1 |
AttentionGQA (attention)
|
Python — Converts cached grouped-query attention into ONNX Attention cache inputs and outputs.
|
2 |
BatchNormalization (normalization)
|
Python — Removes an identity BatchNormalization in inference mode.
|
3 |
BatchNormalizationTraining (normalization)
|
Python — Expands training BatchNormalization into primitive operators.
|
4 |
Cast (canonicalization)
|
Python — Canonicalizes a Cast to Identity when its input already has the target type.
|
5 |
CastCast (canonicalization)
|
Python — Canonicalizes consecutive Cast nodes when a single conversion is equivalent.
|
6 |
CastCastBinary (canonicalization)
|
Python — Moves matching floating-point Cast nodes after a binary arithmetic operation.
|
7 |
CastLayerNormalizationCast (normalization)
|
Python — Moves a normalization from its stash type back to the original element type.
|
8 |
CastOpCast (canonicalization)
|
Python — Moves an operation from a temporary floating-point type to its result type.
|
9 |
ClipClip (canonicalization)
|
Python — Merges two consecutive Clip nodes when one defines the minimum and the other the maximum.
|
10 |
ConcatEmpty (collections)
|
Python — Removes Concat inputs that are empty along the concatenation axis.
|
11 |
ConcatGather (collections)
|
Python — Simplifies Gather(Concat(...), cst_index) when the index is a constant single-element int64 tensor and every Concat input is a 1-D tensor of a statically known size.
|
12 |
ConcatReshape (reshape)
|
Python — Replaces dynamic dimensions in a concatenated Reshape target with one inferred dimension.
|
13 |
ConcatSliceElimination (collections)
|
Python — Removes a Concat whose consumers are exact, non-overlapping Slice nodes that recover every input in order along the concatenation axis.
|
14 |
ConcatTwiceUnary (collections)
|
Python — Pushes a shape-preserving unary operator through a Concat(x, x) so that Unary(Concat(x, x)) becomes Concat(Unary(x), Unary(x)), exposing the shared Unary(x) for common-subexpression elimination.
|
15 |
ConstantToInitializer (canonicalization)
|
Python — Replaces a Constant node by an initializer and an Identity node.
|
16 |
ConvAddFusion (canonicalization)
|
Python — Fuses a channel-wise Add following a Conv into the Conv bias.
|
17 |
ConvBatchNormalizationFusion (canonicalization)
|
Python — Fuses inference BatchNormalization following a Conv into the Conv weights and bias.
|
18 |
ConvBiasNull (canonicalization)
|
Python — Removes a null (all-zero) bias input from a Conv node.
|
19 |
ConvMulFusion (canonicalization)
|
Python — Fuses a scalar or channel-wise Mul following a Conv into its weights and optional bias.
|
20 |
DivMul (algebra)
|
Python — Fuses a reciprocal followed by multiplication into one division.
|
21 |
Dropout (canonicalization)
|
Python — Replaces an eligible Dropout by an Identity node.
|
22 |
Expand (expand)
|
Python — Removes an Expand that does not change the shape of its input.
|
23 |
ExpandBroadcast (expand)
|
Python — Drops an Expand feeding an element-wise binary operator.
|
24 |
ExpandSwap (expand)
|
Python — Moves an Expand past a following unary-like operator.
|
25 |
ExpandUnsqueezeExpand (expand)
|
Python — Fuses Expand, Unsqueeze and Expand into Unsqueeze then Expand.
|
26 |
FunctionAttention (attention)
|
Python — Fuses a masked scaled-dot-product attention graph into a local function.
|
27 |
FunctionAttentionGQA (attention)
|
Python — Moves matching GQA repeat-interleave branches into a LocalAttention function.
|
28 |
FunctionCausalMask (attention)
|
Python — Fuses a causal-mask index construction into a local function.
|
29 |
FunctionCausalMaskMulAdd (attention)
|
Python — Fuses an additive causal-mask index construction into a local function.
|
30 |
FunctionCosSinCache (attention)
|
Python — Fuses cosine and sine rotary-cache generation into a local function.
|
31 |
FunctionHalfRotaryEmbedding (attention)
|
Python — Fuses the canonical half-rotary decomposition into a local function.
|
32 |
GatherConcat (collections)
|
Python — Simplifies Gather(Concat(..., X, ..., axis=0), cst) into Gather(X, cst - offset) when exactly one Concat input X is non-constant, all others are constants of known size, X is 1-D, and every requested index falls inside X’s slice of the concatenation.
|
33 |
GatherGather (collections)
|
Python — Composes two consecutive axis=0 Gather nodes with constant indices into a single Gather.
|
34 |
GatherShape (collections)
|
Python — Simplifies Gather(Shape(X), indices) into Shape(X, start, end) when indices is a constant int64 scalar or a contiguous ascending 1-D range.
|
35 |
GatherSliceToSplit (collections)
|
Python — Replaces compatible sibling scalar Gather and contiguous Slice ranges covering a complete static axis with one Split. Scalar Gather outputs are restored with Squeeze.
|
36 |
GatherToSlice (collections)
|
Python — Replaces a Gather selecting a constant scalar index, a constant single-element vector index, or a constant arithmetic-progression vector of indices (a “range” with a fixed, strictly ascending step) by an equivalent Slice. Slice is generally cheaper to execute than Gather and, unlike Gather, it never triggers a data-dependent gather kernel.
|
37 |
GatherUpstreamPropagation (collections)
|
Python — Moves a scalar or vector constant-index Gather upstream across one compatible producer node, so that the producer runs on the smaller (already gathered) tensor(s) instead of the original full-size ones.
|
38 |
GathersSplit (collections)
|
Python — Merges several sibling Gather nodes reading consecutive constant indices 0, 1, ..., n-1 on the same axis of a common input into a single Split (followed by Squeeze nodes when the indices are scalars).
|
39 |
Gelu (normalization)
|
Python — Fuses the tanh approximation of Gelu.
|
40 |
GemmSumFusion (matmul)
|
Python — Fuses a two-input Sum following an unbiased Gemm into the Gemm bias input.
|
41 |
GemmTranspose (matmul)
|
Python — Exposes a constant Gemm weight transpose while preserving Gemm semantics.
|
42 |
Identity (canonicalization)
|
Python — Replaces no-op operations by an Identity node.
|
43 |
InitializerUnsqueezeCast (canonicalization)
|
Python — Folds an initializer’s Unsqueeze and Cast into the initializer consumed by Add.
|
44 |
LabelEncoderFusion (traditionalml)
|
Python — Composes two consecutive ai.onnx.ml::LabelEncoder mappings, including propagation of the first encoder’s default through the second encoder.
|
45 |
LayerNormalization (normalization)
|
Python — Fuses an explicit layer-normalization decomposition.
|
46 |
LayerNormalizationScale (normalization)
|
Python — Folds a following affine transform into LayerNormalization parameters.
|
47 |
LeakyRelu (normalization)
|
Python — Fuses a Where-based LeakyRelu decomposition.
|
48 |
LinearAttention (attention)
|
Python — Fuses a single-token linear-attention recurrence into LinearAttention.
|
49 |
MatMulAdd (matmul)
|
Python — Fuses an Add bias into a MatMul or Gemm.
|
50 |
MatMulBatchNormalizationFusion (matmul)
|
Python — Folds inference BatchNormalization parameters into a constant MatMul weight.
|
51 |
MatMulReshape2Of3 (matmul)
|
Python — Normalizes adjacent Reshapes around a MatMul to the final batch shape.
|
52 |
MatMulScaleFusion (matmul)
|
Python — Absorbs one scalar Mul or safe Div adjacent to a rank-two MatMul.
|
53 |
MaxRelu (normalization)
|
Python — Replaces a maximum with scalar zero by Relu.
|
54 |
MulMulMatMul (matmul)
|
Python — Moves scalar factors from both MatMul inputs to one multiplication after the MatMul.
|
55 |
MulMulMulScalar (algebra)
|
Python — Combines the scalar constants from three nested Mul or Div nodes.
|
56 |
MulUnsqueezeUnsqueeze (layout)
|
Python — Moves equal, unshared Unsqueeze operations after a Mul.
|
57 |
NotNot (canonicalization)
|
Python — Fuses two consecutive Not nodes into an Identity.
|
58 |
NotWhere (expand)
|
Python — Rewrites Where(Not(c), x, y) into Where(c, y, x).
|
59 |
PadConv (canonicalization)
|
Python — Fuses a Pad node followed by a Conv node into a single Conv node whose pads attribute absorbs the spatial padding.
|
60 |
PadPadFusion (canonicalization)
|
Python — Merges two adjacent constant-mode Pad nodes with equal constant values.
|
61 |
PreShapeNodeElimination (collections)
|
Python — Removes a Cast node whose output only feeds Shape consumers, redirecting each such Shape node directly to the Cast’s input.
|
62 |
RMSNormalization (normalization)
|
Python — Fuses an explicit root-mean-square normalization decomposition.
|
63 |
RMSNormalizationMul (normalization)
|
Python — Folds a constant post-scale into RMSNormalization.
|
64 |
ReduceArgTopK (algebra)
|
Python — Fuses matching ReduceMin/ReduceMax and ArgMin/ArgMax nodes into TopK with K=1.
|
65 |
ReduceReshape (reshape)
|
Python — Fuses a dimension-removing Reshape into a reduction.
|
66 |
ReduceSumNormalize (algebra)
|
Python — Moves a Cast-ReduceSum-Mul-Sub-Cast normalization chain to the result type.
|
67 |
ReluClipFusion (canonicalization)
|
Python — Removes a Relu immediately followed by a Clip whose effective minimum is non-negative.
|
68 |
Reshape (reshape)
|
Python — Removes a Reshape whose constant target already equals its input shape.
|
69 |
Reshape2Of3 (reshape)
|
Python — Moves a binary operation to the common outer shape represented by surrounding Reshapes.
|
70 |
ReshapeMatMulReshape (matmul)
|
Python — Removes batch-flattening Reshapes around a MatMul.
|
71 |
ReshapeReshape (reshape)
|
Python — Composes two consecutive Reshapes into one.
|
72 |
ReshapeReshapeBinary (reshape)
|
Python — Moves equal input Reshapes after an element-wise binary operation.
|
73 |
ReshapeSqueeze (reshape)
|
Python — Fuses a Squeeze into the constant target of a preceding Reshape.
|
74 |
RotaryConcatPart (attention)
|
Python — Collapses two zero-padded rotary branches into Split, Neg, and Concat.
|
75 |
RotaryEmbedding (attention)
|
Python — Replaces a half-rotary local function and doubled caches with RotaryEmbedding.
|
76 |
STFTFusion (canonicalization)
|
Python — Fuses a canonical convolution-based DFT into a standard ONNX STFT.
|
77 |
SameChildren (algebra)
|
Python — Merges pairs of identical sibling nodes, and any identical descendant chains.
|
78 |
SameChildrenFromInput (algebra)
|
Python — Merges identical nodes that consume the same graph input.
|
79 |
SequenceConstructAt (collections)
|
Python — Replaces SequenceConstruct(x0, x1, ...) followed by the matching SequenceAt(seq, 0), SequenceAt(seq, 1), … with one Identity per SequenceAt, forwarding the original tensor directly.
|
80 |
ShapeBasedConcatExpand (expand)
|
Python — Simplifies a dynamic Concat used as an Expand target shape.
|
81 |
ShapeBasedEditDistanceReshape (reshape)
|
Python — Materializes a Concat-built Reshape target from inferred input and output shapes.
|
82 |
ShapeBasedExpandBroadcast (expand)
|
Python — Removes dynamic Expand nodes before a broadcasting binary operator.
|
83 |
ShapeBasedExpandBroadcastMatMul (expand)
|
Python — Removes dynamic Expand nodes before MatMul.
|
84 |
ShapeBasedExpandCastWhereSwap (expand)
|
Python — Moves an Expand after Cast and Where.
|
85 |
ShapeBasedExpandSwap (expand)
|
Python — Moves input Expand nodes after a broadcasting binary operator.
|
86 |
ShapeBasedIdentity (algebra)
|
Python — Replaces a shape-preserving Slice by an Identity.
|
87 |
ShapeBasedMatMulToMul (matmul)
|
Python — Replaces a MatMul with unit reduction dimensions by element-wise Mul.
|
88 |
ShapeBasedReshapeIsSqueeze (reshape)
|
Python — Replaces a Reshape or all-one Expand that only inserts or removes unit dimensions.
|
89 |
ShapeBasedSameChildren (algebra)
|
Python — Merges Expand or Reshape siblings whose inferred output shapes are equal.
|
90 |
ShapeBasedShapeShapeAdd (algebra)
|
Python — Recognizes an Add fed by two Shape nodes but performs no rewrite.
|
91 |
ShapeBasedStaticExpand (expand)
|
Python — Replaces a dynamic Expand target with an equivalent constant target.
|
92 |
ShapeTranspose (collections)
|
Python — Replaces Shape(Transpose(X, perm)) by Gather(Shape(X), perm) so the expensive Transpose on the full data tensor is avoided.
|
93 |
ShapedBasedReshape (reshape)
|
Python — Removes a same-rank Reshape whose target copies every dimension but the last.
|
94 |
SliceConcatToSpaceToDepth (collections)
|
Python — Replaces the canonical rank-4 four-phase slicing and channel concatenation used by focus layers with the standard ONNX SpaceToDepth(blocksize=2).
|
95 |
SliceElimination (collections)
|
Python — Replaces a full-range Slice by an Identity.
|
96 |
SliceSlice (collections)
|
Python — Merges two consecutive Slice nodes acting on disjoint axes into a single Slice whose starts / ends / axes (and steps when present) are the concatenation of the two operand vectors.
|
97 |
SlicesSplit (collections)
|
Python — Merges several sibling Slice nodes that partition a common input along a single axis into contiguous, non-overlapping ranges into one Split.
|
98 |
SoftmaxCrossEntropyLossCast (normalization)
|
Python — Fuses a float16 mean cross-entropy loss decomposition.
|
99 |
SplitConcat (collections)
|
Python — Replaces a Split immediately followed by a Concat that re-joins all of its outputs, in order and on the same axis, with a single Identity.
|
100 |
SplitToSequenceSequenceAt (collections)
|
Python — Replaces SplitToSequence(x, [split], axis=a) followed by the matching SequenceAt(seq, 0), SequenceAt(seq, 1), … with a single Split.
|
101 |
SqueezeAdd (layout)
|
Python — Moves equal Squeeze operations after an Add.
|
102 |
SqueezeBinaryUnsqueeze (layout)
|
Python — Cancels a scalar-producing Squeeze/binary/Unsqueeze chain by expanding the scalar right operand instead.
|
103 |
SqueezeUnsqueeze (unsqueeze)
|
Python — Simplifies a Squeeze/Unsqueeze pair into Identity or Squeeze.
|
104 |
StaticConcatReshape (reshape)
|
Python — Replaces the sole dynamic element of a concatenated Reshape target with [-1].
|
105 |
Sub1Mul (algebra)
|
Python — Rewrites multiplication by 1 - x into a product followed by subtraction.
|
106 |
SwapExpandReshape (expand)
|
Python — Swaps an Expand with a following constant-shape Reshape.
|
107 |
SwapExpandUnsqueeze (expand)
|
Python — Swaps an Expand and a following Unsqueeze.
|
108 |
SwapRangeAddScalar (algebra)
|
Python — Moves a shape-[1] second Add input from after Range into its start and limit inputs.
|
109 |
SwapUnary (algebra)
|
Python — Moves a rank-preserving layout operation after a compatible following operation.
|
110 |
SwapUnsqueezeTranspose (layout)
|
Python — Swaps an unshared Unsqueeze and Transpose, remapping both axes and permutation.
|
111 |
SwitchOrderBinary (algebra)
|
Python — Reassociates nested Add or Mul nodes to reduce broadcasting rank.
|
112 |
SwitchReshapeActivation (matmul)
|
Python — Moves an element-wise activation before a Reshape or Transpose.
|
113 |
TransposeEqualReshape (layout)
|
Python — Replaces a Transpose that only repositions size-one axes around at most one other dimension with a Reshape.
|
114 |
TransposeGather (transpose)
|
Python — Removes an unnecessary Transpose feeding a Gather with a scalar index.
|
115 |
TransposeMatMul (matmul)
|
Python — Absorbs rank-two input Transposes into Gemm transpose attributes.
|
116 |
TransposeReshapeMatMul (matmul)
|
Python — Swaps a last-two-dimension Transpose with a following Reshape on one MatMul input.
|
117 |
TransposeReshapeTranspose (layout)
|
Python — Moves a constant Reshape across one of two adjacent Transpose nodes when the input and target dimensions can be aligned by merging or splitting contiguous dimensions.
|
118 |
TransposeToInitializer (transpose)
|
Python — Folds a Transpose whose input is an initializer.
|
119 |
TransposeTranspose (transpose)
|
Python — Merges two consecutive Transpose nodes into a single one.
|
120 |
TreeEnsemble (traditionalml)
|
Python — Replaces classic TreeEnsembleRegressor and TreeEnsembleClassifier nodes with the unified ai.onnx.ml TreeEnsemble operator from opset 5.
|
121 |
UnsqueezeEqual (expand)
|
Python — Moves Equal(x, c) after a compatible sibling Unsqueeze(x, a) and preserves the rank of Equal(Unsqueeze(x, a), Unsqueeze(y, a)).
|
122 |
UnsqueezeOrSqueezeReshape (reshape)
|
Python — Removes a Squeeze or Unsqueeze immediately before a Reshape.
|
123 |
UnsqueezeReshape (reshape)
|
Python — Collapses a specific Unsqueeze-Reshape sequence into one Unsqueeze.
|
124 |
UnsqueezeShape (collections)
|
Python — Replaces Shape(Unsqueeze(X, axes)) by a Concat of Shape(X) slices interleaved with constant [1] tensors at the inserted axis positions.
|
125 |
UnsqueezeUnsqueeze (unsqueeze)
|
Python — Merges two consecutive Unsqueeze nodes into a single one.
|
126 |
WhereAdd (expand)
|
Python — Rewrites an additive mask or factors a common term from Where branches.
|
