test_onnx_v2.py 21.6 KB
Newer Older
1
import os
2
3
4
5
6
from pathlib import Path
from tempfile import NamedTemporaryFile
from unittest import TestCase
from unittest.mock import patch

lewtun's avatar
lewtun committed
7
8
import pytest

9
from parameterized import parameterized
10
from transformers import AutoConfig, PreTrainedTokenizerBase, is_tf_available, is_torch_available
11
12
13
from transformers.onnx import (
    EXTERNAL_DATA_FORMAT_SIZE_LIMIT,
    OnnxConfig,
lewtun's avatar
lewtun committed
14
    OnnxConfigWithPast,
15
16
17
18
    ParameterFormat,
    export,
    validate_model_outputs,
)
19
20
21
22
23
from transformers.onnx.utils import (
    compute_effective_axis_dimension,
    compute_serialized_parameters_size,
    get_preprocessor,
)
24
from transformers.testing_utils import require_onnx, require_rjieba, require_tf, require_torch, require_vision, slow
25
26


27
if is_torch_available() or is_tf_available():
28
29
    from transformers.onnx.features import FeaturesManager

30
31
32
33
34
if is_torch_available():
    import torch

    from transformers.models.deberta import modeling_deberta

35
36
37
38
39
40
41

@require_onnx
class OnnxUtilsTestCaseV2(TestCase):
    """
    Cover all the utilities involved to export ONNX models
    """

42
43
44
45
46
47
48
49
50
    @require_torch
    @patch("transformers.onnx.convert.is_torch_onnx_dict_inputs_support_available", return_value=False)
    def test_ensure_pytorch_version_ge_1_8_0(self, mock_is_torch_onnx_dict_inputs_support_available):
        """
        Ensure we raise an Exception if the pytorch version is unsupported (< 1.8.0)
        """
        self.assertRaises(AssertionError, export, None, None, None, None, None)
        mock_is_torch_onnx_dict_inputs_support_available.assert_called()

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
    def test_compute_effective_axis_dimension(self):
        """
        When exporting ONNX model with dynamic axis (batch or sequence) we set batch_size and/or sequence_length = -1.
        We cannot generate an effective tensor with axis dim == -1, so we trick by using some "fixed" values
        (> 1 to avoid ONNX squeezing the axis).

        This test ensure we are correctly replacing generated batch / sequence tensor with axis > 1
        """

        # Dynamic axis (batch, no token added by the tokenizer)
        self.assertEqual(compute_effective_axis_dimension(-1, fixed_dimension=2, num_token_to_add=0), 2)

        # Static axis (batch, no token added by the tokenizer)
        self.assertEqual(compute_effective_axis_dimension(0, fixed_dimension=2, num_token_to_add=0), 2)

        # Dynamic axis (sequence, token added by the tokenizer 2 (no pair))
        self.assertEqual(compute_effective_axis_dimension(0, fixed_dimension=8, num_token_to_add=2), 6)
        self.assertEqual(compute_effective_axis_dimension(0, fixed_dimension=8, num_token_to_add=2), 6)

        # Dynamic axis (sequence, token added by the tokenizer 3 (pair))
        self.assertEqual(compute_effective_axis_dimension(0, fixed_dimension=8, num_token_to_add=3), 5)
        self.assertEqual(compute_effective_axis_dimension(0, fixed_dimension=8, num_token_to_add=3), 5)

    def test_compute_parameters_serialized_size(self):
        """
        This test ensures we compute a "correct" approximation of the underlying storage requirement (size) for all the
        parameters for the specified parameter's dtype.
        """
        self.assertEqual(compute_serialized_parameters_size(2, ParameterFormat.Float), 2 * ParameterFormat.Float.size)

    def test_flatten_output_collection_property(self):
        """
        This test ensures we correctly flatten nested collection such as the one we use when returning past_keys.
        past_keys = Tuple[Tuple]

        ONNX exporter will export nested collections as ${collection_name}.${level_idx_0}.${level_idx_1}...${idx_n}
        """
        self.assertEqual(
89
            OnnxConfig.flatten_output_collection_property("past_key", [[0], [1], [2]]),
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
            {
                "past_key.0": 0,
                "past_key.1": 1,
                "past_key.2": 2,
            },
        )


class OnnxConfigTestCaseV2(TestCase):
    """
    Cover the test for models default.

    Default means no specific features is being enabled on the model.
    """

    @patch.multiple(OnnxConfig, __abstractmethods__=set())
    def test_use_external_data_format(self):
        """
        External data format is required only if the serialized size of the parameters if bigger than 2Gb
        """
        TWO_GB_LIMIT = EXTERNAL_DATA_FORMAT_SIZE_LIMIT

        # No parameters
        self.assertFalse(OnnxConfig.use_external_data_format(0))

        # Some parameters
        self.assertFalse(OnnxConfig.use_external_data_format(1))

        # Almost 2Gb parameters
        self.assertFalse(OnnxConfig.use_external_data_format((TWO_GB_LIMIT - 1) // ParameterFormat.Float.size))

        # Exactly 2Gb parameters
        self.assertTrue(OnnxConfig.use_external_data_format(TWO_GB_LIMIT))

        # More than 2Gb parameters
        self.assertTrue(OnnxConfig.use_external_data_format((TWO_GB_LIMIT + 1) // ParameterFormat.Float.size))


class OnnxConfigWithPastTestCaseV2(TestCase):
    """
    Cover the tests for model which have use_cache feature (i.e. "with_past" for ONNX)
    """

133
134
135
136
137
138
    SUPPORTED_WITH_PAST_CONFIGS = {}
    # SUPPORTED_WITH_PAST_CONFIGS = {
    #     ("BART", BartConfig),
    #     ("GPT2", GPT2Config),
    #     # ("T5", T5Config)
    # }
139
140
141
142
143
144
145
146
147

    @patch.multiple(OnnxConfigWithPast, __abstractmethods__=set())
    def test_use_past(self):
        """
        Ensure the use_past variable is correctly being set
        """
        for name, config in OnnxConfigWithPastTestCaseV2.SUPPORTED_WITH_PAST_CONFIGS:
            with self.subTest(name):
                self.assertFalse(
148
149
                    OnnxConfigWithPast.from_model_config(config()).use_past,
                    "OnnxConfigWithPast.from_model_config() should not use_past",
150
151
152
                )

                self.assertTrue(
153
154
                    OnnxConfigWithPast.with_past(config()).use_past,
                    "OnnxConfigWithPast.from_model_config() should use_past",
155
156
157
158
159
160
161
162
163
164
                )

    @patch.multiple(OnnxConfigWithPast, __abstractmethods__=set())
    def test_values_override(self):
        """
        Ensure the use_past variable correctly set the `use_cache` value in model's configuration
        """
        for name, config in OnnxConfigWithPastTestCaseV2.SUPPORTED_WITH_PAST_CONFIGS:
            with self.subTest(name):
                # without past
165
                onnx_config_default = OnnxConfigWithPast.from_model_config(config())
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
                self.assertIsNotNone(onnx_config_default.values_override, "values_override should not be None")
                self.assertIn("use_cache", onnx_config_default.values_override, "use_cache should be present")
                self.assertFalse(
                    onnx_config_default.values_override["use_cache"], "use_cache should be False if not using past"
                )

                # with past
                onnx_config_default = OnnxConfigWithPast.with_past(config())
                self.assertIsNotNone(onnx_config_default.values_override, "values_override should not be None")
                self.assertIn("use_cache", onnx_config_default.values_override, "use_cache should be present")
                self.assertTrue(
                    onnx_config_default.values_override["use_cache"], "use_cache should be False if not using past"
                )


181
182
183
PYTORCH_EXPORT_MODELS = {
    ("albert", "hf-internal-testing/tiny-albert"),
    ("bert", "bert-base-cased"),
184
    ("big-bird", "google/bigbird-roberta-base"),
185
    ("ibert", "kssteven/ibert-roberta-base"),
186
    ("camembert", "camembert-base"),
187
    ("clip", "openai/clip-vit-base-patch32"),
188
    ("convbert", "YituTech/conv-bert-base"),
mrbean's avatar
mrbean committed
189
    ("codegen", "Salesforce/codegen-350M-multi"),
190
191
    ("deberta", "microsoft/deberta-base"),
    ("deberta-v2", "microsoft/deberta-v2-xlarge"),
192
    ("convnext", "facebook/convnext-tiny-224"),
regisss's avatar
regisss committed
193
    ("detr", "facebook/detr-resnet-50"),
194
    ("distilbert", "distilbert-base-cased"),
195
    ("electra", "google/electra-base-generator"),
196
    ("imagegpt", "openai/imagegpt-small"),
regisss's avatar
regisss committed
197
    ("resnet", "microsoft/resnet-50"),
198
    ("roberta", "roberta-base"),
199
    ("roformer", "junnyu/roformer_chinese_base"),
200
    ("squeezebert", "squeezebert/squeezebert-uncased"),
201
    ("mobilebert", "google/mobilebert-uncased"),
202
    ("mobilenet_v2", "google/mobilenet_v2_0.35_96"),
203
    ("mobilevit", "apple/mobilevit-small"),
Ritik Nandwal's avatar
Ritik Nandwal committed
204
    ("xlm", "xlm-clm-ende-1024"),
205
206
    ("xlm-roberta", "xlm-roberta-base"),
    ("layoutlm", "microsoft/layoutlm-base-uncased"),
207
    ("layoutlmv3", "microsoft/layoutlmv3-base"),
208
    ("groupvit", "nvidia/groupvit-gcc-yfcc"),
gcheron's avatar
gcheron committed
209
    ("levit", "facebook/levit-128S"),
210
    ("owlvit", "google/owlvit-base-patch32"),
lewtun's avatar
lewtun committed
211
    ("vit", "google/vit-base-patch16-224"),
212
    ("deit", "facebook/deit-small-patch16-224"),
Jim Rohrer's avatar
Jim Rohrer committed
213
    ("beit", "microsoft/beit-base-patch16-224"),
214
    ("data2vec-text", "facebook/data2vec-text-base"),
215
    ("data2vec-vision", "facebook/data2vec-vision-base"),
216
217
    ("perceiver", "deepmind/language-perceiver", ("masked-lm", "sequence-classification")),
    ("perceiver", "deepmind/vision-perceiver-conv", ("image-classification",)),
218
    ("longformer", "allenai/longformer-base-4096"),
NielsRogge's avatar
NielsRogge committed
219
    ("yolos", "hustvl/yolos-tiny"),
220
    ("segformer", "nvidia/segformer-b0-finetuned-ade-512-512"),
221
    ("swin", "microsoft/swin-tiny-patch4-window7-224"),
222
    ("whisper", "openai/whisper-tiny.en"),
223
224
}

225
226
227
228
PYTORCH_EXPORT_ENCODER_DECODER_MODELS = {
    ("vision-encoder-decoder", "nlpconnect/vit-gpt2-image-captioning"),
}

229
PYTORCH_EXPORT_WITH_PAST_MODELS = {
230
    ("bloom", "bigscience/bloom-560m"),
231
232
233
234
235
236
237
238
    ("gpt2", "gpt2"),
    ("gpt-neo", "EleutherAI/gpt-neo-125M"),
}

PYTORCH_EXPORT_SEQ2SEQ_WITH_PAST_MODELS = {
    ("bart", "facebook/bart-base"),
    ("mbart", "sshleifer/tiny-mbart"),
    ("t5", "t5-small"),
239
    ("marian", "Helsinki-NLP/opus-mt-en-de"),
240
    ("mt5", "google/mt5-base"),
241
    ("m2m-100", "facebook/m2m100_418M"),
242
243
    ("blenderbot-small", "facebook/blenderbot_small-90M"),
    ("blenderbot", "facebook/blenderbot-400M-distill"),
244
    ("bigbird-pegasus", "google/bigbird-pegasus-large-arxiv"),
245
    ("longt5", "google/long-t5-local-base"),
246
247
248
    # Disable for now as it causes fatal error `Floating point exception (core dumped)` and the subsequential tests are
    # not run.
    # ("longt5", "google/long-t5-tglobal-base"),
249
250
}

251
# TODO(lewtun): Include the same model types in `PYTORCH_EXPORT_MODELS` once TensorFlow has parity with the PyTorch model implementations.
252
253
254
TENSORFLOW_EXPORT_DEFAULT_MODELS = {
    ("albert", "hf-internal-testing/tiny-albert"),
    ("bert", "bert-base-cased"),
255
    ("camembert", "camembert-base"),
256
257
258
259
    ("distilbert", "distilbert-base-cased"),
    ("roberta", "roberta-base"),
}

260
261
# TODO(lewtun): Include the same model types in `PYTORCH_EXPORT_WITH_PAST_MODELS` once TensorFlow has parity with the PyTorch model implementations.
TENSORFLOW_EXPORT_WITH_PAST_MODELS = {}
262

263
264
# TODO(lewtun): Include the same model types in `PYTORCH_EXPORT_SEQ2SEQ_WITH_PAST_MODELS` once TensorFlow has parity with the PyTorch model implementations.
TENSORFLOW_EXPORT_SEQ2SEQ_WITH_PAST_MODELS = {}
265

266
267
268

def _get_models_to_test(export_models_list):
    models_to_test = []
269
    if is_torch_available() or is_tf_available():
270
271
272
273
274
275
276
277
278
        for name, model, *features in export_models_list:
            if features:
                feature_config_mapping = {
                    feature: FeaturesManager.get_config(name, feature) for _ in features for feature in _
                }
            else:
                feature_config_mapping = FeaturesManager.get_supported_features_for_model_type(name)

            for feature, onnx_config_class_constructor in feature_config_mapping.items():
279
280
281
282
283
                models_to_test.append((f"{name}_{feature}", name, model, feature, onnx_config_class_constructor))
        return sorted(models_to_test)
    else:
        # Returning some dummy test that should not be ever called because of the @require_torch / @require_tf
        # decorators.
284
285
        # The reason for not returning an empty list is because parameterized.expand complains when it's empty.
        return [("dummy", "dummy", "dummy", "dummy", OnnxConfig.from_model_config)]
286
287
288
289
290
291
292


class OnnxExportTestCaseV2(TestCase):
    """
    Integration tests ensuring supported models are correctly exported
    """

293
294
295
    def _onnx_export(
        self, test_name, name, model_name, feature, onnx_config_class_constructor, device="cpu", framework="pt"
    ):
296
297
        from transformers.onnx import export

298
        model_class = FeaturesManager.get_model_class_for_feature(feature, framework=framework)
lewtun's avatar
lewtun committed
299
        config = AutoConfig.from_pretrained(model_name)
300
        model = model_class.from_config(config)
Yih-Dar's avatar
Yih-Dar committed
301
302
303
304
305
306

        # Dynamic axes aren't supported for YOLO-like models. This means they cannot be exported to ONNX on CUDA devices.
        # See: https://github.com/ultralytics/yolov5/pull/8378
        if model.__class__.__name__.startswith("Yolos") and device != "cpu":
            return

307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
        # ONNX inference fails with the following name, feature, framework parameterizations
        # See: https://github.com/huggingface/transformers/issues/19357
        if (name, feature, framework) in {
            ("deberta-v2", "question-answering", "pt"),
            ("deberta-v2", "multiple-choice", "pt"),
            ("roformer", "multiple-choice", "pt"),
            ("groupvit", "default", "pt"),
            ("perceiver", "masked-lm", "pt"),
            ("perceiver", "sequence-classification", "pt"),
            ("perceiver", "image-classification", "pt"),
            ("bert", "multiple-choice", "tf"),
            ("camembert", "multiple-choice", "tf"),
            ("roberta", "multiple-choice", "tf"),
        }:
            return

323
        onnx_config = onnx_config_class_constructor(model.config)
324

lewtun's avatar
lewtun committed
325
        if is_torch_available():
326
            from transformers.utils import torch_version
lewtun's avatar
lewtun committed
327
328
329

            if torch_version < onnx_config.torch_onnx_minimum_version:
                pytest.skip(
Sylvain Gugger's avatar
Sylvain Gugger committed
330
331
                    "Skipping due to incompatible PyTorch version. Minimum required is"
                    f" {onnx_config.torch_onnx_minimum_version}, got: {torch_version}"
lewtun's avatar
lewtun committed
332
333
                )

334
335
336
337
338
        preprocessor = get_preprocessor(model_name)

        # Useful for causal lm models that do not use pad tokens.
        if isinstance(preprocessor, PreTrainedTokenizerBase) and not getattr(config, "pad_token_id", None):
            config.pad_token_id = preprocessor.eos_token_id
lewtun's avatar
lewtun committed
339

340
341
342
        with NamedTemporaryFile("w") as output:
            try:
                onnx_inputs, onnx_outputs = export(
343
                    preprocessor, model, onnx_config, onnx_config.default_onnx_opset, Path(output.name), device=device
344
345
346
                )
                validate_model_outputs(
                    onnx_config,
lewtun's avatar
lewtun committed
347
                    preprocessor,
348
349
350
351
352
353
354
                    model,
                    Path(output.name),
                    onnx_outputs,
                    onnx_config.atol_for_validation,
                )
            except (RuntimeError, ValueError) as e:
                self.fail(f"{name}, {feature} -> {e}")
355

356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
    def _onnx_export_encoder_decoder_models(
        self, test_name, name, model_name, feature, onnx_config_class_constructor, device="cpu"
    ):
        from transformers import AutoFeatureExtractor, AutoTokenizer
        from transformers.onnx import export

        model_class = FeaturesManager.get_model_class_for_feature(feature)
        config = AutoConfig.from_pretrained(model_name)
        model = model_class.from_config(config)

        onnx_config = onnx_config_class_constructor(model.config)

        if is_torch_available():
            from transformers.utils import torch_version

            if torch_version < onnx_config.torch_onnx_minimum_version:
                pytest.skip(
                    "Skipping due to incompatible PyTorch version. Minimum required is"
                    f" {onnx_config.torch_onnx_minimum_version}, got: {torch_version}"
                )

        encoder_model = model.get_encoder()
        decoder_model = model.get_decoder()

        encoder_onnx_config = onnx_config.get_encoder_config(encoder_model.config)
        decoder_onnx_config = onnx_config.get_decoder_config(encoder_model.config, decoder_model.config, feature)

        preprocessor = AutoFeatureExtractor.from_pretrained(model_name)

        onnx_opset = max(encoder_onnx_config.default_onnx_opset, decoder_onnx_config.default_onnx_opset)

        with NamedTemporaryFile("w") as encoder_output:
            onnx_inputs, onnx_outputs = export(
                preprocessor, encoder_model, encoder_onnx_config, onnx_opset, Path(encoder_output.name), device=device
            )
            validate_model_outputs(
                encoder_onnx_config,
                preprocessor,
                encoder_model,
                Path(encoder_output.name),
                onnx_outputs,
                encoder_onnx_config.atol_for_validation,
            )

        preprocessor = AutoTokenizer.from_pretrained(model_name)

        with NamedTemporaryFile("w") as decoder_output:
403
            _, onnx_outputs = export(
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
                preprocessor,
                decoder_model,
                decoder_onnx_config,
                onnx_config.default_onnx_opset,
                Path(decoder_output.name),
                device=device,
            )
            validate_model_outputs(
                decoder_onnx_config,
                preprocessor,
                decoder_model,
                Path(decoder_output.name),
                onnx_outputs,
                decoder_onnx_config.atol_for_validation,
            )

420
    @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_MODELS))
421
422
    @slow
    @require_torch
lewtun's avatar
lewtun committed
423
    @require_vision
424
    @require_rjieba
425
    def test_pytorch_export(self, test_name, name, model_name, feature, onnx_config_class_constructor):
426
        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor)
427

428
429
430
431
432
433
434
435
    @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_MODELS))
    @slow
    @require_torch
    @require_vision
    @require_rjieba
    def test_pytorch_export_on_cuda(self, test_name, name, model_name, feature, onnx_config_class_constructor):
        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor, device="cuda")

436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
    @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_ENCODER_DECODER_MODELS))
    @slow
    @require_torch
    @require_vision
    @require_rjieba
    def test_pytorch_export_encoder_decoder_models(
        self, test_name, name, model_name, feature, onnx_config_class_constructor
    ):
        self._onnx_export_encoder_decoder_models(test_name, name, model_name, feature, onnx_config_class_constructor)

    @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_ENCODER_DECODER_MODELS))
    @slow
    @require_torch
    @require_vision
    @require_rjieba
    def test_pytorch_export_encoder_decoder_models_on_cuda(
        self, test_name, name, model_name, feature, onnx_config_class_constructor
    ):
        self._onnx_export_encoder_decoder_models(
            test_name, name, model_name, feature, onnx_config_class_constructor, device="cuda"
        )

458
459
460
461
    @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_WITH_PAST_MODELS))
    @slow
    @require_torch
    def test_pytorch_export_with_past(self, test_name, name, model_name, feature, onnx_config_class_constructor):
462
        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor)
463

464
465
466
467
468
469
    @parameterized.expand(_get_models_to_test(PYTORCH_EXPORT_SEQ2SEQ_WITH_PAST_MODELS))
    @slow
    @require_torch
    def test_pytorch_export_seq2seq_with_past(
        self, test_name, name, model_name, feature, onnx_config_class_constructor
    ):
470
471
472
473
474
        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor)

    @parameterized.expand(_get_models_to_test(TENSORFLOW_EXPORT_DEFAULT_MODELS))
    @slow
    @require_tf
lewtun's avatar
lewtun committed
475
    @require_vision
476
    def test_tensorflow_export(self, test_name, name, model_name, feature, onnx_config_class_constructor):
477
        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor, framework="tf")
478

479
    @parameterized.expand(_get_models_to_test(TENSORFLOW_EXPORT_WITH_PAST_MODELS), skip_on_empty=True)
480
481
482
    @slow
    @require_tf
    def test_tensorflow_export_with_past(self, test_name, name, model_name, feature, onnx_config_class_constructor):
483
        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor, framework="tf")
484

485
    @parameterized.expand(_get_models_to_test(TENSORFLOW_EXPORT_SEQ2SEQ_WITH_PAST_MODELS), skip_on_empty=True)
486
487
488
489
490
    @slow
    @require_tf
    def test_tensorflow_export_seq2seq_with_past(
        self, test_name, name, model_name, feature, onnx_config_class_constructor
    ):
491
        self._onnx_export(test_name, name, model_name, feature, onnx_config_class_constructor, framework="tf")
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528


class StableDropoutTestCase(TestCase):
    """Tests export of StableDropout module."""

    @require_torch
    @pytest.mark.filterwarnings("ignore:.*Dropout.*:UserWarning:torch.onnx.*")  # torch.onnx is spammy.
    def test_training(self):
        """Tests export of StableDropout in training mode."""
        devnull = open(os.devnull, "wb")
        # drop_prob must be > 0 for the test to be meaningful
        sd = modeling_deberta.StableDropout(0.1)
        # Avoid warnings in training mode
        do_constant_folding = False
        # Dropout is a no-op in inference mode
        training = torch.onnx.TrainingMode.PRESERVE
        input = (torch.randn(2, 2),)

        torch.onnx.export(
            sd,
            input,
            devnull,
            opset_version=12,  # Minimum supported
            do_constant_folding=do_constant_folding,
            training=training,
        )

        # Expected to fail with opset_version < 12
        with self.assertRaises(Exception):
            torch.onnx.export(
                sd,
                input,
                devnull,
                opset_version=11,
                do_constant_folding=do_constant_folding,
                training=training,
            )