onnx_light.onnx_core.optimization

Optimization workflow

Optimization always operates on a GraphBuilder through GraphGraph:

from onnx_light.onnx_core.optimization import GraphBuilder, GraphGraph

builder = GraphBuilder(model)
graph = GraphGraph(builder)
rewrites, report = graph.optimize(report=True)
optimized_model = builder.to_onnx("model")

Pattern registration

The available patterns are merged by their stable PatternOptimization.name. Builder registrations replace global entries with the same name before selection:

Scope	Registration	Selection
Global	register_pattern()	Makes a pattern available for selection. The standard ONNX patterns are registered globally when this module is imported.
Builder	GraphBuilder.register_pattern()	Overrides a global pattern for optimizers built over that builder.
Graph	GraphGraph(builder, patterns=[...])	Selects only the given names and instances. Explicit instances override earlier entries with the same name and are retained for that optimizer, including recursive subgraphs.

clear_registered_patterns() clears the global registry; reset_registered_patterns() restores the standard patterns.

Pattern selection

GraphGraph(builder, patterns=...) accepts one selector:

Selector	Selected patterns
None (default)	All available device-independent patterns, even if builder.device is already set.
False or []	None, including builder registrations. The optimizer’s ordinary cleanup and constant-folding passes still run.
A concrete Device	All device-independent patterns plus those targeting that exact device. Also sets builder.device and subgraph targets. A different, already-defined device in any of those builders raises ValueError before changing any target.
A regex string or compiled re.Pattern[str]	Available names accepted by fullmatch, including device-specific patterns. Use ".*MatMul.*" to select names containing MatMul.
An iterable of names and/or pattern instances	Only those patterns, including device-specific ones. Names are exact, not regexes. Standard names may also be instantiated when absent from the global registry.

Regexes and explicit lists do not modify builder.device or automatically add global or builder patterns. A regex matching nothing selects no patterns; invalid regexes and invalid selector types raise exceptions. True is not a selector, and Device.kUndefined is not a concrete device: use None for the default selection. During recursive optimization, a subgraph with an undefined device inherits its parent builder’s device; explicitly defined subgraph devices are preserved.

import re
from onnx_light.onnx_core.shape_inference import Device

graph = GraphGraph(builder)                         # device-independent defaults
graph = GraphGraph(builder, patterns=False)         # cleanup only
graph = GraphGraph(builder, patterns=Device.kCPU)    # defaults + CPU patterns
graph = GraphGraph(builder, patterns=r".*MatMul.*")  # full name regex
graph = GraphGraph(builder, patterns=re.compile(r"Cast.*"))
graph = GraphGraph(builder, patterns=["Cast", "TransposeMatMul"])

A pattern’s device defaults to Device.kUndefined, meaning device-independent. Custom Python patterns declare a target with super().__init__(name="MyCPUFusion", device=Device.kCPU); native patterns use the corresponding third PatternOptimization constructor argument. Device equality is exact, not a GPU-family or execution-provider capability query. Existing standard patterns remain device-independent, including patterns whose matching heuristics inspect builder.device.

The former use_global_patterns argument is removed. Replace an explicit list plus use_global_patterns=False with just that list, use False for no patterns, or pass builder.registered_patterns() for builder-only patterns. To combine defaults with explicit patterns, supply the combined list explicitly, for example [*standard_patterns(), custom_pattern] (which intentionally includes every standard pattern, regardless of device).

Registered standard patterns

The following table lists the standard patterns registered when this module is imported. It is generated from the live registry, so it always reflects the currently available patterns.

Class / registered name	Priority	Candidate roots	Transformation
AttentionGQAPattern / AttentionGQA	2		Fuses grouped-query attention cache handling.
BatchNormalizationPattern / BatchNormalization	0	BatchNormalization	Fuses an inference batch-normalization subgraph.
BatchNormalizationTrainingPattern / BatchNormalizationTraining	0	BatchNormalization	Fuses a training batch-normalization subgraph.
CastPattern / Cast	0	Cast	Replaces a type-preserving Cast(to=T) with Identity.
CastCastPattern / CastCast	1	Cast	Collapses two consecutive compatible Cast nodes.
CastCastBinaryPattern / CastCastBinary	1	Add, Div, Mul, Sub	Moves matching floating-point input Cast nodes after a binary operation.
CastLayerNormalizationCastPattern / CastLayerNormalizationCast	1	GroupNormalization, LayerNormalization, RMSNormalization, SimplifiedLayerNormalization	Removes redundant casts surrounding LayerNormalization.
CastOpCastPattern / CastOpCast	1	Add, Div, Mul, MulSigmoid, Neg, Sigmoid, Softmax, Sub	Moves a unary or binary operation to the result Cast type.
ClipClipPattern / ClipClip	1	Clip	Merges two consecutive Clip nodes with complementary bounds.
ConcatEmptyPattern / ConcatEmpty	0	Concat	Drops empty inputs from a Concat node, reducing it to an Identity when a single input remains.
ConcatGatherPattern / ConcatGather	0	Gather	Rewrites a Gather reading a single Concat input into a Gather on that input directly.
ConcatReshapePattern / ConcatReshape	0	Reshape	Simplifies concatenations that construct reshape shapes.
ConcatSliceEliminationPattern / ConcatSliceElimination	0	Concat	Eliminates a Concat followed by exact slices recovering all inputs.
ConcatTwiceUnaryPattern / ConcatTwiceUnary	0	Concat	Pushes a shape-preserving unary op ahead of a Concat(x, x) so the unary op runs once on x.
ConstantToInitializerPattern / ConstantToInitializer	1	Constant	Replaces a Constant node by an initializer and an Identity node.
ConvAddFusionPattern / ConvAddFusion	0	Conv	Folds a channel-wise constant Add following a Conv into the Conv bias.
ConvBatchNormalizationFusionPattern / ConvBatchNormalizationFusion	0	Conv	Folds inference BatchNormalization following a Conv into its weights and bias.
ConvBiasNullPattern / ConvBiasNull	0	Conv	Removes a null (all-zero) bias input from a Conv node.
ConvMulFusionPattern / ConvMulFusion	0	Conv	Folds a scalar or channel-wise constant Mul following a Conv into its constants.
DivMulPattern / DivMul	0	Div	Fuses multiplication by a reciprocal into one division.
DropoutPattern / Dropout	1	Dropout	Replaces an inference Dropout by an Identity node when its mask output is unused and training mode is disabled.
ExpandPattern / Expand	0	Expand	Replaces Expand(x, shape) with Identity(x) when the target shape equals the input shape.
ExpandBroadcastPattern / ExpandBroadcast	0	Expand	Drops an Expand feeding an element-wise binary operator that already broadcasts the pre-expanded input.
ExpandSwapPattern / ExpandSwap	0	Expand	Moves an Expand past a following unary-like operator so the operator runs on the smaller tensor.
ExpandUnsqueezeExpandPattern / ExpandUnsqueezeExpand	0	Expand	Fuses Expand, Unsqueeze and Expand into a single Unsqueeze followed by one Expand.
FunctionAttentionPattern / FunctionAttention	0	Softmax	Replaces a scaled dot-product attention subgraph.
FunctionAttentionGQAPattern / FunctionAttentionGQA	1		Replaces grouped-query attention expressed with local functions.
FunctionCausalMaskPattern / FunctionCausalMask	1	Greater, LessOrEqual	Replaces a causal-mask subgraph with a local function.
FunctionCausalMaskMulAddPattern / FunctionCausalMaskMulAdd	1	Add	Fuses scaling and offset operations into a causal-mask function.
FunctionCosSinCachePattern / FunctionCosSinCache	1	Cos	Replaces cosine and sine cache construction with a local function.
FunctionHalfRotaryEmbeddingPattern / FunctionHalfRotaryEmbedding	1	Split	Replaces half-rotary embedding construction with a local function.
GatherConcatPattern / GatherConcat	0	Gather	Merges a Concat of single-index Gather nodes on a shared input into one Gather node.
GatherGatherPattern / GatherGather	0	Gather	Collapses two consecutive scalar Gather nodes into a single Gather node.
GatherShapePattern / GatherShape	0	Gather	Rewrites a Gather of a scalar index over a Shape node into a narrowed Shape node.
GatherSliceToSplitPattern / GatherSliceToSplit	0	Gather, Slice	Fuses compatible sibling Gather and Slice ranges into one Split.
GatherToSlicePattern / GatherToSlice	0	Gather	Rewrites a Gather selecting a constant scalar, singleton, or arithmetic-progression index into a Slice (plus Squeeze for a scalar index).
GatherUpstreamPropagationPattern / GatherUpstreamPropagation	0	Gather	Moves a scalar or vector constant-index Gather upstream across one compatible producer (pointwise ops, Transpose, Reshape, MatMul batch dims, Softmax, LayerNormalization) so it runs on smaller tensors.
GathersSplitPattern / GathersSplit	0	Gather	Replaces sibling Gather nodes selecting contiguous single indices of a shared input by a single Split node.
GeluPattern / Gelu	0	Mul	Fuses a GELU activation subgraph.
GemmSumFusionPattern / GemmSumFusion	4	Sum	Fuses a two-input Sum into an unbiased Gemm bias input.
GemmTransposePattern / GemmTranspose	1	Gemm	Folds input transposes into a Gemm operation.
IdentityPattern / Identity	0	Add, And, BatchNormalization, Div, Expand, Mul, Or, Reshape, Slice, Sub, Transpose	Replaces no-op arithmetic and layout operations by an Identity node.
InitializerUnsqueezeCastPattern / InitializerUnsqueezeCast	1	Cast	Folds an initializer’s Unsqueeze and Cast into the initializer consumed by Add.
LabelEncoderFusionPattern / LabelEncoderFusion	0	LabelEncoder	Composes consecutive ai.onnx.ml LabelEncoder mappings.
LayerNormalizationPattern / LayerNormalization	1	ReduceMean	Fuses a layer-normalization subgraph.
LayerNormalizationScalePattern / LayerNormalizationScale	1	LayerNormalization	Fuses layer normalization with its scale.
LeakyReluPattern / LeakyRelu	0	Where	Fuses a LeakyRelu activation subgraph.
LinearAttentionPattern / LinearAttention	2	Mul	Fuses a single-token linear-attention recurrence.
MatMulAddPattern / MatMulAdd	3	Gemm, MatMul	Replaces a compatible MatMul and Add with Gemm.
MatMulBatchNormalizationFusionPattern / MatMulBatchNormalizationFusion	5	BatchNormalization	Folds constant inference BatchNormalization parameters into a rank-two MatMul.
MatMulReshape2Of3Pattern / MatMulReshape2Of3	1	FusedMatMul, MatMul	Simplifies compatible reshapes around MatMul.
MatMulScaleFusionPattern / MatMulScaleFusion	4	MatMul	Absorbs one safe scalar Mul or Div adjacent to a rank-two MatMul.
MaxReluPattern / MaxRelu	1	Max	Replaces a compatible maximum with Relu.
MulMulMatMulPattern / MulMulMatMul	1	MatMul	Moves compatible scalar multiplications across MatMul.
MulMulMulScalarPattern / MulMulMulScalar	0	Div, Mul	Combines scalar factors in multiplication chains.
MulUnsqueezeUnsqueezePattern / MulUnsqueezeUnsqueeze	0	Mul	Simplifies multiplication of unsqueezed inputs.
NotNotPattern / NotNot	1	Not	Fuses two consecutive Not nodes into an Identity node.
NotWherePattern / NotWhere	0	Where	Rewrites Where(Not(c), x, y) into Where(c, y, x).
PadConvPattern / PadConv	0	Conv	Folds a Pad node into the pads attribute of a following Conv node.
PadPadFusionPattern / PadPadFusion	1	Pad	Merges adjacent constant-mode Pad nodes with equal values by summing their pads.
PreShapeNodeEliminationPattern / PreShapeNodeElimination	0	Cast	Removes a Cast whose output only feeds Shape nodes, redirecting each Shape node directly to the Cast’s input.
RMSNormalizationPattern / RMSNormalization	1	ReduceMean	Fuses an RMS-normalization subgraph.
RMSNormalizationMulPattern / RMSNormalizationMul	1	RMSNormalization	Fuses RMS normalization with a following scale.
ReduceArgTopKPattern / ReduceArgTopK	0	ArgMax, ArgMin	Simplifies compatible reduction, arg, and TopK operations.
ReduceReshapePattern / ReduceReshape	0	ReduceL1, ReduceL2, ReduceLogSum, ReduceLogSumExp, ReduceMax, ReduceMean, ReduceMin, ReduceProd, ReduceSum, ReduceSumSquare	Simplifies reshape operations around reductions.
ReduceSumNormalizePattern / ReduceSumNormalize	0	ReduceSum	Simplifies reduce-sum normalization subgraphs.
ReluClipFusionPattern / ReluClipFusion	1	Clip	Removes Relu before Clip when the Clip minimum is constant and non-negative.
ReshapePattern / Reshape	0	Reshape	Removes or simplifies redundant reshape operations.
Reshape2Of3Pattern / Reshape2Of3	0	Add, And, BitwiseAnd, BitwiseOr, BitwiseXor, Div, Max, Mean, Min, Mod, Mul, Or, Sub, Sum, Xor	Simplifies two compatible reshapes among three branches.
ReshapeMatMulReshapePattern / ReshapeMatMulReshape	1	MatMul	Simplifies reshape, MatMul, and reshape sequences.
ReshapeReshapePattern / ReshapeReshape	0	Reshape	Collapses consecutive compatible reshapes.
ReshapeReshapeBinaryPattern / ReshapeReshapeBinary	0	Add, And, BitwiseAnd, BitwiseOr, BitwiseXor, Div, Max, Mean, Min, Mod, Mul, Or, Sub, Sum, Xor	Moves compatible reshapes across binary operations.
ReshapeSqueezePattern / ReshapeSqueeze	0	Squeeze	Simplifies a reshape followed by squeeze.
RotaryConcatPartPattern / RotaryConcatPart	1	Add	Simplifies padded rotary concatenation subgraphs.
RotaryEmbeddingPattern / RotaryEmbedding	1	HalfRotaryEmbedding	Fuses a complete rotary-embedding subgraph.
STFTFusionPattern / STFTFusion	0	Transpose	Fuses a canonical convolution-based DFT subgraph into one standard ONNX STFT.
SameChildrenPattern / SameChildren	0		Eliminates equivalent child computations.
SameChildrenFromInputPattern / SameChildrenFromInput	0		Eliminates equivalent computations from one input.
SequenceConstructAtPattern / SequenceConstructAt	0	SequenceConstruct	Replaces a SequenceAt reading a constant index of a SequenceConstruct by the corresponding input tensor.
ShapeBasedConcatExpandPattern / ShapeBasedConcatExpand	0	Expand	Simplifies a dynamic Concat target when Expand changes one dimension.
ShapeBasedEditDistanceReshapePattern / ShapeBasedEditDistanceReshape	0	Reshape	Rewrites reshapes according to the distance between known shapes.
ShapeBasedExpandBroadcastPattern / ShapeBasedExpandBroadcast	0	Add, And, Div, Equal, Greater, GreaterOrEqual, Less, LessOrEqual, Mod, Mul, Or, Sub, Xor	Removes dynamic Expand nodes before a broadcasting binary operator.
ShapeBasedExpandBroadcastMatMulPattern / ShapeBasedExpandBroadcastMatMul	0	MatMul	Removes dynamic Expand nodes from the batch dimensions of MatMul.
ShapeBasedExpandCastWhereSwapPattern / ShapeBasedExpandCastWhereSwap	0	Where	Moves an Expand after a compatible Cast and Where chain.
ShapeBasedExpandSwapPattern / ShapeBasedExpandSwap	0	Add, And, Div, Equal, Greater, GreaterOrEqual, Less, LessOrEqual, Mod, Mul, Or, Sub, Xor	Moves input Expand nodes after a broadcasting binary operator.
ShapeBasedIdentityPattern / ShapeBasedIdentity	0	Slice, Transpose	Eliminates shape-proven identity operations.
ShapeBasedMatMulToMulPattern / ShapeBasedMatMulToMul	1	MatMul	Replaces shape-proven scalar MatMul with Mul.
ShapeBasedReshapeIsSqueezePattern / ShapeBasedReshapeIsSqueeze	0	Expand, Reshape	Replaces eligible reshapes with squeeze.
ShapeBasedSameChildrenPattern / ShapeBasedSameChildren	0	Expand, Reshape	Eliminates shape-equivalent child computations.
ShapeBasedShapeShapeAddPattern / ShapeBasedShapeShapeAdd	0	Add	Exposes the upstream placeholder for additions of two Shape outputs.
ShapeBasedStaticExpandPattern / ShapeBasedStaticExpand	0	Expand	Replaces a dynamic Expand target with an equivalent constant target.
ShapeTransposePattern / ShapeTranspose	0	Shape	Rewrites Shape(Transpose(X, perm)) into Gather(Shape(X), perm).
ShapedBasedReshapePattern / ShapedBasedReshape	0	Reshape	Simplifies reshapes using inferred input and output shapes.
SliceConcatToSpaceToDepthPattern / SliceConcatToSpaceToDepth	0	Concat	Fuses canonical rank-4 phase slices and channel Concat into SpaceToDepth.
SliceEliminationPattern / SliceElimination	1	Slice	Replaces an identity full-range Slice with Identity.
SliceSlicePattern / SliceSlice	0	Slice	Merges two consecutive Slice nodes on distinct axes into one Slice.
SlicesSplitPattern / SlicesSplit	0	Slice	Replaces sibling Slice nodes cutting a shared input into contiguous chunks along one axis by a single Split node.
SoftmaxCrossEntropyLossCastPattern / SoftmaxCrossEntropyLossCast	0	Div	Moves a compatible label cast into SoftmaxCrossEntropyLoss.
SplitConcatPattern / SplitConcat	0	Split	Replaces a Split immediately followed by a Concat that restores the original tensor with an Identity node.
SplitToSequenceSequenceAtPattern / SplitToSequenceSequenceAt	0	SplitToSequence	Replaces a SequenceAt reading a constant index of a SplitToSequence by a single Split output.
SqueezeAddPattern / SqueezeAdd	0	Add	Moves compatible squeeze operations across addition.
SqueezeBinaryUnsqueezePattern / SqueezeBinaryUnsqueeze	0	Unsqueeze	Simplifies squeeze, binary operation, and unsqueeze sequences.
SqueezeUnsqueezePattern / SqueezeUnsqueeze	0	Squeeze, Unsqueeze	Simplifies a Squeeze/Unsqueeze pair into an Identity or a single Squeeze.
StaticConcatReshapePattern / StaticConcatReshape	0	Reshape	Folds static concatenated reshape shapes.
Sub1MulPattern / Sub1Mul	0	Mul	Simplifies multiplication involving one minus a value.
SwapExpandReshapePattern / SwapExpandReshape	0	Reshape	Swaps a supported Expand and constant-shape Reshape pair.
SwapExpandUnsqueezePattern / SwapExpandUnsqueeze	0	Expand	Swaps Expand and a following Unsqueeze so the Unsqueeze runs on the smaller, pre-expansion tensor.
SwapRangeAddScalarPattern / SwapRangeAddScalar	0	Range	Moves scalar addition into compatible range operations.
SwapUnaryPattern / SwapUnary	0	Reshape, Squeeze, Transpose, Unsqueeze	Swaps compatible unary operations.
SwapUnsqueezeTransposePattern / SwapUnsqueezeTranspose	0	Transpose	Swaps compatible unsqueeze and transpose operations.
SwitchOrderBinaryPattern / SwitchOrderBinary	0	Add, Mul	Reorders compatible consecutive binary operations.
SwitchReshapeActivationPattern / SwitchReshapeActivation	1	Cos, Cosh, Elu, Erf, Exp, Gelu, LeakyRelu, Relu, Selu, Sin, Sinh, Softplus, Tan, Tanh	Moves compatible activations before Reshape.
TransposeEqualReshapePattern / TransposeEqualReshape	0	Transpose	Replaces shape-equivalent transposes with reshapes.
TransposeGatherPattern / TransposeGather	0	Gather	Removes or reorders a Transpose feeding a Gather with a scalar index.
TransposeMatMulPattern / TransposeMatMul	1	Gemm, MatMul	Folds compatible transposes into MatMul.
TransposeReshapeMatMulPattern / TransposeReshapeMatMul	1	MatMul	Simplifies transpose and reshape inputs to MatMul.
TransposeReshapeTransposePattern / TransposeReshapeTranspose	0	Transpose	Simplifies transpose, reshape, and transpose sequences.
TransposeToInitializerPattern / TransposeToInitializer	1	Transpose	Folds a Transpose applied to an initializer into a transposed initializer.
TransposeTransposePattern / TransposeTranspose	0	Transpose	Merges two consecutive Transpose nodes into a single Transpose or an Identity when the permutations cancel out.
TreeEnsemblePattern / TreeEnsemble	1	TreeEnsembleClassifier, TreeEnsembleRegressor	Replaces a classic tree ensemble with the unified TreeEnsemble operator.
UnsqueezeEqualPattern / UnsqueezeEqual	0	Equal	Rewrites Equal(Unsqueeze(x), Unsqueeze(y)) into Equal(x, y) when both Unsqueeze nodes use matching constant axes.
UnsqueezeOrSqueezeReshapePattern / UnsqueezeOrSqueezeReshape	0	Reshape	Simplifies reshape operations adjacent to squeeze or unsqueeze.
UnsqueezeReshapePattern / UnsqueezeReshape	0	Reshape	Simplifies an unsqueeze followed by reshape.
UnsqueezeShapePattern / UnsqueezeShape	0	Unsqueeze	Rewrites Shape(Unsqueeze(X, axes)) into a Concat of ranged Shape slices interleaved with constant [1] tensors.
UnsqueezeUnsqueezePattern / UnsqueezeUnsqueeze	0	Unsqueeze	Merges two consecutive Unsqueeze nodes into a single Unsqueeze.
WhereAddPattern / WhereAdd	0	Where	Factors a common additive term from Where branches built with Add.

The runtime list is available through standard_pattern_names(). The complete pattern catalogue in the ByOp catalogue adds the C++ documentation link and the Before/After rewrite graph for every entry.

See How to add a custom graph-rewriting pattern and set its priority for a Python/C++ how-to on writing a custom pattern and choosing its priority, and Optimizing a model with graph-rewriting patterns for a runnable example covering optimization statistics. Replay is demonstrated separately in Replaying graph-rewriting patterns.

Custom Python pattern

import onnx_light.onnx.helper as oh
from onnx_light.onnx_core.optimization import (
    GraphBuilder,
    GraphGraph,
    PatternOptimization,
)

class NegNegPattern(PatternOptimization):
    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, "input is not produced by Neg")
        return self.result([previous, node], insert_at=node)

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

builder = GraphBuilder(model)
builder.register_pattern(NegNegPattern())
graph = GraphGraph(builder)
rewrites = graph.optimize()

API

Graph-pattern optimization with standard and Python-defined patterns.

class onnx_light.onnx_core.optimization.AttentionGQAPattern(*args, **kwargs)

Fuses grouped-query attention cache handling.

class onnx_light.onnx_core.optimization.BatchNormalizationPattern(*args, **kwargs)

Fuses an inference batch-normalization subgraph.

class onnx_light.onnx_core.optimization.BatchNormalizationTrainingPattern(*args, **kwargs)

Fuses a training batch-normalization subgraph.

class onnx_light.onnx_core.optimization.CastCastBinaryPattern(*args, **kwargs)

Moves matching floating-point input Cast nodes after a binary operation.

Cast(x), Cast(y) -> Binary becomes Binary(x, y) -> Cast when precision and use guards allow it.

class onnx_light.onnx_core.optimization.CastCastPattern(*args, **kwargs)

Collapses two consecutive compatible Cast nodes.

x:A -> Cast(B) -> Cast(C) -> y:C becomes one safe Cast(C) or Identity.

class onnx_light.onnx_core.optimization.CastLayerNormalizationCastPattern(*args, **kwargs)

Removes redundant casts surrounding LayerNormalization.

class onnx_light.onnx_core.optimization.CastOpCastPattern(*args, **kwargs)

Moves a unary or binary operation to the result Cast type.

Compatible input Cast nodes and the trailing result Cast are removed or relocated while preserving shared outputs.

class onnx_light.onnx_core.optimization.CastPattern(*args, **kwargs)

Replaces a type-preserving Cast(to=T) with Identity.

x:T -> Cast(to=T) -> y:T becomes x:T -> Identity -> y:T.

class onnx_light.onnx_core.optimization.ClipClipPattern(*args, **kwargs)

Merges two consecutive Clip nodes with complementary bounds.

Clip(x, min) -> Clip(x1, , max) becomes one Clip(x, min, max) when one Clip defines the minimum and the other the maximum.

class onnx_light.onnx_core.optimization.ConcatEmptyPattern(*args, **kwargs)

Drops empty inputs from a Concat node, reducing it to an Identity when a single input remains.

class onnx_light.onnx_core.optimization.ConcatGatherPattern(*args, **kwargs)

Rewrites a Gather reading a single Concat input into a Gather on that input directly.

class onnx_light.onnx_core.optimization.ConcatReshapePattern(*args, **kwargs)

Simplifies concatenations that construct reshape shapes.

class onnx_light.onnx_core.optimization.ConcatSliceEliminationPattern(*args, **kwargs)

Eliminates a Concat followed by exact slices recovering all inputs.

class onnx_light.onnx_core.optimization.ConcatTwiceUnaryPattern(*args, **kwargs)

Pushes a shape-preserving unary op ahead of a Concat(x, x) so the unary op runs once on x.

class onnx_light.onnx_core.optimization.ConstantToInitializerPattern(*args, **kwargs)

Replaces a Constant node by an initializer and an Identity node.

class onnx_light.onnx_core.optimization.ConvAddFusionPattern(*args, **kwargs)

Folds a channel-wise constant Add following a Conv into the Conv bias.

class onnx_light.onnx_core.optimization.ConvBatchNormalizationFusionPattern(*args, **kwargs)

Folds inference BatchNormalization following a Conv into its weights and bias.

class onnx_light.onnx_core.optimization.ConvBiasNullPattern(*args, **kwargs)

Removes a null (all-zero) bias input from a Conv node.

class onnx_light.onnx_core.optimization.ConvMulFusionPattern(*args, **kwargs)

Folds a scalar or channel-wise constant Mul following a Conv into its constants.

class onnx_light.onnx_core.optimization.DivMulPattern(*args, **kwargs)

Fuses multiplication by a reciprocal into one division.

class onnx_light.onnx_core.optimization.DropoutPattern(*args, **kwargs)

Replaces an inference Dropout by an Identity node when its mask output is unused and training mode is disabled.

class onnx_light.onnx_core.optimization.ExpandBroadcastPattern(*args, **kwargs)

Drops an Expand feeding an element-wise binary operator that already broadcasts the pre-expanded input.

class onnx_light.onnx_core.optimization.ExpandPattern(*args, **kwargs)

Replaces Expand(x, shape) with Identity(x) when the target shape equals the input shape.

class onnx_light.onnx_core.optimization.ExpandSwapPattern(*args, **kwargs)

Moves an Expand past a following unary-like operator so the operator runs on the smaller tensor.

class onnx_light.onnx_core.optimization.ExpandUnsqueezeExpandPattern(*args, **kwargs)

Fuses Expand, Unsqueeze and Expand into a single Unsqueeze followed by one Expand.

class onnx_light.onnx_core.optimization.FunctionAttentionGQAPattern(*args, **kwargs)

Replaces grouped-query attention expressed with local functions.

class onnx_light.onnx_core.optimization.FunctionAttentionPattern(*args, **kwargs)

Replaces a scaled dot-product attention subgraph.

class onnx_light.onnx_core.optimization.FunctionCausalMaskMulAddPattern(*args, **kwargs)

Fuses scaling and offset operations into a causal-mask function.

class onnx_light.onnx_core.optimization.FunctionCausalMaskPattern(*args, **kwargs)

Replaces a causal-mask subgraph with a local function.

class onnx_light.onnx_core.optimization.FunctionCosSinCachePattern(*args, **kwargs)

Replaces cosine and sine cache construction with a local function.

class onnx_light.onnx_core.optimization.FunctionHalfRotaryEmbeddingPattern(*args, **kwargs)

Replaces half-rotary embedding construction with a local function.

class onnx_light.onnx_core.optimization.GatherConcatPattern(*args, **kwargs)

Merges a Concat of single-index Gather nodes on a shared input into one Gather node.

class onnx_light.onnx_core.optimization.GatherGatherPattern(*args, **kwargs)

Collapses two consecutive scalar Gather nodes into a single Gather node.

class onnx_light.onnx_core.optimization.GatherShapePattern(*args, **kwargs)

Rewrites a Gather of a scalar index over a Shape node into a narrowed Shape node.

class onnx_light.onnx_core.optimization.GatherSliceToSplitPattern(*args, **kwargs)

Fuses compatible sibling Gather and Slice ranges into one Split.

class onnx_light.onnx_core.optimization.GatherToSlicePattern(*args, **kwargs)

Rewrites a Gather selecting a constant scalar, singleton, or arithmetic-progression index into a Slice (plus Squeeze for a scalar index).

class onnx_light.onnx_core.optimization.GatherUpstreamPropagationPattern(*args, **kwargs)

Moves a scalar or vector constant-index Gather upstream across one compatible producer (pointwise ops, Transpose, Reshape, MatMul batch dims, Softmax, LayerNormalization) so it runs on smaller tensors.

class onnx_light.onnx_core.optimization.GathersSplitPattern(*args, **kwargs)

Replaces sibling Gather nodes selecting contiguous single indices of a shared input by a single Split node.

class onnx_light.onnx_core.optimization.GeluPattern(*args, **kwargs)

Fuses a GELU activation subgraph.

class onnx_light.onnx_core.optimization.GemmSumFusionPattern(*args, **kwargs)

Fuses a two-input Sum into an unbiased Gemm bias input.

class onnx_light.onnx_core.optimization.GemmTransposePattern(*args, **kwargs)

Folds input transposes into a Gemm operation.

class onnx_light.onnx_core.optimization.GraphBuilder(name: str = 'graph', schema_lookup: ~collections.abc.Callable[[str], list[~onnx_light.onnx_py._onnxpyprotoop.onnx_op.LightOpSchema]] | None = <function _default_schema_lookup>)

Incrementally builds an ONNX graph, model or function.

See onnx_light.onnx_core.graph_builder for details. By default the builder validates nodes and resolves opsets using the built-in ONNX operator schemas; pass schema_lookup=None to disable this, or a custom op_type -> list[LightOpSchema] callable to use different schemas.

clear_registered_patterns() → None

Removes every builder-local pattern.

init(value: ndarray, name: str | None = None, *, copy: bool = True) → str

Adds a NumPy initializer and returns its final name.

By default, copies the payload. With copy=False, borrows C-contiguous, dtype-aligned, little-endian storage without conversion. Supports bool, 8/16/32/64-bit integers, float16/32/64 and complex64/128; rejects other dtypes and layouts. Retains the array until the last borrowed payload owner releases it, including models exported from this builder.

Writable arrays remain writable and mutations are visible to all owners. The caller must finish mutations before optimization or creating an execution session, which may cache derived values, and must not resize or reallocate storage while borrowed. Use copy=True for an independent snapshot; setting an array read-only does not freeze its aliases.

inp(name: str, elem_type: int, shape: list[str | int | None]) → str

Declares and returns a compact graph input.

property op: _OperatorProxy

Returns the cached compact operator proxy.

out(name: str, elem_type: int | None = None, shape: list[str | int | None] | None = None) → str

Declares and returns a compact graph output.

register_pattern(pattern: PatternOptimization) → None

Registers or replaces a pattern for this builder.

registered_pattern_names() → tuple[str, ...]

Returns builder-local pattern names in registration order.

registered_patterns() → tuple[PatternOptimization, ...]

Returns builder-local patterns in registration order.

set_opset_version(domain: str, version: int) → None

Records an explicit opset version.

unregister_pattern(name: str) → bool

Removes a builder-local pattern and returns whether it existed.

class onnx_light.onnx_core.optimization.GraphGraph(builder: GraphBuilder, patterns: Iterable[str | PatternOptimization] | str | Pattern[str] | Device | Literal[False] | None = None)

Indexes a builder and selects the patterns used for rewriting.

patterns=None selects all registered device-independent patterns. False selects none. A concrete Device includes independent patterns and patterns targeting that exact device, and sets builder.device; a conflicting builder device raises ValueError.

A string or compiled regex selects registered names using fullmatch. An iterable selects only its exact names and pattern instances. These explicit selections can include device-specific patterns without changing the builder device. Builder registrations override global registrations before selection; repeated explicit names keep the last instance. Disabling patterns does not disable the optimizer’s cleanup passes.

class onnx_light.onnx_core.optimization.IdentityPattern(*args, **kwargs)

Replaces no-op arithmetic and layout operations by an Identity node.

class onnx_light.onnx_core.optimization.InitializerUnsqueezeCastPattern(*args, **kwargs)

Folds an initializer’s Unsqueeze and Cast into the initializer consumed by Add.

class onnx_light.onnx_core.optimization.LabelEncoderFusionPattern(*args, **kwargs)

Composes consecutive ai.onnx.ml LabelEncoder mappings.

class onnx_light.onnx_core.optimization.LayerNormalizationPattern(*args, **kwargs)

Fuses a layer-normalization subgraph.

class onnx_light.onnx_core.optimization.LayerNormalizationScalePattern(*args, **kwargs)

Fuses layer normalization with its scale.

class onnx_light.onnx_core.optimization.LeakyReluPattern(*args, **kwargs)

Fuses a LeakyRelu activation subgraph.

class onnx_light.onnx_core.optimization.LinearAttentionPattern(*args, **kwargs)

Fuses a single-token linear-attention recurrence.

class onnx_light.onnx_core.optimization.MatMulAddPattern(*args, **kwargs)

Replaces a compatible MatMul and Add with Gemm.

class onnx_light.onnx_core.optimization.MatMulBatchNormalizationFusionPattern(*args, **kwargs)

Folds constant inference BatchNormalization parameters into a rank-two MatMul.

class onnx_light.onnx_core.optimization.MatMulReshape2Of3Pattern(*args, **kwargs)

Simplifies compatible reshapes around MatMul.

class onnx_light.onnx_core.optimization.MatMulScaleFusionPattern(*args, **kwargs)

Absorbs one safe scalar Mul or Div adjacent to a rank-two MatMul.

class onnx_light.onnx_core.optimization.MaxReluPattern(*args, **kwargs)

Replaces a compatible maximum with Relu.

class onnx_light.onnx_core.optimization.MulMulMatMulPattern(*args, **kwargs)

Moves compatible scalar multiplications across MatMul.

class onnx_light.onnx_core.optimization.MulMulMulScalarPattern(*args, **kwargs)

Combines scalar factors in multiplication chains.

class onnx_light.onnx_core.optimization.MulUnsqueezeUnsqueezePattern(*args, **kwargs)

Simplifies multiplication of unsqueezed inputs.

class onnx_light.onnx_core.optimization.NotNotPattern(*args, **kwargs)

Fuses two consecutive Not nodes into an Identity node.

class onnx_light.onnx_core.optimization.NotWherePattern(*args, **kwargs)

Rewrites Where(Not(c), x, y) into Where(c, y, x).

class onnx_light.onnx_core.optimization.PadConvPattern(*args, **kwargs)

Folds a Pad node into the pads attribute of a following Conv node.

class onnx_light.onnx_core.optimization.PadPadFusionPattern(*args, **kwargs)

Merges adjacent constant-mode Pad nodes with equal values by summing their pads.

class onnx_light.onnx_core.optimization.PreShapeNodeEliminationPattern(*args, **kwargs)

Removes a Cast whose output only feeds Shape nodes, redirecting each Shape node directly to the Cast’s input.

class onnx_light.onnx_core.optimization.RMSNormalizationMulPattern(*args, **kwargs)

Fuses RMS normalization with a following scale.

class onnx_light.onnx_core.optimization.RMSNormalizationPattern(*args, **kwargs)

Fuses an RMS-normalization subgraph.

class onnx_light.onnx_core.optimization.ReduceArgTopKPattern(*args, **kwargs)

Simplifies compatible reduction, arg, and TopK operations.

class onnx_light.onnx_core.optimization.ReduceReshapePattern(*args, **kwargs)

Simplifies reshape operations around reductions.

class onnx_light.onnx_core.optimization.ReduceSumNormalizePattern(*args, **kwargs)

Simplifies reduce-sum normalization subgraphs.

class onnx_light.onnx_core.optimization.ReluClipFusionPattern(*args, **kwargs)

Removes Relu before Clip when the Clip minimum is constant and non-negative.

class onnx_light.onnx_core.optimization.Reshape2Of3Pattern(*args, **kwargs)

Simplifies two compatible reshapes among three branches.

class onnx_light.onnx_core.optimization.ReshapeMatMulReshapePattern(*args, **kwargs)

Simplifies reshape, MatMul, and reshape sequences.

class onnx_light.onnx_core.optimization.ReshapePattern(*args, **kwargs)

Removes or simplifies redundant reshape operations.

class onnx_light.onnx_core.optimization.ReshapeReshapeBinaryPattern(*args, **kwargs)

Moves compatible reshapes across binary operations.

class onnx_light.onnx_core.optimization.ReshapeReshapePattern(*args, **kwargs)

Collapses consecutive compatible reshapes.

class onnx_light.onnx_core.optimization.ReshapeSqueezePattern(*args, **kwargs)

Simplifies a reshape followed by squeeze.

class onnx_light.onnx_core.optimization.RotaryConcatPartPattern(*args, **kwargs)

Simplifies padded rotary concatenation subgraphs.

class onnx_light.onnx_core.optimization.RotaryEmbeddingPattern(*args, **kwargs)

Fuses a complete rotary-embedding subgraph.

class onnx_light.onnx_core.optimization.STFTFusionPattern(*args, **kwargs)

Fuses a canonical convolution-based DFT subgraph into one standard ONNX STFT.

class onnx_light.onnx_core.optimization.SameChildrenFromInputPattern(*args, **kwargs)

Eliminates equivalent computations from one input.

class onnx_light.onnx_core.optimization.SameChildrenPattern(*args, **kwargs)

Eliminates equivalent child computations.

class onnx_light.onnx_core.optimization.SequenceConstructAtPattern(*args, **kwargs)

Replaces a SequenceAt reading a constant index of a SequenceConstruct by the corresponding input tensor.

class onnx_light.onnx_core.optimization.ShapeBasedConcatExpandPattern(*args, **kwargs)

Simplifies a dynamic Concat target when Expand changes one dimension.

class onnx_light.onnx_core.optimization.ShapeBasedEditDistanceReshapePattern(*args, **kwargs)

Rewrites reshapes according to the distance between known shapes.

class onnx_light.onnx_core.optimization.ShapeBasedExpandBroadcastMatMulPattern(*args, **kwargs)

Removes dynamic Expand nodes from the batch dimensions of MatMul.

class onnx_light.onnx_core.optimization.ShapeBasedExpandBroadcastPattern(*args, **kwargs)

Removes dynamic Expand nodes before a broadcasting binary operator.

class onnx_light.onnx_core.optimization.ShapeBasedExpandCastWhereSwapPattern(*args, **kwargs)

Moves an Expand after a compatible Cast and Where chain.

class onnx_light.onnx_core.optimization.ShapeBasedExpandSwapPattern(*args, **kwargs)

Moves input Expand nodes after a broadcasting binary operator.

class onnx_light.onnx_core.optimization.ShapeBasedIdentityPattern(*args, **kwargs)

Eliminates shape-proven identity operations.

class onnx_light.onnx_core.optimization.ShapeBasedMatMulToMulPattern(*args, **kwargs)

Replaces shape-proven scalar MatMul with Mul.

class onnx_light.onnx_core.optimization.ShapeBasedReshapeIsSqueezePattern(*args, **kwargs)

Replaces eligible reshapes with squeeze.

class onnx_light.onnx_core.optimization.ShapeBasedSameChildrenPattern(*args, **kwargs)

Eliminates shape-equivalent child computations.

class onnx_light.onnx_core.optimization.ShapeBasedShapeShapeAddPattern(*args, **kwargs)

Exposes the upstream placeholder for additions of two Shape outputs.

class onnx_light.onnx_core.optimization.ShapeBasedStaticExpandPattern(*args, **kwargs)

Replaces a dynamic Expand target with an equivalent constant target.

class onnx_light.onnx_core.optimization.ShapeTransposePattern(*args, **kwargs)

Rewrites Shape(Transpose(X, perm)) into Gather(Shape(X), perm).

class onnx_light.onnx_core.optimization.ShapedBasedReshapePattern(*args, **kwargs)

Simplifies reshapes using inferred input and output shapes.

class onnx_light.onnx_core.optimization.SliceConcatToSpaceToDepthPattern(*args, **kwargs)

Fuses canonical rank-4 phase slices and channel Concat into SpaceToDepth.

class onnx_light.onnx_core.optimization.SliceEliminationPattern(*args, **kwargs)

Replaces an identity full-range Slice with Identity.

class onnx_light.onnx_core.optimization.SliceSlicePattern(*args, **kwargs)

Merges two consecutive Slice nodes on distinct axes into one Slice.

class onnx_light.onnx_core.optimization.SlicesSplitPattern(*args, **kwargs)

Replaces sibling Slice nodes cutting a shared input into contiguous chunks along one axis by a single Split node.

class onnx_light.onnx_core.optimization.SoftmaxCrossEntropyLossCastPattern(*args, **kwargs)

Moves a compatible label cast into SoftmaxCrossEntropyLoss.

class onnx_light.onnx_core.optimization.SplitConcatPattern(*args, **kwargs)

Replaces a Split immediately followed by a Concat that restores the original tensor with an Identity node.

class onnx_light.onnx_core.optimization.SplitToSequenceSequenceAtPattern(*args, **kwargs)

Replaces a SequenceAt reading a constant index of a SplitToSequence by a single Split output.

class onnx_light.onnx_core.optimization.SqueezeAddPattern(*args, **kwargs)

Moves compatible squeeze operations across addition.

class onnx_light.onnx_core.optimization.SqueezeBinaryUnsqueezePattern(*args, **kwargs)

Simplifies squeeze, binary operation, and unsqueeze sequences.

class onnx_light.onnx_core.optimization.SqueezeUnsqueezePattern(*args, **kwargs)

Simplifies a Squeeze/Unsqueeze pair into an Identity or a single Squeeze.

class onnx_light.onnx_core.optimization.StaticConcatReshapePattern(*args, **kwargs)

Folds static concatenated reshape shapes.

class onnx_light.onnx_core.optimization.Sub1MulPattern(*args, **kwargs)

Simplifies multiplication involving one minus a value.

class onnx_light.onnx_core.optimization.SwapExpandReshapePattern(*args, **kwargs)

Swaps a supported Expand and constant-shape Reshape pair.

class onnx_light.onnx_core.optimization.SwapExpandUnsqueezePattern(*args, **kwargs)

Swaps Expand and a following Unsqueeze so the Unsqueeze runs on the smaller, pre-expansion tensor.

class onnx_light.onnx_core.optimization.SwapRangeAddScalarPattern(*args, **kwargs)

Moves scalar addition into compatible range operations.

class onnx_light.onnx_core.optimization.SwapUnaryPattern(*args, **kwargs)

Swaps compatible unary operations.

class onnx_light.onnx_core.optimization.SwapUnsqueezeTransposePattern(*args, **kwargs)

Swaps compatible unsqueeze and transpose operations.

class onnx_light.onnx_core.optimization.SwitchOrderBinaryPattern(*args, **kwargs)

Reorders compatible consecutive binary operations.

class onnx_light.onnx_core.optimization.SwitchReshapeActivationPattern(*args, **kwargs)

Moves compatible activations before Reshape.

class onnx_light.onnx_core.optimization.TransposeEqualReshapePattern(*args, **kwargs)

Replaces shape-equivalent transposes with reshapes.

class onnx_light.onnx_core.optimization.TransposeGatherPattern(*args, **kwargs)

Removes or reorders a Transpose feeding a Gather with a scalar index.

class onnx_light.onnx_core.optimization.TransposeMatMulPattern(*args, **kwargs)

Folds compatible transposes into MatMul.

class onnx_light.onnx_core.optimization.TransposeReshapeMatMulPattern(*args, **kwargs)

Simplifies transpose and reshape inputs to MatMul.

class onnx_light.onnx_core.optimization.TransposeReshapeTransposePattern(*args, **kwargs)

Simplifies transpose, reshape, and transpose sequences.

class onnx_light.onnx_core.optimization.TransposeToInitializerPattern(*args, **kwargs)

Folds a Transpose applied to an initializer into a transposed initializer.

class onnx_light.onnx_core.optimization.TransposeTransposePattern(*args, **kwargs)

Merges two consecutive Transpose nodes into a single Transpose or an Identity when the permutations cancel out.

class onnx_light.onnx_core.optimization.TreeEnsemblePattern(*args, **kwargs)

Replaces a classic tree ensemble with the unified TreeEnsemble operator.

class onnx_light.onnx_core.optimization.UnsqueezeEqualPattern(*args, **kwargs)

Rewrites Equal(Unsqueeze(x), Unsqueeze(y)) into Equal(x, y) when both Unsqueeze nodes use matching constant axes.

class onnx_light.onnx_core.optimization.UnsqueezeOrSqueezeReshapePattern(*args, **kwargs)

Simplifies reshape operations adjacent to squeeze or unsqueeze.

class onnx_light.onnx_core.optimization.UnsqueezeReshapePattern(*args, **kwargs)

Simplifies an unsqueeze followed by reshape.

class onnx_light.onnx_core.optimization.UnsqueezeShapePattern(*args, **kwargs)

Rewrites Shape(Unsqueeze(X, axes)) into a Concat of ranged Shape slices interleaved with constant [1] tensors.

class onnx_light.onnx_core.optimization.UnsqueezeUnsqueezePattern(*args, **kwargs)

Merges two consecutive Unsqueeze nodes into a single Unsqueeze.

class onnx_light.onnx_core.optimization.WhereAddPattern(*args, **kwargs)

Factors a common additive term from Where branches built with Add.

onnx_light.onnx_core.optimization.clear_registered_patterns() → None

Removes every globally registered pattern, including standard patterns.

onnx_light.onnx_core.optimization.register_pattern(pattern: PatternOptimization) → None

Registers or replaces a process-global pattern.

onnx_light.onnx_core.optimization.registered_pattern_names() → tuple[str, ...]

Returns global pattern names in registration order.

onnx_light.onnx_core.optimization.registered_patterns() → tuple[PatternOptimization, ...]

Returns global patterns in registration order.

onnx_light.onnx_core.optimization.render_rst_standard_patterns_table() → str

Renders the standard ONNX patterns as a reST list-table.

The table is generated from the patterns returned by standard_patterns(), so it stays in sync with the registered patterns without any manual maintenance.

Returns:

The list-table directive as a reST string.

onnx_light.onnx_core.optimization.replay(model: ModelProto, rewrites: Iterable[LocalRewriting], schema_lookup: SchemaLookup | None = <function _default_schema_lookup>) → GraphProto

Replays captured rewrites and returns the resulting graph.

onnx_light.onnx_core.optimization.reset_registered_patterns() → None

Restores the global registry to the standard ONNX patterns.

onnx_light.onnx_core.optimization.standard_pattern_names() → list[str]

Returns the standard ONNX pattern names.

onnx_light.onnx_core.optimization.standard_patterns(names: Iterable[str] | None = None) → list[PatternOptimization]

Creates the selected standard ONNX patterns.

onnx_light.onnx_core.optimization.unregister_pattern(name: str) → bool

Removes a global pattern and returns whether it existed.