onnx-ir 0.21.0

ONNX-IR is a pure Rust library for parsing ONNX models into an intermediate representation that can be used to generate code for various ML/DL frameworks
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275

#!/usr/bin/env -S uv run
# /// script
# dependencies = [
#   "onnx>=1.15.0",
# ]
# ///

"""
Generate ONNX models for testing the constant folding simplification pass.

Models exercise various constant folding scenarios at the IR level:
1. Cascading shape->gather->mul fold
2. Chained arithmetic (add then sub on constants)
3. Dynamic input blocking fold (negative test)
"""

import onnx
from onnx import helper, TensorProto


def save(model, name):
    onnx.checker.check_model(model)
    path = f"../fixtures/{name}"
    onnx.save(model, path)
    print(f"  {path}")


def constant_fold_cascade():
    """Shape->Gather(dim1) * Shape->Gather(dim2) folds to a single constant.

    x: [2, 3, 4]
    dim1 = Shape(x)[1] = 3   (folded by constant_shape)
    dim2 = Shape(x)[2] = 4   (folded by constant_shape)
    product = Mul(dim1, dim2) = 12  (folded by constant_fold)
    """
    graph = helper.make_graph(
        name="main_graph",
        nodes=[
            helper.make_node("Shape", ["x"], ["shape1"]),
            helper.make_node("Shape", ["x"], ["shape2"]),
            helper.make_node(
                "Constant", [], ["idx1"],
                value=helper.make_tensor("v", TensorProto.INT64, [], [1]),
            ),
            helper.make_node(
                "Constant", [], ["idx2"],
                value=helper.make_tensor("v", TensorProto.INT64, [], [2]),
            ),
            helper.make_node("Gather", ["shape1", "idx1"], ["dim1"], axis=0),
            helper.make_node("Gather", ["shape2", "idx2"], ["dim2"], axis=0),
            helper.make_node("Mul", ["dim1", "dim2"], ["product"]),
        ],
        inputs=[
            helper.make_value_info(
                "x",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[2, 3, 4]),
            ),
        ],
        outputs=[
            helper.make_value_info(
                "product",
                helper.make_tensor_type_proto(TensorProto.INT64, shape=[]),
            ),
        ],
    )
    save(
        helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 16)]),
        "constant_fold_cascade.onnx",
    )


def constant_fold_chain():
    """Chained arithmetic on constants: Add(2, 3) -> Sub(result, 1) -> 4.

    All inputs are initializer constants, so the entire chain should fold.
    """
    graph = helper.make_graph(
        name="main_graph",
        nodes=[
            helper.make_node("Add", ["a", "b"], ["sum"]),
            helper.make_node("Sub", ["sum", "c"], ["result"]),
        ],
        inputs=[],
        outputs=[
            helper.make_value_info(
                "result",
                helper.make_tensor_type_proto(TensorProto.INT64, shape=[]),
            ),
        ],
        initializer=[
            helper.make_tensor("a", TensorProto.INT64, [], [2]),
            helper.make_tensor("b", TensorProto.INT64, [], [3]),
            helper.make_tensor("c", TensorProto.INT64, [], [1]),
        ],
    )
    save(
        helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 16)]),
        "constant_fold_chain.onnx",
    )


def constant_fold_blocked():
    """Dynamic input blocks folding: Add(dynamic_x, const) cannot fold.

    Only the Neg(const) can be folded; the Add must remain.
    """
    graph = helper.make_graph(
        name="main_graph",
        nodes=[
            # Neg(-5) = 5, foldable
            helper.make_node("Neg", ["offset"], ["neg_offset"]),
            # Add(x, 5), NOT foldable (x is dynamic)
            helper.make_node("Add", ["x", "neg_offset"], ["result"]),
        ],
        inputs=[
            helper.make_value_info(
                "x",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[2, 3]),
            ),
        ],
        outputs=[
            helper.make_value_info(
                "result",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[2, 3]),
            ),
        ],
        initializer=[
            helper.make_tensor("offset", TensorProto.FLOAT, [], [-5.0]),
        ],
    )
    save(
        helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 16)]),
        "constant_fold_blocked.onnx",
    )


def constant_fold_concat():
    """Concat of constant arrays folds into a single constant.

    Shape(x)[0:2] ++ Shape(y)[0:1] -> [2, 3, 4]
    Uses Shape->Slice to produce constant arrays, then Concat merges them.
    """
    graph = helper.make_graph(
        name="main_graph",
        nodes=[
            helper.make_node("Shape", ["x"], ["shape_x"]),
            helper.make_node("Shape", ["y"], ["shape_y"]),
            helper.make_node(
                "Constant", [], ["starts_a"],
                value=helper.make_tensor("v", TensorProto.INT64, [1], [0]),
            ),
            helper.make_node(
                "Constant", [], ["ends_a"],
                value=helper.make_tensor("v", TensorProto.INT64, [1], [2]),
            ),
            helper.make_node(
                "Constant", [], ["starts_b"],
                value=helper.make_tensor("v", TensorProto.INT64, [1], [0]),
            ),
            helper.make_node(
                "Constant", [], ["ends_b"],
                value=helper.make_tensor("v", TensorProto.INT64, [1], [1]),
            ),
            helper.make_node("Slice", ["shape_x", "starts_a", "ends_a"], ["slice_x"]),
            helper.make_node("Slice", ["shape_y", "starts_b", "ends_b"], ["slice_y"]),
            helper.make_node("Concat", ["slice_x", "slice_y"], ["result"], axis=0),
        ],
        inputs=[
            helper.make_value_info(
                "x",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[2, 3, 4]),
            ),
            helper.make_value_info(
                "y",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[5, 6]),
            ),
        ],
        outputs=[
            helper.make_value_info(
                "result",
                helper.make_tensor_type_proto(TensorProto.INT64, shape=[3]),
            ),
        ],
    )
    save(
        helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 16)]),
        "constant_fold_concat.onnx",
    )


def constant_fold_cast():
    """Shape->Gather folds to i64 constant, then Cast(to=FLOAT) folds to f32.

    x: [2, 3, 4]
    dim = Shape(x)[1] = 3   (folded by constant_shape)
    float_dim = Cast(dim, to=FLOAT) = 3.0  (folded by constant_fold)
    """
    graph = helper.make_graph(
        name="main_graph",
        nodes=[
            helper.make_node("Shape", ["x"], ["shape1"]),
            helper.make_node(
                "Constant", [], ["idx"],
                value=helper.make_tensor("v", TensorProto.INT64, [], [1]),
            ),
            helper.make_node("Gather", ["shape1", "idx"], ["dim"], axis=0),
            helper.make_node("Cast", ["dim"], ["float_dim"], to=TensorProto.FLOAT),
        ],
        inputs=[
            helper.make_value_info(
                "x",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[2, 3, 4]),
            ),
        ],
        outputs=[
            helper.make_value_info(
                "float_dim",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[]),
            ),
        ],
    )
    save(
        helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 16)]),
        "constant_fold_cast.onnx",
    )


def constant_fold_sqrt():
    """Shape->Gather->Cast->Sqrt cascade: computes sqrt(head_dim) at compile time.

    x: [2, 3, 64]
    dim = Shape(x)[2] = 64       (folded by constant_shape)
    float_dim = Cast(dim) = 64.0  (folded by constant_fold: Cast)
    scale = Sqrt(float_dim) = 8.0 (folded by constant_fold: Sqrt)
    """
    graph = helper.make_graph(
        name="main_graph",
        nodes=[
            helper.make_node("Shape", ["x"], ["shape1"]),
            helper.make_node(
                "Constant", [], ["idx"],
                value=helper.make_tensor("v", TensorProto.INT64, [], [2]),
            ),
            helper.make_node("Gather", ["shape1", "idx"], ["dim"], axis=0),
            helper.make_node("Cast", ["dim"], ["float_dim"], to=TensorProto.FLOAT),
            helper.make_node("Sqrt", ["float_dim"], ["scale"]),
        ],
        inputs=[
            helper.make_value_info(
                "x",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[2, 3, 64]),
            ),
        ],
        outputs=[
            helper.make_value_info(
                "scale",
                helper.make_tensor_type_proto(TensorProto.FLOAT, shape=[]),
            ),
        ],
    )
    save(
        helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 16)]),
        "constant_fold_sqrt.onnx",
    )


if __name__ == "__main__":
    print("Generating constant fold test models:")
    constant_fold_cascade()
    constant_fold_chain()
    constant_fold_blocked()
    constant_fold_concat()
    constant_fold_cast()
    constant_fold_sqrt()
    print("Done.")