23 and one output.

Apply the custom pattern with the standard patterns#

builder = GraphBuilder(model)
graph = GraphGraph(builder, [*standard_patterns(["Cast"]), NegNegPattern()])
rewrites = graph.optimize()
fully_optimized = builder.to_onnx("model")

print(pretty_onnx(fully_optimized))
print([rewrite.pattern_name for rewrite in rewrites])
assert [node.op_type for node in fully_optimized.graph.node] == ["Identity"]

opset: domain='ai.onnx' version=18
graph: name='agraph'
input: float[4] x
0: Identity(x) -> y
output: float[4] y
['Cast', 'Cast', 'RemoveIdentityNodes', 'NegNeg', 'RemoveIdentityNodes']

Total running time of the script: (0 minutes 0.006 seconds)