test_modeling_common.py 144 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Aymeric Augustin's avatar
Aymeric Augustin committed
15

16
import copy
17
import gc
18
import inspect
19
import json
20
import os
21
import os.path
22
import pickle
Aymeric Augustin's avatar
Aymeric Augustin committed
23
import random
24
import sys
25
import tempfile
thomwolf's avatar
thomwolf committed
26
import unittest
27
import unittest.mock as mock
28
import warnings
29
from pathlib import Path
NielsRogge's avatar
NielsRogge committed
30
from typing import Dict, List, Tuple
thomwolf's avatar
thomwolf committed
31

32
33
34
import numpy as np

import transformers
35
from huggingface_hub import HfFolder, delete_repo, set_access_token
36
from huggingface_hub.file_download import http_get
Sylvain Gugger's avatar
Sylvain Gugger committed
37
from requests.exceptions import HTTPError
38
39
40
41
42
43
44
45
from transformers import (
    AutoConfig,
    AutoModel,
    AutoModelForSequenceClassification,
    PretrainedConfig,
    is_torch_available,
    logging,
)
46
from transformers.models.auto import get_values
Sylvain Gugger's avatar
Sylvain Gugger committed
47
from transformers.testing_utils import (
48
    TOKEN,
Sylvain Gugger's avatar
Sylvain Gugger committed
49
50
    USER,
    CaptureLogger,
51
    TestCasePlus,
52
    is_flaky,
53
54
    is_pt_flax_cross_test,
    is_pt_tf_cross_test,
Sylvain Gugger's avatar
Sylvain Gugger committed
55
    is_staging_test,
56
    require_accelerate,
57
    require_safetensors,
Sylvain Gugger's avatar
Sylvain Gugger committed
58
    require_torch,
59
    require_torch_gpu,
Sylvain Gugger's avatar
Sylvain Gugger committed
60
    require_torch_multi_gpu,
61
    require_usr_bin_time,
Sylvain Gugger's avatar
Sylvain Gugger committed
62
63
64
    slow,
    torch_device,
)
65
from transformers.utils import (
66
67
    SAFE_WEIGHTS_INDEX_NAME,
    SAFE_WEIGHTS_NAME,
68
69
    WEIGHTS_INDEX_NAME,
    WEIGHTS_NAME,
70
    is_accelerate_available,
71
72
73
74
75
    is_flax_available,
    is_tf_available,
    is_torch_fx_available,
)
from transformers.utils.generic import ModelOutput
76

Aymeric Augustin's avatar
Aymeric Augustin committed
77

78
79
sys.path.append(str(Path(__file__).parent.parent / "utils"))

80
from test_module.custom_configuration import CustomConfig, NoSuperInitConfig  # noqa E402
81
82


83
84
85
86
if is_accelerate_available():
    from accelerate.utils import compute_module_sizes


87
if is_torch_available():
88
    import torch
89
    from torch import nn
thomwolf's avatar
thomwolf committed
90

91
    from test_module.custom_modeling import CustomModel, NoSuperInitModel
92
    from transformers import (
93
        BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
94
        MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING,
95
        MODEL_FOR_AUDIO_XVECTOR_MAPPING,
96
        MODEL_FOR_BACKBONE_MAPPING,
NielsRogge's avatar
NielsRogge committed
97
        MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
98
        MODEL_FOR_CAUSAL_LM_MAPPING,
99
        MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING,
100
        MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
NielsRogge's avatar
NielsRogge committed
101
        MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING,
102
        MODEL_FOR_MASKED_LM_MAPPING,
103
        MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
104
        MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
105
        MODEL_FOR_QUESTION_ANSWERING_MAPPING,
NielsRogge's avatar
NielsRogge committed
106
        MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING,
107
108
109
        MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
        MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
        MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
NielsRogge's avatar
NielsRogge committed
110
        MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING,
111
        MODEL_MAPPING,
112
        AdaptiveEmbedding,
113
114
        AutoModelForCausalLM,
        AutoTokenizer,
115
116
117
        BertConfig,
        BertModel,
        PreTrainedModel,
118
        T5Config,
119
        T5ForConditionalGeneration,
120
    )
Sylvain Gugger's avatar
Sylvain Gugger committed
121
    from transformers.modeling_utils import shard_checkpoint
thomwolf's avatar
thomwolf committed
122

Sylvain Gugger's avatar
Sylvain Gugger committed
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
    # Fake pretrained models for tests
    class BaseModel(PreTrainedModel):
        config_class = PretrainedConfig

        def __init__(self, config):
            super().__init__(config)
            self.linear = nn.Linear(4, 5)
            self.linear_2 = nn.Linear(5, 6)

        def forward(self, x):
            return self.linear_2(self.linear(x))

    class ModelWithHead(PreTrainedModel):
        base_model_prefix = "base"
        config_class = PretrainedConfig

        def _init_weights(self, module):
            pass

        def __init__(self, config):
            super().__init__(config)
            self.base = BaseModel(config)
            # linear is a common name between Base and Head on purpose.
            self.linear = nn.Linear(6, 3)
            self.linear2 = nn.Linear(3, 5)

        def forward(self, x):
            return self.linear2(self.linear(self.base(x)))


153
154
155
if is_tf_available():
    import tensorflow as tf

156
157
158
159
160
161
162
if is_flax_available():
    import jax.numpy as jnp
    from transformers.modeling_flax_pytorch_utils import (
        convert_pytorch_state_dict_to_flax,
        load_flax_weights_in_pytorch_model,
    )

163
if is_torch_fx_available():
164
    from transformers.utils.fx import symbolic_trace
165

166

167
168
169
def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
170
        if "_range" in key or "_std" in key or "initializer_factor" in key or "layer_scale" in key:
Lysandre Debut's avatar
Lysandre Debut committed
171
            setattr(configs_no_init, key, 1e-10)
172
173
    return configs_no_init

thomwolf's avatar
thomwolf committed
174

175
TINY_T5 = "patrickvonplaten/t5-tiny-random"
176
TINY_BERT_FOR_TOKEN_CLASSIFICATION = "hf-internal-testing/tiny-bert-for-token-classification"
177
178


179
180
181
182
183
@require_torch
class ModelTesterMixin:

    model_tester = None
    all_model_classes = ()
184
    all_generative_model_classes = ()
185
    fx_compatible = False
Patrick von Platen's avatar
Patrick von Platen committed
186
187
188
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
189
    test_resize_position_embeddings = False
Patrick von Platen's avatar
Patrick von Platen committed
190
    test_head_masking = True
191
    test_mismatched_shapes = True
192
    test_missing_keys = True
193
    test_model_parallel = False
194
    is_encoder_decoder = False
195
    has_attentions = True
196
    model_split_percents = [0.5, 0.7, 0.9]
197

198
199
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
200
        if model_class in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
201
            inputs_dict = {
202
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
203
                if isinstance(v, torch.Tensor) and v.ndim > 1
Sylvain Gugger's avatar
Sylvain Gugger committed
204
                else v
205
206
                for k, v in inputs_dict.items()
            }
207
208
        elif model_class in get_values(MODEL_FOR_AUDIO_XVECTOR_MAPPING):
            inputs_dict.pop("attention_mask")
209
210

        if return_labels:
211
            if model_class in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
212
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
213
214
215
216
            elif model_class in [
                *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING),
                *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING),
            ]:
217
218
219
220
221
222
                inputs_dict["start_positions"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
                inputs_dict["end_positions"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
223
            elif model_class in [
224
225
226
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING),
                *get_values(MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
NielsRogge's avatar
NielsRogge committed
227
                *get_values(MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING),
228
                *get_values(MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING),
229
            ]:
230
231
232
233
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
            elif model_class in [
234
235
                *get_values(MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING),
                *get_values(MODEL_FOR_CAUSAL_LM_MAPPING),
NielsRogge's avatar
NielsRogge committed
236
                *get_values(MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING),
237
238
                *get_values(MODEL_FOR_MASKED_LM_MAPPING),
                *get_values(MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING),
239
240
241
242
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
NielsRogge's avatar
NielsRogge committed
243
244
245
246
247
            elif model_class in get_values(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING):
                num_patches = self.model_tester.image_size // self.model_tester.patch_size
                inputs_dict["bool_masked_pos"] = torch.zeros(
                    (self.model_tester.batch_size, num_patches**2), dtype=torch.long, device=torch_device
                )
NielsRogge's avatar
NielsRogge committed
248
249
250
251
252
            elif model_class in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING):
                batch_size, num_channels, height, width = inputs_dict["pixel_values"].shape
                inputs_dict["labels"] = torch.zeros(
                    [self.model_tester.batch_size, height, width], device=torch_device
                ).long()
253

254
255
        return inputs_dict

Patrick von Platen's avatar
Patrick von Platen committed
256
    def test_save_load(self):
257
258
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

259
260
261
262
263
264
265
266
267
268
        def check_save_load(out1, out2):
            # make sure we don't have nans
            out_2 = out2.cpu().numpy()
            out_2[np.isnan(out_2)] = 0

            out_1 = out1.cpu().numpy()
            out_1[np.isnan(out_1)] = 0
            max_diff = np.amax(np.abs(out_1 - out_2))
            self.assertLessEqual(max_diff, 1e-5)

269
270
271
272
273
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
274
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
275

276
            with tempfile.TemporaryDirectory() as tmpdirname:
277
278
                model.save_pretrained(tmpdirname)
                model = model_class.from_pretrained(tmpdirname)
279
                model.to(torch_device)
280
                with torch.no_grad():
281
                    second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
thomwolf's avatar
thomwolf committed
282

283
284
285
286
287
            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_save_load(tensor1, tensor2)
            else:
                check_save_load(first, second)
288

289
    def test_save_load_keys_to_ignore_on_save(self):
290
291
292
293
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
294
295
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
296
297
298
                continue

            # check the keys are in the original state_dict
299
            for k in _keys_to_ignore_on_save:
300
                self.assertIn(k, model.state_dict().keys(), "\n".join(model.state_dict().keys()))
301
302
303
304
305
306

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                output_model_file = os.path.join(tmpdirname, WEIGHTS_NAME)
                state_dict_saved = torch.load(output_model_file)
307
                for k in _keys_to_ignore_on_save:
308
                    self.assertNotIn(k, state_dict_saved.keys(), "\n".join(state_dict_saved.keys()))
309

Sylvain Gugger's avatar
Sylvain Gugger committed
310
311
312
                # Test we can load the state dict in the model, necessary for the checkpointing API in Trainer.
                load_result = model.load_state_dict(state_dict_saved, strict=False)
                self.assertTrue(
313
314
                    len(load_result.missing_keys) == 0
                    or set(load_result.missing_keys) == set(model._keys_to_ignore_on_save)
Sylvain Gugger's avatar
Sylvain Gugger committed
315
316
317
                )
                self.assertTrue(len(load_result.unexpected_keys) == 0)

318
319
320
321
322
323
324
325
326
327
328
    def test_gradient_checkpointing_backward_compatibility(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if not model_class.supports_gradient_checkpointing:
                continue

            config.gradient_checkpointing = True
            model = model_class(config)
            self.assertTrue(model.is_gradient_checkpointing)

329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
    def test_gradient_checkpointing_enable_disable(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if not model_class.supports_gradient_checkpointing:
                continue

            # at init model should have gradient checkpointing disabled
            model = model_class(config)
            self.assertFalse(model.is_gradient_checkpointing)

            # check enable works
            model.gradient_checkpointing_enable()
            self.assertTrue(model.is_gradient_checkpointing)

            # check disable works
            model.gradient_checkpointing_disable()
            self.assertFalse(model.is_gradient_checkpointing)

348
349
350
351
352
353
    def _mock_init_weights(self, module):
        if hasattr(module, "weight") and module.weight is not None:
            module.weight.data.fill_(3)
        if hasattr(module, "bias") and module.bias is not None:
            module.bias.data.fill_(3)

354
    @is_flaky()
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
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
    def test_save_load_fast_init_from_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        base_class = MODEL_MAPPING[config.__class__]

        if isinstance(base_class, tuple):
            base_class = base_class[0]

        for model_class in self.all_model_classes:
            if model_class == base_class:
                continue

            # make a copy of model class to not break future tests
            # from https://stackoverflow.com/questions/9541025/how-to-copy-a-python-class
            class CopyClass(model_class):
                pass

            model_class_copy = CopyClass

            # make sure that all keys are expected for test
            model_class_copy._keys_to_ignore_on_load_missing = []

            # make init deterministic, but make sure that
            # non-initialized weights throw errors nevertheless
            model_class_copy._init_weights = self._mock_init_weights

            model = base_class(config)
            state_dict = model.state_dict()

            # this will often delete a single weight of a multi-weight module
            # to test an edge case
            random_key_to_del = random.choice(list(state_dict.keys()))
            del state_dict[random_key_to_del]

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                torch.save(state_dict, os.path.join(tmpdirname, "pytorch_model.bin"))

                model_fast_init = model_class_copy.from_pretrained(tmpdirname)
                model_slow_init = model_class_copy.from_pretrained(tmpdirname, _fast_init=False)

                for key in model_fast_init.state_dict().keys():
                    max_diff = (model_slow_init.state_dict()[key] - model_fast_init.state_dict()[key]).sum().item()
                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")

    def test_save_load_fast_init_to_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        base_class = MODEL_MAPPING[config.__class__]

        if isinstance(base_class, tuple):
            base_class = base_class[0]

        for model_class in self.all_model_classes:

            if model_class == base_class:
                continue

            # make a copy of model class to not break future tests
            # from https://stackoverflow.com/questions/9541025/how-to-copy-a-python-class
            class CopyClass(base_class):
                pass

            base_class_copy = CopyClass

            # make sure that all keys are expected for test
            base_class_copy._keys_to_ignore_on_load_missing = []

            # make init deterministic, but make sure that
            # non-initialized weights throw errors nevertheless
            base_class_copy._init_weights = self._mock_init_weights

            model = model_class(config)
            state_dict = model.state_dict()

            # this will often delete a single weight of a multi-weight module
            # to test an edge case
            random_key_to_del = random.choice(list(state_dict.keys()))
            del state_dict[random_key_to_del]

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.config.save_pretrained(tmpdirname)
                torch.save(state_dict, os.path.join(tmpdirname, "pytorch_model.bin"))

                model_fast_init = base_class_copy.from_pretrained(tmpdirname)
                model_slow_init = base_class_copy.from_pretrained(tmpdirname, _fast_init=False)

                for key in model_fast_init.state_dict().keys():
443
444
445
                    max_diff = torch.max(
                        torch.abs(model_slow_init.state_dict()[key] - model_fast_init.state_dict()[key])
                    ).item()
446
447
                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")

Patrick von Platen's avatar
Patrick von Platen committed
448
    def test_initialization(self):
449
450
451
452
453
454
455
456
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        configs_no_init = _config_zero_init(config)
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            for name, param in model.named_parameters():
                if param.requires_grad:
                    self.assertIn(
Lysandre Debut's avatar
Lysandre Debut committed
457
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
458
                        [0.0, 1.0],
459
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
460
                    )
thomwolf's avatar
thomwolf committed
461

Patrick von Platen's avatar
Patrick von Platen committed
462
    def test_determinism(self):
463
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
464
465
466
467
468
469
470
471
472

        def check_determinism(first, second):
            out_1 = first.cpu().numpy()
            out_2 = second.cpu().numpy()
            out_1 = out_1[~np.isnan(out_1)]
            out_2 = out_2[~np.isnan(out_2)]
            max_diff = np.amax(np.abs(out_1 - out_2))
            self.assertLessEqual(max_diff, 1e-5)

473
474
475
476
477
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
478
479
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
                second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
Weizhen's avatar
Weizhen committed
480

481
482
483
484
485
            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_determinism(tensor1, tensor2)
            else:
                check_determinism(first, second)
486

487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
    def test_forward_signature(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            signature = inspect.signature(model.forward)
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
            arg_names = [*signature.parameters.keys()]

            if model.config.is_encoder_decoder:
                expected_arg_names = [
                    "input_ids",
                    "attention_mask",
                    "decoder_input_ids",
                    "decoder_attention_mask",
                ]
503
                expected_arg_names.extend(
504
505
                    ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
                    if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
506
507
508
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
509
510
511
512
            else:
                expected_arg_names = ["input_ids"]
                self.assertListEqual(arg_names[:1], expected_arg_names)

513
514
515
516
517
    def test_training(self):
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
518
519
520
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.return_dict = True

521
522
523
524
            if model_class in [
                *get_values(MODEL_MAPPING),
                *get_values(MODEL_FOR_BACKBONE_MAPPING),
            ]:
525
                continue
526

527
528
529
530
531
532
533
534
            model = model_class(config)
            model.to(torch_device)
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

    def test_training_gradient_checkpointing(self):
535
        if not self.model_tester.is_training:
536
537
538
            return

        for model_class in self.all_model_classes:
539
540
541
542
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.use_cache = False
            config.return_dict = True

543
544
545
546
            if (
                model_class in [*get_values(MODEL_MAPPING), *get_values(MODEL_FOR_BACKBONE_MAPPING)]
                or not model_class.supports_gradient_checkpointing
            ):
547
548
549
                continue
            model = model_class(config)
            model.to(torch_device)
550
            model.gradient_checkpointing_enable()
551
552
553
554
555
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

Patrick von Platen's avatar
Patrick von Platen committed
556
    def test_attention_outputs(self):
557
558
559
        if not self.has_attentions:
            self.skipTest(reason="Model does not output attentions")

560
561
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True
562

563
564
565
566
567
568
569
570
571
572
573
574
        seq_len = getattr(self.model_tester, "seq_length", None)
        decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
        decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
        chunk_length = getattr(self.model_tester, "chunk_length", None)
        if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
            encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes

        for model_class in self.all_model_classes:
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = False
575
            config.return_dict = True
576
577
578
579
580
581
582
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
583

584
585
586
587
588
589
590
591
592
593
            # check that output_attentions also work using config
            del inputs_dict["output_attentions"]
            config.output_attentions = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
thomwolf's avatar
thomwolf committed
594

595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
            if chunk_length is not None:
                self.assertListEqual(
                    list(attentions[0].shape[-4:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
                )
            else:
                self.assertListEqual(
                    list(attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                )
            out_len = len(outputs)

            if self.is_encoder_decoder:
                correct_outlen = 5

                # loss is at first position
                if "labels" in inputs_dict:
                    correct_outlen += 1  # loss is added to beginning
                # Question Answering model returns start_logits and end_logits
614
615
616
617
                if model_class in [
                    *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING),
                    *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING),
                ]:
618
619
620
621
622
623
624
625
626
627
628
629
630
631
                    correct_outlen += 1  # start_logits and end_logits instead of only 1 output
                if "past_key_values" in outputs:
                    correct_outlen += 1  # past_key_values have been returned

                self.assertEqual(out_len, correct_outlen)

                # decoder attentions
                decoder_attentions = outputs.decoder_attentions
                self.assertIsInstance(decoder_attentions, (list, tuple))
                self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
                self.assertListEqual(
                    list(decoder_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
                )
632

633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
                # cross attentions
                cross_attentions = outputs.cross_attentions
                self.assertIsInstance(cross_attentions, (list, tuple))
                self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
                self.assertListEqual(
                    list(cross_attentions[0].shape[-3:]),
                    [
                        self.model_tester.num_attention_heads,
                        decoder_seq_length,
                        encoder_key_length,
                    ],
                )

            # Check attention is always last and order is fine
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            if hasattr(self.model_tester, "num_hidden_states_types"):
                added_hidden_states = self.model_tester.num_hidden_states_types
            elif self.is_encoder_decoder:
                added_hidden_states = 2
            else:
                added_hidden_states = 1
            self.assertEqual(out_len + added_hidden_states, len(outputs))

            self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions

            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
            if chunk_length is not None:
                self.assertListEqual(
                    list(self_attentions[0].shape[-4:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
                )
            else:
                self.assertListEqual(
                    list(self_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                )
thomwolf's avatar
thomwolf committed
676

677
    @slow
678
    def test_torchscript_simple(self):
679
680
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        self._create_and_check_torchscript(config, inputs_dict)
thomwolf's avatar
thomwolf committed
681

682
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
683
    def test_torchscript_output_attentions(self):
684
685
686
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_attentions = True
        self._create_and_check_torchscript(config, inputs_dict)
thomwolf's avatar
thomwolf committed
687

688
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
689
    def test_torchscript_output_hidden_state(self):
690
691
692
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
        self._create_and_check_torchscript(config, inputs_dict)
thomwolf's avatar
thomwolf committed
693

694
695
696
697
698
699
700
    # This is copied from `torch/testing/_internal/jit_utils.py::clear_class_registry`
    def clear_torch_jit_class_registry(self):

        torch._C._jit_clear_class_registry()
        torch.jit._recursive.concrete_type_store = torch.jit._recursive.ConcreteTypeStore()
        torch.jit._state._clear_class_state()

701
    def _create_and_check_torchscript(self, config, inputs_dict):
Patrick von Platen's avatar
Patrick von Platen committed
702
        if not self.test_torchscript:
703
            return
704

705
706
707
708
709
710
        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        configs_no_init.torchscript = True
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
711
            inputs = self._prepare_for_class(inputs_dict, model_class)
thomwolf's avatar
thomwolf committed
712

713
714
            main_input_name = model_class.main_input_name

715
            try:
716
                if model.config.is_encoder_decoder:
717
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
718
                    main_input = inputs[main_input_name]
719
720
721
                    attention_mask = inputs["attention_mask"]
                    decoder_input_ids = inputs["decoder_input_ids"]
                    decoder_attention_mask = inputs["decoder_attention_mask"]
722
                    model(main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
723
                    traced_model = torch.jit.trace(
724
                        model, (main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
725
                    )
726
727
728
729
                elif "bbox" in inputs and "image" in inputs:  # LayoutLMv2 requires additional inputs
                    input_ids = inputs["input_ids"]
                    bbox = inputs["bbox"]
                    image = inputs["image"].tensor
730
                    model(input_ids, bbox, image)
731
732
733
                    traced_model = torch.jit.trace(
                        model, (input_ids, bbox, image), check_trace=False
                    )  # when traced model is checked, an error is produced due to name mangling
734
                else:
735
                    main_input = inputs[main_input_name]
736
                    model(main_input)
737
                    traced_model = torch.jit.trace(model, main_input)
738
739
            except RuntimeError:
                self.fail("Couldn't trace module.")
thomwolf's avatar
thomwolf committed
740

741
            with tempfile.TemporaryDirectory() as tmp_dir_name:
742
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
thomwolf's avatar
thomwolf committed
743

744
                try:
745
                    torch.jit.save(traced_model, pt_file_name)
746
747
                except Exception:
                    self.fail("Couldn't save module.")
thomwolf's avatar
thomwolf committed
748

749
750
751
752
                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")
LysandreJik's avatar
LysandreJik committed
753

754
755
            model.to(torch_device)
            model.eval()
thomwolf's avatar
thomwolf committed
756

757
758
            loaded_model.to(torch_device)
            loaded_model.eval()
thomwolf's avatar
thomwolf committed
759

760
761
762
            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

763
764
765
766
767
768
769
770
771
            non_persistent_buffers = {}
            for key in loaded_model_state_dict.keys():
                if key not in model_state_dict.keys():
                    non_persistent_buffers[key] = loaded_model_state_dict[key]

            loaded_model_state_dict = {
                key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
            }

772
            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
thomwolf's avatar
thomwolf committed
773

774
775
776
777
778
779
780
781
782
783
784
            model_buffers = list(model.buffers())
            for non_persistent_buffer in non_persistent_buffers.values():
                found_buffer = False
                for i, model_buffer in enumerate(model_buffers):
                    if torch.equal(non_persistent_buffer, model_buffer):
                        found_buffer = True
                        break

                self.assertTrue(found_buffer)
                model_buffers.pop(i)

785
            models_equal = True
786
            for layer_name, p1 in model_state_dict.items():
787
788
789
790
                if layer_name in loaded_model_state_dict:
                    p2 = loaded_model_state_dict[layer_name]
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False
thomwolf's avatar
thomwolf committed
791

792
            self.assertTrue(models_equal)
thomwolf's avatar
thomwolf committed
793

794
795
796
797
            # Avoid memory leak. Without this, each call increase RAM usage by ~20MB.
            # (Even with this call, there are still memory leak by ~0.04MB)
            self.clear_torch_jit_class_registry()

798
799
800
801
802
803
804
805
    def test_torch_fx(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        self._create_and_check_torch_fx_tracing(config, inputs_dict)

    def test_torch_fx_output_loss(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        self._create_and_check_torch_fx_tracing(config, inputs_dict, output_loss=True)

806
807
    def _create_and_check_torch_fx_tracing(self, config, inputs_dict, output_loss=False):
        if not is_torch_fx_available() or not self.fx_compatible:
808
809
810
811
812
            return

        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        configs_no_init.return_dict = False

813
        for model_class in self.all_model_classes:
814
815
816
817
818
819
820
821
822
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=output_loss)

            try:
                if model.config.is_encoder_decoder:
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
                    labels = inputs.get("labels", None)
823
824
825
                    input_names = [
                        "attention_mask",
                        "decoder_attention_mask",
826
                        "decoder_input_ids",
827
                        "input_features",
828
829
                        "input_ids",
                        "input_values",
830
                    ]
831
832
                    if labels is not None:
                        input_names.append("labels")
833

834
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
835
                    input_names = list(filtered_inputs.keys())
836

837
                    model_output = model(**filtered_inputs)
838

839
                    traced_model = symbolic_trace(model, input_names)
840
                    traced_output = traced_model(**filtered_inputs)
841
                else:
842
843
844
845
                    input_names = [
                        "attention_mask",
                        "bbox",
                        "input_features",
846
847
848
849
850
851
                        "input_ids",
                        "input_values",
                        "pixel_values",
                        "token_type_ids",
                        "visual_feats",
                        "visual_pos",
852
                    ]
853

854
                    labels = inputs.get("labels", None)
855
856
                    start_positions = inputs.get("start_positions", None)
                    end_positions = inputs.get("end_positions", None)
857
858
                    if labels is not None:
                        input_names.append("labels")
859
860
861
862
                    if start_positions is not None:
                        input_names.append("start_positions")
                    if end_positions is not None:
                        input_names.append("end_positions")
863

864
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
865
                    input_names = list(filtered_inputs.keys())
866

867
868
                    if isinstance(model, tuple(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING.values())) and (
                        not hasattr(model.config, "problem_type") or model.config.problem_type is None
869
870
871
                    ):
                        model.config.problem_type = "single_label_classification"

872
                    traced_model = symbolic_trace(model, input_names)
873
                    traced_output = traced_model(**filtered_inputs)
874
                    model_output = model(**filtered_inputs)
875

876
            except Exception as e:
877
                self.fail(f"Couldn't trace module: {e}")
878

879
880
881
882
883
884
885
886
887
888
889
890
891
            def flatten_output(output):
                flatten = []
                for x in output:
                    if isinstance(x, (tuple, list)):
                        flatten += flatten_output(x)
                    elif not isinstance(x, torch.Tensor):
                        continue
                    else:
                        flatten.append(x)
                return flatten

            model_output = flatten_output(model_output)
            traced_output = flatten_output(traced_output)
892
            num_outputs = len(model_output)
893
894
895
896
897
898

            for i in range(num_outputs):
                self.assertTrue(
                    torch.allclose(model_output[i], traced_output[i]),
                    f"traced {i}th output doesn't match model {i}th output for {model_class}",
                )
899

900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
            # Test that the model can be serialized and restored properly
            with tempfile.TemporaryDirectory() as tmp_dir_name:
                pkl_file_name = os.path.join(tmp_dir_name, "model.pkl")
                try:
                    with open(pkl_file_name, "wb") as f:
                        pickle.dump(traced_model, f)
                    with open(pkl_file_name, "rb") as f:
                        loaded = pickle.load(f)
                except Exception as e:
                    self.fail(f"Couldn't serialize / deserialize the traced model: {e}")

                loaded_output = loaded(**filtered_inputs)
                loaded_output = flatten_output(loaded_output)

                for i in range(num_outputs):
                    self.assertTrue(
                        torch.allclose(model_output[i], loaded_output[i]),
                        f"serialized model {i}th output doesn't match model {i}th output for {model_class}",
                    )

920
921
922
923
            # Avoid memory leak. Without this, each call increase RAM usage by ~20MB.
            # (Even with this call, there are still memory leak by ~0.04MB)
            self.clear_torch_jit_class_registry()

Patrick von Platen's avatar
Patrick von Platen committed
924
925
    def test_headmasking(self):
        if not self.test_head_masking:
926
            return
927

928
929
930
        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()
LysandreJik's avatar
LysandreJik committed
931

932
        inputs_dict["output_attentions"] = True
933
934
935
936
937
938
        config.output_hidden_states = True
        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
LysandreJik's avatar
LysandreJik committed
939

940
941
942
            # Prepare head_mask
            # Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior)
            head_mask = torch.ones(
Lysandre's avatar
Lysandre committed
943
944
945
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
946
947
948
949
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
950
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
951
            inputs["head_mask"] = head_mask
952
953
954
955
956
            if model.config.is_encoder_decoder:
                signature = inspect.signature(model.forward)
                arg_names = [*signature.parameters.keys()]
                if "decoder_head_mask" in arg_names:  # necessary diferentiation because of T5 model
                    inputs["decoder_head_mask"] = head_mask
957
958
                if "cross_attn_head_mask" in arg_names:
                    inputs["cross_attn_head_mask"] = head_mask
959
            outputs = model(**inputs, return_dict=True)
960
961
962
963
964
965
966
967
968

            # Test that we can get a gradient back for importance score computation
            output = sum(t.sum() for t in outputs[0])
            output = output.sum()
            output.backward()
            multihead_outputs = head_mask.grad

            self.assertIsNotNone(multihead_outputs)
            self.assertEqual(len(multihead_outputs), self.model_tester.num_hidden_layers)
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989

            def check_attentions_validity(attentions):
                # Remove Nan
                for t in attentions:
                    self.assertLess(
                        torch.sum(torch.isnan(t)), t.numel() / 4
                    )  # Check we don't have more than 25% nans (arbitrary)
                attentions = [
                    t.masked_fill(torch.isnan(t), 0.0) for t in attentions
                ]  # remove them (the test is less complete)

                self.assertAlmostEqual(attentions[0][..., 0, :, :].flatten().sum().item(), 0.0)
                self.assertNotEqual(attentions[0][..., -1, :, :].flatten().sum().item(), 0.0)
                if len(attentions) > 2:  # encoder-decoder models have only 2 layers in each module
                    self.assertNotEqual(attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)
                self.assertAlmostEqual(attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0)
                self.assertNotEqual(attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)

            if model.config.is_encoder_decoder:
                check_attentions_validity(outputs.encoder_attentions)
                check_attentions_validity(outputs.decoder_attentions)
990
                check_attentions_validity(outputs.cross_attentions)
991
992
            else:
                check_attentions_validity(outputs.attentions)
993

Patrick von Platen's avatar
Patrick von Platen committed
994
995
    def test_head_pruning(self):
        if not self.test_pruning:
996
997
998
            return

        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
999
1000
1001
1002
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1003

1004
1005
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1006

1007
            inputs_dict["output_attentions"] = True
1008
1009
1010
1011
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1012
1013
1014
1015
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1016
1017
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
1018
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1019

1020
            attentions = outputs[-1]
1021

1022
1023
1024
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
LysandreJik's avatar
LysandreJik committed
1025

Patrick von Platen's avatar
Patrick von Platen committed
1026
1027
    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
1028
            return
LysandreJik's avatar
LysandreJik committed
1029

1030
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1031
1032
1033
1034
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1035
1036
1037

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1038

1039
            inputs_dict["output_attentions"] = True
1040
1041
1042
1043
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1044
1045
1046
1047
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1048
            model.prune_heads(heads_to_prune)
1049

1050
            with tempfile.TemporaryDirectory() as temp_dir_name:
1051
1052
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1053
                model.to(torch_device)
1054

1055
            with torch.no_grad():
1056
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1057
1058
1059
1060
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1061

Patrick von Platen's avatar
Patrick von Platen committed
1062
1063
    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
1064
            return
1065

1066
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1067
1068
1069
1070
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1071

1072
1073
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1074

1075
            inputs_dict["output_attentions"] = True
1076
            config.output_hidden_states = False
1077

1078
1079
1080
1081
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1082
            config.pruned_heads = heads_to_prune
1083

1084
1085
1086
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1087

1088
            with torch.no_grad():
1089
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1090
            attentions = outputs[-1]
1091

1092
1093
1094
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1095

Patrick von Platen's avatar
Patrick von Platen committed
1096
1097
    def test_head_pruning_integration(self):
        if not self.test_pruning:
1098
            return
1099

1100
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1101
1102
1103
1104
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1105

1106
1107
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1108

1109
            inputs_dict["output_attentions"] = True
1110
            config.output_hidden_states = False
1111

1112
1113
            heads_to_prune = {0: [0], 1: [1, 2]}
            config.pruned_heads = heads_to_prune
1114

1115
1116
1117
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1118

1119
            with torch.no_grad():
1120
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1121
            attentions = outputs[-1]
1122

1123
1124
1125
1126
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
thomwolf's avatar
thomwolf committed
1127

1128
            with tempfile.TemporaryDirectory() as temp_dir_name:
1129
1130
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1131
                model.to(torch_device)
thomwolf's avatar
thomwolf committed
1132

1133
            with torch.no_grad():
1134
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1135
            attentions = outputs[-1]
LysandreJik's avatar
LysandreJik committed
1136

1137
1138
1139
1140
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
thomwolf's avatar
thomwolf committed
1141

1142
1143
            heads_to_prune = {0: [0], 2: [1, 2]}
            model.prune_heads(heads_to_prune)
1144

1145
            with torch.no_grad():
1146
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1147
            attentions = outputs[-1]
1148

1149
1150
1151
1152
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
1153

1154
            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
thomwolf's avatar
thomwolf committed
1155

Patrick von Platen's avatar
Patrick von Platen committed
1156
    def test_hidden_states_output(self):
Joseph Liu's avatar
Joseph Liu committed
1157
        def check_hidden_states_output(inputs_dict, config, model_class):
1158
            model = model_class(config)
1159
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
1160
            model.eval()
Joseph Liu's avatar
Joseph Liu committed
1161

thomwolf's avatar
thomwolf committed
1162
            with torch.no_grad():
1163
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1164
1165

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
1166

Sylvain Gugger's avatar
Sylvain Gugger committed
1167
1168
1169
1170
            expected_num_layers = getattr(
                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
            )
            self.assertEqual(len(hidden_states), expected_num_layers)
1171

Patrick von Platen's avatar
Patrick von Platen committed
1172
1173
1174
1175
1176
1177
1178
            if hasattr(self.model_tester, "encoder_seq_length"):
                seq_length = self.model_tester.encoder_seq_length
                if hasattr(self.model_tester, "chunk_length") and self.model_tester.chunk_length > 1:
                    seq_length = seq_length * self.model_tester.chunk_length
            else:
                seq_length = self.model_tester.seq_length

1179
            self.assertListEqual(
Lysandre's avatar
Lysandre committed
1180
1181
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
1182
            )
thomwolf's avatar
thomwolf committed
1183

1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
            if config.is_encoder_decoder:
                hidden_states = outputs.decoder_hidden_states

                self.assertIsInstance(hidden_states, (list, tuple))
                self.assertEqual(len(hidden_states), expected_num_layers)
                seq_len = getattr(self.model_tester, "seq_length", None)
                decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)

                self.assertListEqual(
                    list(hidden_states[0].shape[-2:]),
                    [decoder_seq_length, self.model_tester.hidden_size],
                )

Joseph Liu's avatar
Joseph Liu committed
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            inputs_dict["output_hidden_states"] = True
            check_hidden_states_output(inputs_dict, config, model_class)

            # check that output_hidden_states also work using config
            del inputs_dict["output_hidden_states"]
            config.output_hidden_states = True

            check_hidden_states_output(inputs_dict, config, model_class)

1209
1210
1211
    def test_retain_grad_hidden_states_attentions(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
1212
        config.output_attentions = self.has_attentions
1213
1214
1215
1216
1217
1218
1219
1220
1221

        # no need to test all models as different heads yield the same functionality
        model_class = self.all_model_classes[0]
        model = model_class(config)
        model.to(torch_device)

        inputs = self._prepare_for_class(inputs_dict, model_class)

        outputs = model(**inputs)
1222

1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
        output = outputs[0]

        if config.is_encoder_decoder:
            # Seq2Seq models
            encoder_hidden_states = outputs.encoder_hidden_states[0]
            encoder_hidden_states.retain_grad()

            decoder_hidden_states = outputs.decoder_hidden_states[0]
            decoder_hidden_states.retain_grad()

1233
1234
1235
1236
1237
1238
1239
1240
1241
            if self.has_attentions:
                encoder_attentions = outputs.encoder_attentions[0]
                encoder_attentions.retain_grad()

                decoder_attentions = outputs.decoder_attentions[0]
                decoder_attentions.retain_grad()

                cross_attentions = outputs.cross_attentions[0]
                cross_attentions.retain_grad()
1242
1243
1244
1245
1246

            output.flatten()[0].backward(retain_graph=True)

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)
1247
1248
1249
1250
1251

            if self.has_attentions:
                self.assertIsNotNone(encoder_attentions.grad)
                self.assertIsNotNone(decoder_attentions.grad)
                self.assertIsNotNone(cross_attentions.grad)
1252
1253
1254
1255
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            hidden_states.retain_grad()
1256
1257
1258
1259

            if self.has_attentions:
                attentions = outputs.attentions[0]
                attentions.retain_grad()
1260
1261
1262
1263

            output.flatten()[0].backward(retain_graph=True)

            self.assertIsNotNone(hidden_states.grad)
1264
1265
1266

            if self.has_attentions:
                self.assertIsNotNone(attentions.grad)
1267

Pradhy729's avatar
Pradhy729 committed
1268
    def test_feed_forward_chunking(self):
Lysandre's avatar
Lysandre committed
1269
1270
1271
1272
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Pradhy729's avatar
Pradhy729 committed
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
        for model_class in self.all_model_classes:
            torch.manual_seed(0)
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            hidden_states_no_chunk = model(**self._prepare_for_class(inputs_dict, model_class))[0]

            torch.manual_seed(0)
            config.chunk_size_feed_forward = 1
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            hidden_states_with_chunk = model(**self._prepare_for_class(inputs_dict, model_class))[0]
            self.assertTrue(torch.allclose(hidden_states_no_chunk, hidden_states_with_chunk, atol=1e-3))

1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
    def test_resize_position_vector_embeddings(self):
        if not self.test_resize_position_embeddings:
            return

        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)

            if self.model_tester.is_training is False:
                model.eval()

            max_position_embeddings = config.max_position_embeddings

            # Retrieve the embeddings and clone theme
            if model.config.is_encoder_decoder:
                encoder_model_embed, decoder_model_embed = model.get_position_embeddings()
                encoder_cloned_embeddings = encoder_model_embed.weight.clone()
                decoder_cloned_embeddings = decoder_model_embed.weight.clone()
            else:
                model_embed = model.get_position_embeddings()
                cloned_embeddings = model_embed.weight.clone()

            # Check that resizing the position embeddings with a larger max_position_embeddings increases
            # the model's postion embeddings size
            model.resize_position_embeddings(max_position_embeddings + 10)
            self.assertEqual(model.config.max_position_embeddings, max_position_embeddings + 10)

            # Check that it actually resizes the embeddings matrix
            if model.config.is_encoder_decoder:
                encoder_model_embed, decoder_model_embed = model.get_position_embeddings()
                self.assertEqual(encoder_model_embed.weight.shape[0], encoder_cloned_embeddings.shape[0] + 10)
                self.assertEqual(decoder_model_embed.weight.shape[0], decoder_cloned_embeddings.shape[0] + 10)
            else:
                model_embed = model.get_position_embeddings()
                self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)

            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the position embeddings with a smaller max_position_embeddings decreases
            # the model's max_position_embeddings
            model.resize_position_embeddings(max_position_embeddings - 5)
            self.assertEqual(model.config.max_position_embeddings, max_position_embeddings - 5)

            # Check that it actually resizes the embeddings matrix
            if model.config.is_encoder_decoder:
                encoder_model_embed, decoder_model_embed = model.get_position_embeddings()
                self.assertEqual(encoder_model_embed.weight.shape[0], encoder_cloned_embeddings.shape[0] - 5)
                self.assertEqual(decoder_model_embed.weight.shape[0], decoder_cloned_embeddings.shape[0] - 5)
            else:
                model_embed = model.get_position_embeddings()
                self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 5)

            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
            models_equal = True

            if model.config.is_encoder_decoder:
                for p1, p2 in zip(encoder_cloned_embeddings, encoder_model_embed.weight):
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False
                for p1, p2 in zip(decoder_cloned_embeddings, decoder_model_embed.weight):
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False
            else:
                for p1, p2 in zip(cloned_embeddings, model_embed.weight):
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False

            self.assertTrue(models_equal)

Patrick von Platen's avatar
Patrick von Platen committed
1370
    def test_resize_tokens_embeddings(self):
Lysandre's avatar
Lysandre committed
1371
1372
1373
1374
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Patrick von Platen's avatar
Patrick von Platen committed
1375
        if not self.test_resize_embeddings:
1376
1377
1378
1379
1380
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
1381
            model.to(torch_device)
1382

Patrick von Platen's avatar
Patrick von Platen committed
1383
1384
1385
            if self.model_tester.is_training is False:
                model.eval()

1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
            model_vocab_size = config.vocab_size
            # Retrieve the embeddings and clone theme
            model_embed = model.resize_token_embeddings(model_vocab_size)
            cloned_embeddings = model_embed.weight.clone()

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
1396
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
1397
            model(**self._prepare_for_class(inputs_dict, model_class))
1398
1399
1400
1401
1402
1403
1404

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)

1405
1406
1407
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
1408
1409
1410
1411

            # make sure that decoder_input_ids are resized as well
            if "decoder_input_ids" in inputs_dict:
                inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
1412
            model(**self._prepare_for_class(inputs_dict, model_class))
1413

1414
1415
1416
1417
1418
1419
1420
1421
            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
            models_equal = True
            for p1, p2 in zip(cloned_embeddings, model_embed.weight):
                if p1.data.ne(p2.data).sum() > 0:
                    models_equal = False

            self.assertTrue(models_equal)

1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
    def test_resize_embeddings_untied(self):
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
        if not self.test_resize_embeddings:
            return

        original_config.tie_word_embeddings = False

        # if model cannot untied embeddings -> leave test
        if original_config.tie_word_embeddings:
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config).to(torch_device)

            # if no output embeddings -> leave test
            if model.get_output_embeddings() is None:
                continue

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_vocab_size = config.vocab_size
            model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
            if "decoder_input_ids" in inputs_dict:
                inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

Patrick von Platen's avatar
Patrick von Platen committed
1473
    def test_model_common_attributes(self):
1474
1475
1476
1477
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1478
1479
            self.assertIsInstance(model.get_input_embeddings(), (nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(nn.Embedding(10, 10))
1480
            x = model.get_output_embeddings()
1481
            self.assertTrue(x is None or isinstance(x, nn.Linear))
1482

1483
1484
1485
1486
1487
1488
1489
    def test_model_main_input_name(self):
        for model_class in self.all_model_classes:
            model_signature = inspect.signature(getattr(model_class, "forward"))
            # The main input is the name of the argument after `self`
            observed_main_input_name = list(model_signature.parameters.keys())[1]
            self.assertEqual(model_class.main_input_name, observed_main_input_name)

1490
    def test_correct_missing_keys(self):
1491
1492
        if not self.test_missing_keys:
            return
1493
1494
1495
1496
1497
1498
1499
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            base_model_prefix = model.base_model_prefix

            if hasattr(model, base_model_prefix):
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513

                extra_params = {k: v for k, v in model.named_parameters() if not k.startswith(base_model_prefix)}
                extra_params.update({k: v for k, v in model.named_buffers() if not k.startswith(base_model_prefix)})
                # Some models define this as None
                if model._keys_to_ignore_on_load_missing:
                    for key in model._keys_to_ignore_on_load_missing:
                        extra_params.pop(key, None)

                if not extra_params:
                    # In that case, we *are* on a head model, but every
                    # single key is not actual parameters and this is
                    # tested in `test_tied_model_weights_key_ignore` test.
                    continue

1514
1515
1516
                with tempfile.TemporaryDirectory() as temp_dir_name:
                    model.base_model.save_pretrained(temp_dir_name)
                    model, loading_info = model_class.from_pretrained(temp_dir_name, output_loading_info=True)
1517
                    self.assertGreater(len(loading_info["missing_keys"]), 0, model.__class__.__name__)
1518

1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
    def test_tie_model_weights(self):
        if not self.test_torchscript:
            return

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def check_same_values(layer_1, layer_2):
            equal = True
            for p1, p2 in zip(layer_1.weight, layer_2.weight):
                if p1.data.ne(p2.data).sum() > 0:
                    equal = False
            return equal

        for model_class in self.all_model_classes:
            config.torchscript = True
            model_not_tied = model_class(config)
            if model_not_tied.get_output_embeddings() is None:
                continue

            config_tied = copy.deepcopy(config)
            config_tied.torchscript = False
            model_tied = model_class(config_tied)
            params_tied = list(model_tied.parameters())
            # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(check_same_values(embeddings, decoding))

            # # Check that after modification, they remain the same.
            # embeddings.weight.data.div_(2)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
            # self.assertTrue(check_same_values(embeddings, decoding))

            # # Check that after modification, they remain the same.
            # decoding.weight.data.div_(4)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
            # self.assertTrue(check_same_values(embeddings, decoding))

            # Check that after resize they remain tied.
            model_tied.resize_token_embeddings(config.vocab_size + 10)
            params_tied_2 = list(model_tied.parameters())
            self.assertEqual(len(params_tied_2), len(params_tied))

            # decoding.weight.data.mul_(20)
            # # Check that the embedding layer and decoding layer are the same in size and in value
            # self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
            # self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))

1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
    def test_tied_model_weights_key_ignore(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        for model_class in self.all_model_classes:
            model_tied = model_class(config)
            with tempfile.TemporaryDirectory() as d:
                model_tied.save_pretrained(d)

                # We are nuking ALL weights on file, so every parameter should
                # yell on load. We're going to detect if we yell too much, or too little.
                with open(os.path.join(d, "pytorch_model.bin"), "wb") as f:
                    torch.save({}, f)
                model_reloaded, infos = model_class.from_pretrained(d, output_loading_info=True)

                # ! Actually we could use `state_dict()` and check iteratively the tensors which are the same (for instance using `tensor.data_ptr()`). to detect the duplicates.
                # ```python
                # model = GPT2LMHeadModel.from_pretrained("gpt2")
                # "lm_head.weight" in model.state_dict().keys()  # True
                # "lm_head.weight" in model.named_parameters() # False
                # In [6]: model.lm_head.weight.data_ptr()
                # Out[6]: 139901378371648
                # In [9]: model.transformer.wte.weight.data_ptr()
                # Out[9]: 139901378371648  # Same PTR, it's the same DATA ! we would need to check for stride too to be 100% accurate.
                # ```

                prefix = f"{model_reloaded.base_model_prefix}."
                params = dict(model_reloaded.named_parameters())
                params.update(dict(model_reloaded.named_buffers()))
                # param_names = set(k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys())
                param_names = set(k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys())

                missing_keys = set(infos["missing_keys"])

                extra_missing = missing_keys - param_names
                # missed_missing = param_names - missing_keys

                self.assertEqual(
                    extra_missing,
                    set(),
                    f"This model {model_class.__name__} might be missing some `keys_to_ignore`: {extra_missing}",
                )

                # self.assertEqual(
                #     missed_missing,
                #     set(),
                #     f"This model {model_class.__name__} ignores keys {missed_missing} but they look like real"
                #     " parameters",
                # )

1615
1616
1617
    def test_model_outputs_equivalence(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

Sam Shleifer's avatar
Sam Shleifer committed
1618
1619
1620
1621
        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

1622
1623
1624
1625
1626
1627
1628
1629
1630
        def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
            with torch.no_grad():
                tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
                dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()

                def recursive_check(tuple_object, dict_object):
                    if isinstance(tuple_object, (List, Tuple)):
                        for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
                            recursive_check(tuple_iterable_value, dict_iterable_value)
NielsRogge's avatar
NielsRogge committed
1631
1632
1633
1634
1635
                    elif isinstance(tuple_object, Dict):
                        for tuple_iterable_value, dict_iterable_value in zip(
                            tuple_object.values(), dict_object.values()
                        ):
                            recursive_check(tuple_iterable_value, dict_iterable_value)
1636
1637
1638
1639
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
Sam Shleifer's avatar
Sam Shleifer committed
1640
1641
1642
                            torch.allclose(
                                set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
                            ),
Sylvain Gugger's avatar
Sylvain Gugger committed
1643
1644
1645
1646
1647
1648
                            msg=(
                                "Tuple and dict output are not equal. Difference:"
                                f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
                                f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
                                f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
                            ),
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
                        )

                recursive_check(tuple_output, dict_output)

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

1674
1675
1676
1677
            if self.has_attentions:
                tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
                dict_inputs = self._prepare_for_class(inputs_dict, model_class)
                check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})
1678

1679
1680
1681
1682
1683
1684
1685
1686
1687
                tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})

                tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
                check_equivalence(
                    model, tuple_inputs, dict_inputs, {"output_hidden_states": True, "output_attentions": True}
                )
1688

1689
1690
1691
1692
    # Don't copy this method to model specific test file!
    # TODO: remove this method once the issues are all fixed!
    def _make_attention_mask_non_null(self, inputs_dict):
        """Make sure no sequence has all zeros as attention mask"""
1693

1694
1695
1696
        for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
            if k in inputs_dict:
                attention_mask = inputs_dict[k]
1697

1698
1699
1700
1701
1702
1703
                # Make sure no all 0s attention masks - to avoid failure at this moment.
                # Put `1` at the beginning of sequences to make it still work when combining causal attention masks.
                # TODO: remove this line once a fix regarding large negative values for attention mask is done.
                attention_mask = torch.cat(
                    [torch.ones_like(attention_mask[:, :1], dtype=attention_mask.dtype), attention_mask[:, 1:]], dim=-1
                )
1704

1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
                # Here we make the first sequence with all 0s as attention mask.
                # Currently, this will fail for `TFWav2Vec2Model`. This is caused by the different large negative
                # values, like `1e-4`, `1e-9`, `1e-30` and `-inf` for attention mask across models/frameworks.
                # TODO: enable this block once the large negative values thing is cleaned up.
                # (see https://github.com/huggingface/transformers/issues/14859)
                # attention_mask = torch.cat(
                #     [torch.zeros_like(attention_mask[:1], dtype=attention_mask.dtype), attention_mask[1:]],
                #     dim=0
                # )

                inputs_dict[k] = attention_mask

    # Don't copy this method to model specific test file!
    # TODO: remove this method once the issues are all fixed!
    def _postprocessing_to_ignore_test_cases(self, tf_outputs, pt_outputs, model_class):
        """For temporarily ignoring some failed test cases (issues to be fixed)"""

        tf_keys = set([k for k, v in tf_outputs.items() if v is not None])
        pt_keys = set([k for k, v in pt_outputs.items() if v is not None])

        key_differences = tf_keys.symmetric_difference(pt_keys)

        if model_class.__name__ in [
            "FlaubertWithLMHeadModel",
            "FunnelForPreTraining",
            "ElectraForPreTraining",
            "XLMWithLMHeadModel",
            "TransfoXLLMHeadModel",
        ]:
            for k in key_differences:
                if k in ["loss", "losses"]:
                    tf_keys.discard(k)
                    pt_keys.discard(k)
        elif model_class.__name__.startswith("GPT2"):
            # `TFGPT2` has `past_key_values` as a tensor while `GPT2` has it as a tuple.
            tf_keys.discard("past_key_values")
            pt_keys.discard("past_key_values")

        # create new outputs from the remaining fields
        new_tf_outputs = type(tf_outputs)(**{k: tf_outputs[k] for k in tf_keys})
        new_pt_outputs = type(pt_outputs)(**{k: pt_outputs[k] for k in pt_keys})

        return new_tf_outputs, new_pt_outputs

    # Copied from tests.test_modeling_tf_common.TFModelTesterMixin.check_pt_tf_outputs
    def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
1751
        """Check the outputs from PyTorch and TensorFlow models are close enough. Checks are done in a recursive way.
1752

1753
1754
1755
1756
1757
1758
1759
1760
        Args:
            model_class: The class of the model that is currently testing. For example, `TFBertModel`,
                TFBertForMaskedLM`, `TFBertForSequenceClassification`, etc. Mainly used for providing more informative
                error messages.
            name (`str`): The name of the output. For example, `output.hidden_states`, `output.attentions`, etc.
            attributes (`Tuple[str]`): The names of the output's element if the output is a tuple/list with each element
                being a named field in the output.
        """
1761

1762
1763
1764
        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")
1765

1766
1767
1768
1769
1770
1771
        # Allow `ModelOutput` (e.g. `CLIPOutput` has `text_model_output` and `vision_model_output`).
        if isinstance(tf_outputs, ModelOutput):
            self.assertTrue(
                isinstance(pt_outputs, ModelOutput),
                f"{name}: `pt_outputs` should an instance of `ModelOutput` when `tf_outputs` is",
            )
1772

1773
1774
1775
            # Don't copy this block to model specific test file!
            # TODO: remove this method and this line after issues are fixed
            tf_outputs, pt_outputs = self._postprocessing_to_ignore_test_cases(tf_outputs, pt_outputs, model_class)
1776

1777
1778
            tf_keys = [k for k, v in tf_outputs.items() if v is not None]
            pt_keys = [k for k, v in pt_outputs.items() if v is not None]
1779

1780
            self.assertEqual(tf_keys, pt_keys, f"{name}: Output keys differ between TF and PyTorch")
1781

1782
            # convert to the case of `tuple`
1783
            # appending each key to the current (string) `name`
1784
1785
1786
1787
            attributes = tuple([f"{name}.{k}" for k in tf_keys])
            self.check_pt_tf_outputs(
                tf_outputs.to_tuple(), pt_outputs.to_tuple(), model_class, tol=tol, name=name, attributes=attributes
            )
1788

1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
        # Allow `list` (e.g. `TransfoXLModelOutput.mems` is a list of tensors.)
        elif type(tf_outputs) in [tuple, list]:
            self.assertEqual(type(tf_outputs), type(pt_outputs), f"{name}: Output types differ between TF and PyTorch")
            self.assertEqual(len(tf_outputs), len(pt_outputs), f"{name}: Output lengths differ between TF and PyTorch")

            if attributes is not None:
                # case 1: each output has assigned name (e.g. a tuple form of a `ModelOutput`)
                self.assertEqual(
                    len(attributes),
                    len(tf_outputs),
1799
                    f"{name}: The tuple `attributes` should have the same length as `tf_outputs`",
1800
                )
1801
            else:
1802
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
1803
                attributes = tuple([f"{name}_{idx}" for idx in range(len(tf_outputs))])
1804

1805
1806
            for tf_output, pt_output, attr in zip(tf_outputs, pt_outputs, attributes):
                self.check_pt_tf_outputs(tf_output, pt_output, model_class, tol=tol, name=attr)
1807

1808
1809
1810
1811
        elif isinstance(tf_outputs, tf.Tensor):
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `tf_outputs` is"
            )
1812

1813
1814
            tf_outputs = tf_outputs.numpy()
            pt_outputs = pt_outputs.detach().to("cpu").numpy()
1815

1816
1817
1818
            self.assertEqual(
                tf_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between TF and PyTorch"
            )
1819

1820
1821
1822
1823
            # deal with NumPy's scalars to make replacing nan values by 0 work.
            if np.isscalar(tf_outputs):
                tf_outputs = np.array([tf_outputs])
                pt_outputs = np.array([pt_outputs])
1824

1825
1826
            tf_nans = np.isnan(tf_outputs)
            pt_nans = np.isnan(pt_outputs)
1827

1828
1829
1830
1831
            pt_outputs[tf_nans] = 0
            tf_outputs[tf_nans] = 0
            pt_outputs[pt_nans] = 0
            tf_outputs[pt_nans] = 0
1832

1833
            max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
1834
            self.assertLessEqual(max_diff, tol, f"{name}: Difference between PyTorch and TF is {max_diff} (>= {tol}).")
1835
1836
        else:
            raise ValueError(
1837
                "`tf_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `tf.Tensor`. Got"
Sylvain Gugger's avatar
Sylvain Gugger committed
1838
                f" {type(tf_outputs)} instead."
1839
1840
            )

1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
    def prepare_tf_inputs_from_pt_inputs(self, pt_inputs_dict):

        tf_inputs_dict = {}
        for key, tensor in pt_inputs_dict.items():
            # skip key that does not exist in tf
            if type(tensor) == bool:
                tf_inputs_dict[key] = tensor
            elif key == "input_values":
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            elif key == "pixel_values":
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            elif key == "input_features":
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            # other general float inputs
            elif tensor.is_floating_point():
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
            else:
                tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.int32)
1859

1860
        return tf_inputs_dict
1861

1862
    def check_pt_tf_models(self, tf_model, pt_model, pt_inputs_dict):
1863

1864
        tf_inputs_dict = self.prepare_tf_inputs_from_pt_inputs(pt_inputs_dict)
1865

1866
1867
1868
1869
        # send pytorch inputs to the correct device
        pt_inputs_dict = {
            k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs_dict.items()
        }
1870

1871
1872
        # send pytorch model to the correct device
        pt_model.to(torch_device)
1873

1874
1875
        # Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
        pt_model.eval()
1876

1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
        with torch.no_grad():
            pt_outputs = pt_model(**pt_inputs_dict)
        tf_outputs = tf_model(tf_inputs_dict)

        # tf models returned loss is usually a tensor rather than a scalar.
        # (see `hf_compute_loss`: it uses `tf.keras.losses.Reduction.NONE`)
        # Change it here to a scalar to match PyTorch models' loss
        tf_loss = getattr(tf_outputs, "loss", None)
        if tf_loss is not None:
            tf_outputs.loss = tf.math.reduce_mean(tf_loss)

        self.check_pt_tf_outputs(tf_outputs, pt_outputs, type(pt_model))

    @is_pt_tf_cross_test
    def test_pt_tf_model_equivalence(self):
        import transformers
1893
1894
1895

        for model_class in self.all_model_classes:

1896
1897
1898
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

            tf_model_class_name = "TF" + model_class.__name__  # Add the "TF" at the beginning
1899
            if not hasattr(transformers, tf_model_class_name):
1900
                # transformers does not have this model in TF version yet
1901
1902
                return

1903
1904
1905
            # Output all for aggressive testing
            config.output_hidden_states = True
            config.output_attentions = self.has_attentions
1906

1907
1908
1909
1910
            # Make sure no sequence has all zeros as attention mask, otherwise some tests fail due to the inconsistency
            # of the usage `1e-4`, `1e-9`, `1e-30`, `-inf`.
            # TODO: Use a uniform value for all models, make sure all tests pass without this processing, and remove it.
            self._make_attention_mask_non_null(inputs_dict)
1911
1912

            tf_model_class = getattr(transformers, tf_model_class_name)
1913
1914

            pt_model = model_class(config)
1915
1916
1917
1918
1919
1920
1921
1922
1923
            tf_model = tf_model_class(config)

            pt_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
            pt_inputs_dict_with_labels = self._prepare_for_class(
                inputs_dict,
                model_class,
                # Not all models accept "labels" in the forward pass (yet :) )
                return_labels=True if "labels" in inspect.signature(model_class.forward).parameters.keys() else False,
            )
1924
1925
1926
1927
1928
1929
1930
1931
1932

            # make sure only tf inputs are forward that actually exist in function args
            tf_input_keys = set(inspect.signature(tf_model.call).parameters.keys())

            # remove all head masks
            tf_input_keys.discard("head_mask")
            tf_input_keys.discard("cross_attn_head_mask")
            tf_input_keys.discard("decoder_head_mask")

1933
            pt_inputs_dict = {k: v for k, v in pt_inputs_dict.items() if k in tf_input_keys}
1934
1935
1936
1937
1938
1939
            pt_inputs_dict_with_labels = {k: v for k, v in pt_inputs_dict_with_labels.items() if k in tf_input_keys}

            # For some models (e.g. base models), there is no label returned.
            # Set the input dict to `None` to avoid check outputs twice for the same input dicts.
            if set(pt_inputs_dict_with_labels.keys()).symmetric_difference(pt_inputs_dict.keys()):
                pt_inputs_dict_with_labels = None
1940
1941

            # Check we can load pt model in tf and vice-versa with model => model functions
1942
1943
            # Here requires `tf_inputs_dict` to build `tf_model`
            tf_inputs_dict = self.prepare_tf_inputs_from_pt_inputs(pt_inputs_dict)
1944
            tf_model = transformers.load_pytorch_model_in_tf2_model(tf_model, pt_model, tf_inputs=tf_inputs_dict)
1945
            pt_model = transformers.load_tf2_model_in_pytorch_model(pt_model, tf_model)
1946

1947
1948
1949
1950
1951
            # Original test: check without `labels`
            self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict)
            # check with `labels`
            if pt_inputs_dict_with_labels:
                self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict_with_labels)
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962

            # Check we can load pt model in tf and vice-versa with checkpoint => model functions
            with tempfile.TemporaryDirectory() as tmpdirname:
                pt_checkpoint_path = os.path.join(tmpdirname, "pt_model.bin")
                torch.save(pt_model.state_dict(), pt_checkpoint_path)
                tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(tf_model, pt_checkpoint_path)

                tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
                tf_model.save_weights(tf_checkpoint_path)
                pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(pt_model, tf_checkpoint_path)

1963
1964
1965
1966
1967
            # Original test: check without `labels`
            self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict)
            # check with `labels`
            if pt_inputs_dict_with_labels:
                self.check_pt_tf_models(tf_model, pt_model, pt_inputs_dict_with_labels)
1968
1969
1970
1971
1972

    def assert_almost_equals(self, a: np.ndarray, b: np.ndarray, tol: float):
        diff = np.abs((a - b)).max()
        self.assertLessEqual(diff, tol, f"Difference between torch and flax is {diff} (>= {tol}).")

1973
    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
1974
1975
1976
1977
1978
1979
1980
1981
1982
        """
        Args:
            model_class: The class of the model that is currently testing. For example, ..., etc.
            Currently unused, but it could make debugging easier and faster.

            names: A string, or a list of strings. These specify what fx_outputs/pt_outputs represent in the model outputs.
                Currently unused, but in the future, we could use this information to make the error message clearer
                by giving the name(s) of the output tensor(s) with large difference(s) between PT and Flax.
        """
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022

        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")

        # Allow `ModelOutput` (e.g. `CLIPOutput` has `text_model_output` and `vision_model_output`).
        if isinstance(fx_outputs, ModelOutput):
            self.assertTrue(
                isinstance(pt_outputs, ModelOutput),
                f"{name}: `pt_outputs` should an instance of `ModelOutput` when `fx_outputs` is",
            )

            fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
            pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

            self.assertEqual(fx_keys, pt_keys, f"{name}: Output keys differ between Flax and PyTorch")

            # convert to the case of `tuple`
            # appending each key to the current (string) `name`
            attributes = tuple([f"{name}.{k}" for k in fx_keys])
            self.check_pt_flax_outputs(
                fx_outputs.to_tuple(), pt_outputs.to_tuple(), model_class, tol=tol, name=name, attributes=attributes
            )

        # Allow `list` (e.g. `TransfoXLModelOutput.mems` is a list of tensors.)
        elif type(fx_outputs) in [tuple, list]:
            self.assertEqual(
                type(fx_outputs), type(pt_outputs), f"{name}: Output types differ between Flax and PyTorch"
            )
            self.assertEqual(
                len(fx_outputs), len(pt_outputs), f"{name}: Output lengths differ between Flax and PyTorch"
            )

            if attributes is not None:
                # case 1: each output has assigned name (e.g. a tuple form of a `ModelOutput`)
                self.assertEqual(
                    len(attributes),
                    len(fx_outputs),
                    f"{name}: The tuple `attributes` should have the same length as `fx_outputs`",
                )
2023
            else:
2024
2025
2026
2027
2028
2029
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
                attributes = tuple([f"{name}_{idx}" for idx in range(len(fx_outputs))])

            for fx_output, pt_output, attr in zip(fx_outputs, pt_outputs, attributes):
                self.check_pt_flax_outputs(fx_output, pt_output, model_class, tol=tol, name=attr)

2030
        elif isinstance(fx_outputs, jnp.ndarray):
2031
2032
2033
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `fx_outputs` is"
            )
2034
2035
2036
2037
2038

            # Using `np.asarray` gives `ValueError: assignment destination is read-only` at the line `fx_outputs[fx_nans] = 0`.
            fx_outputs = np.array(fx_outputs)
            pt_outputs = pt_outputs.detach().to("cpu").numpy()

2039
2040
2041
2042
2043
2044
2045
2046
2047
            self.assertEqual(
                fx_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between Flax and PyTorch"
            )

            # deal with NumPy's scalars to make replacing nan values by 0 work.
            if np.isscalar(fx_outputs):
                fx_outputs = np.array([fx_outputs])
                pt_outputs = np.array([pt_outputs])

2048
2049
2050
2051
2052
2053
2054
2055
            fx_nans = np.isnan(fx_outputs)
            pt_nans = np.isnan(pt_outputs)

            pt_outputs[fx_nans] = 0
            fx_outputs[fx_nans] = 0
            pt_outputs[pt_nans] = 0
            fx_outputs[pt_nans] = 0

2056
2057
2058
2059
            max_diff = np.amax(np.abs(fx_outputs - pt_outputs))
            self.assertLessEqual(
                max_diff, tol, f"{name}: Difference between PyTorch and Flax is {max_diff} (>= {tol})."
            )
2060
2061
        else:
            raise ValueError(
2062
2063
                "`fx_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `jnp.ndarray`. Got"
                f" {type(fx_outputs)} instead."
2064
2065
            )

2066
2067
2068
2069
2070
2071
2072
2073
2074
    @is_pt_flax_cross_test
    def test_equivalence_pt_to_flax(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            with self.subTest(model_class.__name__):
                fx_model_class_name = "Flax" + model_class.__name__

                if not hasattr(transformers, fx_model_class_name):
2075
                    # no flax model exists for this class
2076
2077
                    return

2078
2079
2080
2081
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2082
2083
                fx_model_class = getattr(transformers, fx_model_class_name)

2084
2085
2086
2087
2088
2089
                # load PyTorch class
                pt_model = model_class(config).eval()
                # Flax models don't use the `use_cache` option and cache is not returned as a default.
                # So we disable `use_cache` here for PyTorch model.
                pt_model.config.use_cache = False

2090
2091
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2092

2093
2094
2095
2096
2097
2098
2099
2100
2101
                # make sure only flax inputs are forward that actually exist in function args
                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()

                # prepare inputs
                pt_inputs = self._prepare_for_class(inputs_dict, model_class)

                # remove function args that don't exist in Flax
                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}

2102
2103
2104
2105
2106
2107
2108
2109
                # send pytorch inputs to the correct device
                pt_inputs = {
                    k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs.items()
                }

                # convert inputs to Flax
                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}

2110
2111
2112
                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
                fx_model.params = fx_state

2113
2114
2115
                # send pytorch model to the correct device
                pt_model.to(torch_device)

2116
                with torch.no_grad():
2117
2118
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)
2119

2120
2121
2122
2123
                fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
2124
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2125
2126
2127
2128
2129

                with tempfile.TemporaryDirectory() as tmpdirname:
                    pt_model.save_pretrained(tmpdirname)
                    fx_model_loaded = fx_model_class.from_pretrained(tmpdirname, from_pt=True)

2130
2131
2132
2133
2134
2135
                fx_outputs_loaded = fx_model_loaded(**fx_inputs)

                fx_keys = tuple([k for k, v in fx_outputs_loaded.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
2136
                self.check_pt_flax_outputs(fx_outputs_loaded, pt_outputs, model_class)
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149

    @is_pt_flax_cross_test
    def test_equivalence_flax_to_pt(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            with self.subTest(model_class.__name__):
                fx_model_class_name = "Flax" + model_class.__name__

                if not hasattr(transformers, fx_model_class_name):
                    # no flax model exists for this class
                    return

2150
2151
2152
2153
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2154
2155
                fx_model_class = getattr(transformers, fx_model_class_name)

2156
2157
2158
2159
2160
2161
                # load PyTorch class
                pt_model = model_class(config).eval()
                # Flax models don't use the `use_cache` option and cache is not returned as a default.
                # So we disable `use_cache` here for PyTorch model.
                pt_model.config.use_cache = False

2162
2163
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2164

2165
2166
2167
2168
2169
2170
2171
2172
2173
                # make sure only flax inputs are forward that actually exist in function args
                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()

                # prepare inputs
                pt_inputs = self._prepare_for_class(inputs_dict, model_class)

                # remove function args that don't exist in Flax
                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}

2174
2175
2176
2177
                # send pytorch inputs to the correct device
                pt_inputs = {
                    k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs.items()
                }
2178

2179
                # convert inputs to Flax
2180
2181
                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}

2182
2183
2184
2185
2186
2187
2188
                pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)

                # make sure weights are tied in PyTorch
                pt_model.tie_weights()

                # send pytorch model to the correct device
                pt_model.to(torch_device)
2189

2190
2191
2192
2193
2194
2195
2196
2197
                with torch.no_grad():
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)

                fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
2198
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2199
2200
2201
2202
2203

                with tempfile.TemporaryDirectory() as tmpdirname:
                    fx_model.save_pretrained(tmpdirname)
                    pt_model_loaded = model_class.from_pretrained(tmpdirname, from_flax=True)

2204
2205
2206
2207
                # send pytorch model to the correct device
                pt_model_loaded.to(torch_device)
                pt_model_loaded.eval()

2208
                with torch.no_grad():
2209
                    pt_outputs_loaded = pt_model_loaded(**pt_inputs)
2210

2211
2212
2213
2214
                fx_keys = tuple([k for k, v in fx_outputs.items() if v is not None])
                pt_keys = tuple([k for k, v in pt_outputs_loaded.items() if v is not None])

                self.assertEqual(fx_keys, pt_keys)
2215
                self.check_pt_flax_outputs(fx_outputs, pt_outputs_loaded, model_class)
2216

Patrick von Platen's avatar
Patrick von Platen committed
2217
    def test_inputs_embeds(self):
2218
2219
2220
2221
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
2222
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
2223
            model.eval()
2224

2225
            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
2226

2227
2228
2229
2230
2231
2232
2233
2234
2235
            if not self.is_encoder_decoder:
                input_ids = inputs["input_ids"]
                del inputs["input_ids"]
            else:
                encoder_input_ids = inputs["input_ids"]
                decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
                del inputs["input_ids"]
                inputs.pop("decoder_input_ids", None)

2236
2237
            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
2238
                inputs["inputs_embeds"] = wte(input_ids)
2239
            else:
2240
2241
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
2242

thomwolf's avatar
thomwolf committed
2243
            with torch.no_grad():
Weizhen's avatar
Weizhen committed
2244
                model(**inputs)[0]
2245

2246
2247
    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
2248
2249
2250
2251
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        # some params shouldn't be scattered by nn.DataParallel
        # so just remove them if they are present.
2252
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask", "cross_attn_head_mask"]
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
        for k in blacklist_non_batched_params:
            inputs_dict.pop(k, None)

        # move input tensors to cuda:O
        for k, v in inputs_dict.items():
            if torch.is_tensor(v):
                inputs_dict[k] = v.to(0)

        for model_class in self.all_model_classes:
            model = model_class(config=config)
            model.to(0)
            model.eval()

            # Wrap model in nn.DataParallel
2267
            model = nn.DataParallel(model)
2268
            with torch.no_grad():
2269
                _ = model(**self._prepare_for_class(inputs_dict, model_class))
2270

2271
2272
2273
    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
2274
            return
2275

2276
        # a candidate for testing_utils
2277
        def get_current_gpu_memory_use():
Patrick von Platen's avatar
Patrick von Platen committed
2278
            """returns a list of cuda memory allocations per GPU in MBs"""
2279
2280
2281
2282
2283

            per_device_memory = []
            for id in range(torch.cuda.device_count()):
                with torch.cuda.device(id):
                    per_device_memory.append(torch.cuda.memory_allocated() >> 20)
2284
2285
2286
2287
2288
2289
2290
2291
2292

            return per_device_memory

        # Needs a large model to see the difference.
        config = self.model_tester.get_large_model_config()

        for model_class in self.all_parallelizable_model_classes:
            torch.cuda.empty_cache()

2293
2294
2295
            # 1. single gpu memory load + unload + memory measurements
            # Retrieve initial memory usage (can easily be ~0.6-1.5GB if cuda-kernels have been preloaded by previous tests)
            memory_at_start = get_current_gpu_memory_use()
2296

2297
2298
            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
2299
2300
2301
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

2302
2303
2304
            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

2305
            del model
2306
            gc.collect()
2307
2308
            torch.cuda.empty_cache()

2309
2310
2311
            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()
2312
2313

            # Spread model layers over multiple devices
2314
            model = model_class(config)
2315
2316
2317
2318
            model.parallelize()
            memory_after_parallelization = get_current_gpu_memory_use()

            # Assert that the memory use on all devices is higher than it was when loaded only on CPU
2319
            for n in range(len(model.device_map.keys())):
2320
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])
2321

2322
            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
2323
2324
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

2325
2326
            # Assert that the memory use of device 1 is higher than it was when the entire model was loaded
            # on device 0 and device 1 wasn't used at all
2327
2328
2329
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
2330
            gc.collect()
2331
2332
2333
2334
2335
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
2336
            return
2337
2338
2339
2340
2341
2342

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_parallelizable_model_classes:
            inputs_dict = self._prepare_for_class(inputs_dict, model_class)

2343
            def cast_to_device(dictionary, device):
2344
2345
2346
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
2347
                        output[k] = v.to(device)
2348
2349
2350
2351
2352
                    else:
                        output[k] = v

                return output

2353
2354
2355
2356
2357
2358
            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))
2359
2360
2361
2362
2363
2364
2365
2366

            for value, parallel_value in zip(output, parallel_output):
                if isinstance(value, torch.Tensor):
                    self.assertTrue(torch.allclose(value, parallel_value.to("cpu"), atol=1e-7))
                elif isinstance(value, (Tuple, List)):
                    for value_, parallel_value_ in zip(value, parallel_value):
                        self.assertTrue(torch.allclose(value_, parallel_value_.to("cpu"), atol=1e-7))

2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
    @require_torch_multi_gpu
    def test_model_parallel_beam_search(self):
        if not self.test_model_parallel:
            return

        all_generative_and_parallelizable_model_classes = tuple(
            set(self.all_generative_model_classes).intersection(self.all_parallelizable_model_classes)
        )

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in all_generative_and_parallelizable_model_classes:
            inputs_dict = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config)

            def cast_to_device(dictionary, device):
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
                        output[k] = v.to(device)
                    else:
                        output[k] = v

                return output

            model.parallelize()
            model.generate(**cast_to_device(inputs_dict, "cuda:0"), num_beams=2)

2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
    def check_device_map_is_respected(self, model, device_map):
        for param_name, param in model.named_parameters():
            # Find device in device_map
            while len(param_name) > 0 and param_name not in device_map:
                param_name = ".".join(param_name.split(".")[:-1])
            if param_name not in device_map:
                raise ValueError("device map is incomplete, it does not contain any device for `param_name`.")

            param_device = device_map[param_name]
            if param_device in ["cpu", "disk"]:
                self.assertEqual(param.device, torch.device("meta"))
            else:
                self.assertEqual(param.device, torch.device(param_device))

Sylvain Gugger's avatar
Sylvain Gugger committed
2409
2410
2411
2412
2413
2414
2415
2416
2417
    @require_accelerate
    @require_torch_gpu
    def test_disk_offload(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class._no_split_modules is None:
                continue

2418
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2419
2420
            model = model_class(config).eval()
            model = model.to(torch_device)
2421
            torch.manual_seed(0)
2422
            base_output = model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2423
2424

            model_size = compute_module_sizes(model)[""]
2425
            max_size = int(self.model_split_percents[0] * model_size)
Sylvain Gugger's avatar
Sylvain Gugger committed
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(tmp_dir)

                max_memory = {0: max_size, "cpu": max_size}
                with self.assertRaises(ValueError):
                    # This errors out cause it's missing an offload folder
                    new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)

                new_model = model_class.from_pretrained(
                    tmp_dir, device_map="auto", max_memory=max_memory, offload_folder=tmp_dir
                )

                self.check_device_map_is_respected(new_model, new_model.hf_device_map)
2439
                torch.manual_seed(0)
2440
                new_output = new_model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2441
2442
2443

                self.assertTrue(torch.allclose(base_output[0], new_output[0]))

2444
2445
2446
2447
2448
2449
2450
2451
2452
    @require_accelerate
    @require_torch_gpu
    def test_cpu_offload(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class._no_split_modules is None:
                continue

2453
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2454
2455
            model = model_class(config).eval()
            model = model.to(torch_device)
2456
2457

            torch.manual_seed(0)
2458
            base_output = model(**inputs_dict_class)
2459
2460
2461

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
2462
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(tmp_dir)

                for max_size in max_gpu_sizes:
                    max_memory = {0: max_size, "cpu": model_size * 2}
                    new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
                    # Making sure part of the model will actually end up offloaded
                    self.assertSetEqual(set(new_model.hf_device_map.values()), {0, "cpu"})

                    self.check_device_map_is_respected(new_model, new_model.hf_device_map)
2473
2474

                    torch.manual_seed(0)
2475
                    new_output = new_model(**inputs_dict_class)
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487

                    self.assertTrue(torch.allclose(base_output[0], new_output[0]))

    @require_accelerate
    @require_torch_multi_gpu
    def test_model_parallelism(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class._no_split_modules is None:
                continue

2488
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2489
2490
            model = model_class(config).eval()
            model = model.to(torch_device)
2491
2492

            torch.manual_seed(0)
2493
            base_output = model(**inputs_dict_class)
2494
2495
2496

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
2497
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(tmp_dir)

                for max_size in max_gpu_sizes:
                    max_memory = {0: max_size, 1: model_size * 2, "cpu": model_size * 2}
                    new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
                    # Making sure part of the model will actually end up offloaded
                    self.assertSetEqual(set(new_model.hf_device_map.values()), {0, 1})

                    self.check_device_map_is_respected(new_model, new_model.hf_device_map)
2508
2509

                    torch.manual_seed(0)
2510
                    new_output = new_model(**inputs_dict_class)
2511
2512
2513

                    self.assertTrue(torch.allclose(base_output[0], new_output[0]))

2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
    def test_problem_types(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        problem_types = [
            {"title": "multi_label_classification", "num_labels": 2, "dtype": torch.float},
            {"title": "single_label_classification", "num_labels": 1, "dtype": torch.long},
            {"title": "regression", "num_labels": 1, "dtype": torch.float},
        ]

        for model_class in self.all_model_classes:
2524
2525
2526
2527
            if model_class not in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
            ]:
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
                continue

            for problem_type in problem_types:
                with self.subTest(msg=f"Testing {model_class} with {problem_type['title']}"):

                    config.problem_type = problem_type["title"]
                    config.num_labels = problem_type["num_labels"]

                    model = model_class(config)
                    model.to(torch_device)
                    model.train()

                    inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)

                    if problem_type["num_labels"] > 1:
                        inputs["labels"] = inputs["labels"].unsqueeze(1).repeat(1, problem_type["num_labels"])

                    inputs["labels"] = inputs["labels"].to(problem_type["dtype"])

2547
2548
2549
2550
2551
2552
                    # This tests that we do not trigger the warning form PyTorch "Using a target size that is different
                    # to the input size. This will likely lead to incorrect results due to broadcasting. Please ensure
                    # they have the same size." which is a symptom something in wrong for the regression problem.
                    # See https://github.com/huggingface/transformers/issues/11780
                    with warnings.catch_warnings(record=True) as warning_list:
                        loss = model(**inputs).loss
2553
2554
2555
2556
2557
                    for w in warning_list:
                        if "Using a target size that is different to the input size" in str(w.message):
                            raise ValueError(
                                f"Something is going wrong in the regression problem: intercepted {w.message}"
                            )
2558

2559
2560
                    loss.backward()

2561
    def test_load_with_mismatched_shapes(self):
2562
2563
        if not self.test_mismatched_shapes:
            return
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if model_class not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING):
                continue

            with self.subTest(msg=f"Testing {model_class}"):
                with tempfile.TemporaryDirectory() as tmp_dir:
                    model = model_class(config)
                    model.save_pretrained(tmp_dir)

                    # Fails when we don't set ignore_mismatched_sizes=True
2576
                    with self.assertRaises(RuntimeError):
2577
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
2578
2579
                    with self.assertRaises(RuntimeError):
                        new_model_without_prefix = AutoModel.from_pretrained(tmp_dir, vocab_size=10)
2580
2581

                    logger = logging.get_logger("transformers.modeling_utils")
2582

2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
                    with CaptureLogger(logger) as cl:
                        new_model = AutoModelForSequenceClassification.from_pretrained(
                            tmp_dir, num_labels=42, ignore_mismatched_sizes=True
                        )
                    self.assertIn("the shapes did not match", cl.out)
                    new_model.to(torch_device)
                    inputs = self._prepare_for_class(inputs_dict, model_class)
                    logits = new_model(**inputs).logits
                    self.assertEqual(logits.shape[1], 42)

2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
                    with CaptureLogger(logger) as cl:
                        new_model_without_prefix = AutoModel.from_pretrained(
                            tmp_dir, vocab_size=10, ignore_mismatched_sizes=True
                        )
                    self.assertIn("the shapes did not match", cl.out)
                    input_ids = ids_tensor((2, 8), 10)
                    new_model_without_prefix.to(torch_device)
                    if self.is_encoder_decoder:
                        new_model_without_prefix(input_ids, decoder_input_ids=input_ids)
                    else:
                        new_model_without_prefix(input_ids)

2605

2606
global_rng = random.Random()
thomwolf's avatar
thomwolf committed
2607
2608


thomwolf's avatar
thomwolf committed
2609
def ids_tensor(shape, vocab_size, rng=None, name=None):
2610
    #  Creates a random int32 tensor of the shape within the vocab size
thomwolf's avatar
thomwolf committed
2611
    if rng is None:
2612
        rng = global_rng
thomwolf's avatar
thomwolf committed
2613

thomwolf's avatar
thomwolf committed
2614
2615
2616
    total_dims = 1
    for dim in shape:
        total_dims *= dim
thomwolf's avatar
thomwolf committed
2617

thomwolf's avatar
thomwolf committed
2618
2619
2620
    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))
thomwolf's avatar
thomwolf committed
2621

2622
    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
thomwolf's avatar
thomwolf committed
2623
2624


2625
2626
2627
2628
2629
2630
2631
def random_attention_mask(shape, rng=None, name=None):
    attn_mask = ids_tensor(shape, vocab_size=2, rng=None, name=None)
    # make sure that at least one token is attended to for each batch
    attn_mask[:, -1] = 1
    return attn_mask


2632
def floats_tensor(shape, scale=1.0, rng=None, name=None):
Patrick von Platen's avatar
Patrick von Platen committed
2633
    """Creates a random float32 tensor"""
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
    if rng is None:
        rng = global_rng

    total_dims = 1
    for dim in shape:
        total_dims *= dim

    values = []
    for _ in range(total_dims):
        values.append(rng.random() * scale)

2645
    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()
2646
2647


2648
2649
2650
2651
2652
2653
2654
2655
2656
def check_models_equal(model1, model2):
    models_are_equal = True
    for model1_p, model2_p in zip(model1.parameters(), model2.parameters()):
        if model1_p.data.ne(model2_p.data).sum() > 0:
            models_are_equal = False

    return models_are_equal


2657
@require_torch
2658
class ModelUtilsTest(TestCasePlus):
2659
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
2660
    def test_model_from_pretrained(self):
2661
        for model_name in BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
thomwolf's avatar
thomwolf committed
2662
2663
2664
2665
2666
2667
2668
2669
            config = BertConfig.from_pretrained(model_name)
            self.assertIsNotNone(config)
            self.assertIsInstance(config, PretrainedConfig)

            model = BertModel.from_pretrained(model_name)
            model, loading_info = BertModel.from_pretrained(model_name, output_loading_info=True)
            self.assertIsNotNone(model)
            self.assertIsInstance(model, PreTrainedModel)
Lysandre Debut's avatar
Lysandre Debut committed
2670
2671
2672
2673
2674

            self.assertEqual(len(loading_info["missing_keys"]), 0)
            self.assertEqual(len(loading_info["unexpected_keys"]), 8)
            self.assertEqual(len(loading_info["mismatched_keys"]), 0)
            self.assertEqual(len(loading_info["error_msgs"]), 0)
thomwolf's avatar
thomwolf committed
2675
2676

            config = BertConfig.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
Lysandre Debut's avatar
Lysandre Debut committed
2677
2678
2679
2680

            # Not sure this is the intended behavior. TODO fix Lysandre & Thom
            config.name_or_path = model_name

thomwolf's avatar
thomwolf committed
2681
2682
2683
            model = BertModel.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
            self.assertEqual(model.config.output_hidden_states, True)
            self.assertEqual(model.config, config)
2684

2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
    def test_model_from_pretrained_subfolder(self):
        config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
        model = BertModel(config)

        subfolder = "bert"
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(os.path.join(tmp_dir, subfolder))

            with self.assertRaises(OSError):
                _ = BertModel.from_pretrained(tmp_dir)

            model_loaded = BertModel.from_pretrained(tmp_dir, subfolder=subfolder)

        self.assertTrue(check_models_equal(model, model_loaded))

    def test_model_from_pretrained_subfolder_sharded(self):
        config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
        model = BertModel(config)

        subfolder = "bert"
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(os.path.join(tmp_dir, subfolder), max_shard_size="10KB")

            with self.assertRaises(OSError):
                _ = BertModel.from_pretrained(tmp_dir)

            model_loaded = BertModel.from_pretrained(tmp_dir, subfolder=subfolder)

        self.assertTrue(check_models_equal(model, model_loaded))

    def test_model_from_pretrained_hub_subfolder(self):
        subfolder = "bert"
        model_id = "hf-internal-testing/tiny-random-bert-subfolder"
        with self.assertRaises(OSError):
            _ = BertModel.from_pretrained(model_id)

        model = BertModel.from_pretrained(model_id, subfolder=subfolder)

        self.assertIsNotNone(model)

    def test_model_from_pretrained_hub_subfolder_sharded(self):
        subfolder = "bert"
        model_id = "hf-internal-testing/tiny-random-bert-sharded-subfolder"
        with self.assertRaises(OSError):
            _ = BertModel.from_pretrained(model_id)

        model = BertModel.from_pretrained(model_id, subfolder=subfolder)

        self.assertIsNotNone(model)

2735
2736
2737
2738
    def test_model_from_pretrained_with_different_pretrained_model_name(self):
        model = T5ForConditionalGeneration.from_pretrained(TINY_T5)
        self.assertIsNotNone(model)

2739
2740
        logger = logging.get_logger("transformers.configuration_utils")
        with CaptureLogger(logger) as cl:
2741
            BertModel.from_pretrained(TINY_T5)
2742
        self.assertTrue("You are using a model of type t5 to instantiate a model of type bert" in cl.out)
Sylvain Gugger's avatar
Sylvain Gugger committed
2743

2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
    @require_torch
    def test_model_from_config_torch_dtype(self):
        # test that the model can be instantiated with dtype of user's choice - as long as it's a
        # float dtype. To make it happen config.torch_dtype needs to be set before instantiating the
        # model from the config object.

        config = T5Config.from_pretrained(TINY_T5)
        model = AutoModel.from_config(config)
        # XXX: isn't supported
        # model = T5ForConditionalGeneration.from_config(config)
        self.assertEqual(model.dtype, torch.float32)

        model = AutoModel.from_config(config, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

        # torch.set_default_dtype() supports only float dtypes, so will fail with non-float type
        with self.assertRaises(ValueError):
            model = AutoModel.from_config(config, torch_dtype=torch.int64)

    @require_torch
    def test_model_from_pretrained_torch_dtype(self):
        # test that the model can be instantiated with dtype of either
2766
2767
        # 1. explicit from_pretrained's torch_dtype argument
        # 2. via autodiscovery by looking at model weights (torch_dtype="auto")
2768
        # so if a model.half() was saved, we want it to be instantiated as such.
2769
2770
        #
        # test an explicit model class, but also AutoModel separately as the latter goes through a different code path
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
        model_path = self.get_auto_remove_tmp_dir()

        # baseline - we know TINY_T5 is fp32 model
        model = T5ForConditionalGeneration.from_pretrained(TINY_T5)
        self.assertEqual(model.dtype, torch.float32)

        # test the default fp32 save_pretrained => from_pretrained cycle
        model.save_pretrained(model_path)
        model = T5ForConditionalGeneration.from_pretrained(model_path)
        self.assertEqual(model.dtype, torch.float32)
        # test with auto-detection
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float32)

        # test forced loading in fp16 (even though the weights are in fp32)
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

        # test fp16 save_pretrained, loaded with auto-detection
        model = model.half()
        model.save_pretrained(model_path)
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype="auto")
2793
        self.assertEqual(model.config.torch_dtype, torch.float16)
2794
2795
        self.assertEqual(model.dtype, torch.float16)

2796
2797
2798
2799
2800
        # tests `config.torch_dtype` saving
        with open(f"{model_path}/config.json") as f:
            config_dict = json.load(f)
        self.assertEqual(config_dict["torch_dtype"], "float16")

2801
2802
2803
2804
        # test fp16 save_pretrained, loaded with the explicit fp16
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

2805
2806
2807
2808
2809
2810
2811
2812
        # test AutoModel separately as it goes through a different path
        # test auto-detection
        model = AutoModel.from_pretrained(TINY_T5, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float32)
        # test forcing an explicit dtype
        model = AutoModel.from_pretrained(TINY_T5, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

2813
2814
2815
2816
        # test model whose first param is not of a floating type, but int
        model = AutoModel.from_pretrained(TINY_BERT_FOR_TOKEN_CLASSIFICATION, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float32)

2817
2818
2819
2820
2821
2822
2823
    def test_no_super_init_config_and_model(self):
        config = NoSuperInitConfig(attribute=32)
        model = NoSuperInitModel(config)

        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir)

2824
2825
2826
2827
            new_model = NoSuperInitModel.from_pretrained(tmp_dir)

        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))
2828

Sylvain Gugger's avatar
Sylvain Gugger committed
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
    def test_shard_checkpoint(self):
        # This is the model we will use, total size 340,000 bytes.
        model = torch.nn.Sequential(
            torch.nn.Linear(100, 200, bias=False),  # size 80,000
            torch.nn.Linear(200, 200, bias=False),  # size 160,000
            torch.nn.Linear(200, 100, bias=False),  # size 80,000
            torch.nn.Linear(100, 50, bias=False),  # size 20,000
        )
        state_dict = model.state_dict()

        with self.subTest("No shard when max size is bigger than model size"):
            shards, index = shard_checkpoint(state_dict)
            self.assertIsNone(index)
            self.assertDictEqual(shards, {WEIGHTS_NAME: state_dict})

        with self.subTest("Test sharding, no weights bigger than max size"):
            shards, index = shard_checkpoint(state_dict, max_shard_size="300kB")
            # Split is first two layers then last two.
            self.assertDictEqual(
                index,
                {
                    "metadata": {"total_size": 340000},
                    "weight_map": {
                        "0.weight": "pytorch_model-00001-of-00002.bin",
                        "1.weight": "pytorch_model-00001-of-00002.bin",
                        "2.weight": "pytorch_model-00002-of-00002.bin",
                        "3.weight": "pytorch_model-00002-of-00002.bin",
                    },
                },
            )

            shard1 = {"0.weight": state_dict["0.weight"], "1.weight": state_dict["1.weight"]}
            shard2 = {"2.weight": state_dict["2.weight"], "3.weight": state_dict["3.weight"]}
            self.assertDictEqual(
                shards, {"pytorch_model-00001-of-00002.bin": shard1, "pytorch_model-00002-of-00002.bin": shard2}
            )

        with self.subTest("Test sharding with weights bigger than max size"):
            shards, index = shard_checkpoint(state_dict, max_shard_size="100kB")
            # Split is first layer, second layer then last 2.
            self.assertDictEqual(
                index,
                {
                    "metadata": {"total_size": 340000},
                    "weight_map": {
                        "0.weight": "pytorch_model-00001-of-00003.bin",
                        "1.weight": "pytorch_model-00002-of-00003.bin",
                        "2.weight": "pytorch_model-00003-of-00003.bin",
                        "3.weight": "pytorch_model-00003-of-00003.bin",
                    },
                },
            )

            shard1 = {"0.weight": state_dict["0.weight"]}
            shard2 = {"1.weight": state_dict["1.weight"]}
            shard3 = {"2.weight": state_dict["2.weight"], "3.weight": state_dict["3.weight"]}
            self.assertDictEqual(
                shards,
                {
                    "pytorch_model-00001-of-00003.bin": shard1,
                    "pytorch_model-00002-of-00003.bin": shard2,
                    "pytorch_model-00003-of-00003.bin": shard3,
                },
            )

    def test_checkpoint_sharding_local(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        with tempfile.TemporaryDirectory() as tmp_dir:
            # We use the same folder for various sizes to make sure a new save erases the old checkpoint.
            for max_size in ["50kB", "50kiB", "100kB", "100kiB", "200kB", "200kiB"]:
                model.save_pretrained(tmp_dir, max_shard_size=max_size)

                # Get each shard file and its size
                shard_to_size = {}
                for shard in os.listdir(tmp_dir):
                    if shard.endswith(".bin"):
                        shard_file = os.path.join(tmp_dir, shard)
                        shard_to_size[shard_file] = os.path.getsize(shard_file)

                index_file = os.path.join(tmp_dir, WEIGHTS_INDEX_NAME)
                # Check there is an index but no regular weight file
                self.assertTrue(os.path.isfile(index_file))
                self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))

                # Check a file is bigger than max_size only when it has a single weight
                for shard_file, size in shard_to_size.items():
                    if max_size.endswith("kiB"):
                        max_size_int = int(max_size[:-3]) * 2**10
                    else:
                        max_size_int = int(max_size[:-2]) * 10**3
                    # Note: pickle adds some junk so the weight of the file can end up being slightly bigger than
                    # the size asked for (since we count parameters)
                    if size >= max_size_int + 50000:
                        state_dict = torch.load(shard_file)
                        self.assertEqual(len(state_dict), 1)

                # Check the index and the shard files found match
                with open(index_file, "r", encoding="utf-8") as f:
                    index = json.loads(f.read())

                all_shards = set(index["weight_map"].values())
                shards_found = set(f for f in os.listdir(tmp_dir) if f.endswith(".bin"))
                self.assertSetEqual(all_shards, shards_found)

                # Finally, check the model can be reloaded
                new_model = BertModel.from_pretrained(tmp_dir)
                for p1, p2 in zip(model.parameters(), new_model.parameters()):
                    self.assertTrue(torch.allclose(p1, p2))

    def test_checkpoint_sharding_from_hub(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded")
        # the model above is the same as the model below, just a sharded version.
        ref_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
        for p1, p2 in zip(model.parameters(), ref_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

2946
    @require_accelerate
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
    def test_from_pretrained_low_cpu_mem_usage_functional(self):
        # test that we can use `from_pretrained(..., low_cpu_mem_usage=True)` with normal and
        # sharded models

        mnames = [
            "hf-internal-testing/tiny-random-bert-sharded",
            "hf-internal-testing/tiny-random-bert",
        ]
        for mname in mnames:
            _ = BertModel.from_pretrained(mname, low_cpu_mem_usage=True)

    @require_usr_bin_time
2959
    @require_accelerate
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
    def test_from_pretrained_low_cpu_mem_usage_measured(self):
        # test that `from_pretrained(..., low_cpu_mem_usage=True)` uses less cpu memory than default

        mname = "bert-base-cased"

        preamble = "from transformers import AutoModel"
        one_liner_str = f'{preamble}; AutoModel.from_pretrained("{mname}", low_cpu_mem_usage=False)'
        max_rss_normal = self.python_one_liner_max_rss(one_liner_str)
        # print(f"{max_rss_normal=}")

        one_liner_str = f'{preamble};  AutoModel.from_pretrained("{mname}", low_cpu_mem_usage=True)'
        max_rss_low_mem = self.python_one_liner_max_rss(one_liner_str)
        # print(f"{max_rss_low_mem=}")

        diff_bytes = max_rss_normal - max_rss_low_mem
        diff_percent = diff_bytes / max_rss_low_mem
        # print(f"{diff_bytes=}, {diff_percent=}")
        # ideally we would compare that the diff is close to ~1x checkpoint size in bytes, but
        # measuring cpu memory on linux is very tricky and inconsistent, so instead let's check that
        # it's at least 15% less cpu memory consumed

        self.assertGreater(
            diff_percent,
            0.15,
            "should use less CPU memory for low_cpu_mem_usage=True, "
            f"but got max_rss_normal={max_rss_normal} and max_rss_low_mem={max_rss_low_mem}",
        )

        # if you want to compare things manually, let's first look at the size of the model in bytes
        # model = BertModel.from_pretrained(mname, low_cpu_mem_usage=False)
        # total_numel = sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values())
        # total_bytes = total_numel * 4  # 420MB
        # Now the diff_bytes should be very close to total_bytes, but the reports are inconsistent.
        # The easiest way to test this is to switch the model and torch.load to do all the work on
        # gpu - that way one can measure exactly the total and peak memory used. Perhaps once we add
        # functionality to load models directly on gpu, this test can be rewritten to use torch's
        # cuda memory tracking and then we should be able to do a much more precise test.

2998
    @require_accelerate
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
    @require_torch_multi_gpu
    @slow
    def test_model_parallelism_gpt2(self):
        device_map = {"transformer.wte": 0, "transformer.wpe": 0, "lm_head": 0, "transformer.ln_f": 1}
        for i in range(12):
            device_map[f"transformer.h.{i}"] = 0 if i <= 5 else 1

        model = AutoModelForCausalLM.from_pretrained("gpt2", device_map=device_map)

        tokenizer = AutoTokenizer.from_pretrained("gpt2")
        inputs = tokenizer("Hello, my name is", return_tensors="pt")
        output = model.generate(inputs["input_ids"].to(0))

        text_output = tokenizer.decode(output[0].tolist())
        self.assertEqual(text_output, "Hello, my name is John. I'm a writer, and I'm a writer. I'm")

Sylvain Gugger's avatar
Sylvain Gugger committed
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
    @require_accelerate
    @require_torch_gpu
    def test_from_pretrained_disk_offload_task_model(self):
        model = AutoModel.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        device_map = {
            "transformer.wte": 0,
            "transformer.wpe": 0,
            "transformer.h.0": "cpu",
            "transformer.h.1": "cpu",
            "transformer.h.2": "cpu",
            "transformer.h.3": "disk",
            "transformer.h.4": "disk",
            "transformer.ln_f": 0,
            "lm_head": 0,
        }
        with tempfile.TemporaryDirectory() as tmp_dir:
            inputs = torch.tensor([[1, 2, 3]]).to(0)

            model.save_pretrained(tmp_dir)
            new_model = AutoModelForCausalLM.from_pretrained(tmp_dir).to(0)
            outputs1 = new_model.to(0)(inputs)

            offload_folder = os.path.join(tmp_dir, "offload")
            new_model_with_offload = AutoModelForCausalLM.from_pretrained(
                tmp_dir, device_map=device_map, offload_folder=offload_folder
            )
            outputs2 = new_model_with_offload(inputs)

            self.assertTrue(torch.allclose(outputs1.logits.cpu(), outputs2.logits.cpu()))

            # With state dict temp offload
            offload_folder = os.path.join(tmp_dir, "offload")
            new_model_with_offload = AutoModelForCausalLM.from_pretrained(
                tmp_dir,
                device_map=device_map,
                offload_folder=offload_folder,
                offload_state_dict=True,
            )
            outputs2 = new_model_with_offload(inputs)

            self.assertTrue(torch.allclose(outputs1.logits.cpu(), outputs2.logits.cpu()))

3057
3058
3059
3060
    def test_cached_files_are_used_when_internet_is_down(self):
        # A mock response for an HTTP head request to emulate server down
        response_mock = mock.Mock()
        response_mock.status_code = 500
3061
        response_mock.headers = {}
3062
        response_mock.raise_for_status.side_effect = HTTPError
3063
        response_mock.json.return_value = {}
3064
3065
3066
3067
3068

        # Download this model to make sure it's in the cache.
        _ = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        # Under the mock environment we get a 500 error when trying to reach the model.
3069
        with mock.patch("requests.request", return_value=response_mock) as mock_head:
3070
3071
3072
3073
            _ = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
            # This check we did call the fake head request
            mock_head.assert_called()

3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
    def test_load_from_one_file(self):
        try:
            tmp_file = tempfile.mktemp()
            with open(tmp_file, "wb") as f:
                http_get(
                    "https://huggingface.co/hf-internal-testing/tiny-random-bert/resolve/main/pytorch_model.bin", f
                )

            config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
            _ = BertModel.from_pretrained(tmp_file, config=config)
        finally:
            os.remove(tmp_file)

    def test_legacy_load_from_url(self):
        # This test is for deprecated behavior and can be removed in v5
        config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
        _ = BertModel.from_pretrained(
            "https://huggingface.co/hf-internal-testing/tiny-random-bert/resolve/main/pytorch_model.bin", config=config
        )

3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
    @require_safetensors
    def test_safetensors_save_and_load(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, safe_serialization=True)
            # No pytorch_model.bin file, only a model.safetensors
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))

            new_model = BertModel.from_pretrained(tmp_dir)

            # Check models are equal
            for p1, p2 in zip(model.parameters(), new_model.parameters()):
                self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_safetensors_load_from_hub(self):
        safetensors_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors")
        pytorch_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")

        # Check models are equal
        for p1, p2 in zip(safetensors_model.parameters(), pytorch_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_safetensors_save_and_load_sharded(self):
        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, safe_serialization=True, max_shard_size="100kB")
            # No pytorch_model.bin index file, only a model.safetensors index
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_INDEX_NAME)))
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)))
            # No regular weights file
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))
            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))

            new_model = BertModel.from_pretrained(tmp_dir)

            # Check models are equal
            for p1, p2 in zip(model.parameters(), new_model.parameters()):
                self.assertTrue(torch.allclose(p1, p2))

    @require_safetensors
    def test_safetensors_load_from_hub_sharded(self):
        safetensors_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded-safetensors")
        pytorch_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded")

        # Check models are equal
        for p1, p2 in zip(safetensors_model.parameters(), pytorch_model.parameters()):
            self.assertTrue(torch.allclose(p1, p2))

Sylvain Gugger's avatar
Sylvain Gugger committed
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
    def test_base_model_to_head_model_load(self):
        base_model = BaseModel(PretrainedConfig())
        with tempfile.TemporaryDirectory() as tmp_dir:
            base_model.save_pretrained(tmp_dir)

            # Can load a base model in a model with head
            model = ModelWithHead.from_pretrained(tmp_dir)
            for p1, p2 in zip(model.base.parameters(), base_model.parameters()):
                self.assertTrue(torch.allclose(p1, p2))

            # It doesn't work if the state dict has a mix of keys of the head and base without prefix though.
            base_state_dict = base_model.state_dict()
            head_state_dict = model.state_dict()
            base_state_dict["linear2.weight"] = head_state_dict["linear2.weight"]
            base_state_dict["linear2.bias"] = head_state_dict["linear2.bias"]
            torch.save(base_state_dict, os.path.join(tmp_dir, WEIGHTS_NAME))

            with self.assertRaisesRegex(
                ValueError, "The state dictionary of the model you are trying to load is corrupted."
            ):
                _ = ModelWithHead.from_pretrained(tmp_dir)

Sylvain Gugger's avatar
Sylvain Gugger committed
3167
3168
3169
3170
3171
3172

@require_torch
@is_staging_test
class ModelPushToHubTester(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
3173
3174
3175
        cls._token = TOKEN
        set_access_token(TOKEN)
        HfFolder.save_token(TOKEN)
Sylvain Gugger's avatar
Sylvain Gugger committed
3176
3177
3178
3179

    @classmethod
    def tearDownClass(cls):
        try:
3180
            delete_repo(token=cls._token, repo_id="test-model")
Sylvain Gugger's avatar
Sylvain Gugger committed
3181
3182
3183
3184
        except HTTPError:
            pass

        try:
3185
            delete_repo(token=cls._token, repo_id="valid_org/test-model-org")
Sylvain Gugger's avatar
Sylvain Gugger committed
3186
3187
3188
        except HTTPError:
            pass

3189
        try:
3190
            delete_repo(token=cls._token, repo_id="test-dynamic-model")
3191
3192
3193
        except HTTPError:
            pass

Sylvain Gugger's avatar
Sylvain Gugger committed
3194
3195
3196
3197
3198
    def test_push_to_hub(self):
        config = BertConfig(
            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
        )
        model = BertModel(config)
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
        model.push_to_hub("test-model", use_auth_token=self._token)

        new_model = BertModel.from_pretrained(f"{USER}/test-model")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=self._token, repo_id="test-model")

        # Push to hub via save_pretrained
Sylvain Gugger's avatar
Sylvain Gugger committed
3209
        with tempfile.TemporaryDirectory() as tmp_dir:
3210
            model.save_pretrained(tmp_dir, repo_id="test-model", push_to_hub=True, use_auth_token=self._token)
Sylvain Gugger's avatar
Sylvain Gugger committed
3211

3212
3213
3214
        new_model = BertModel.from_pretrained(f"{USER}/test-model")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))
Sylvain Gugger's avatar
Sylvain Gugger committed
3215
3216
3217
3218
3219
3220

    def test_push_to_hub_in_organization(self):
        config = BertConfig(
            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
        )
        model = BertModel(config)
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
        model.push_to_hub("valid_org/test-model-org", use_auth_token=self._token)

        new_model = BertModel.from_pretrained("valid_org/test-model-org")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=self._token, repo_id="valid_org/test-model-org")

        # Push to hub via save_pretrained
Sylvain Gugger's avatar
Sylvain Gugger committed
3231
3232
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(
3233
                tmp_dir, push_to_hub=True, use_auth_token=self._token, repo_id="valid_org/test-model-org"
Sylvain Gugger's avatar
Sylvain Gugger committed
3234
3235
            )

3236
3237
3238
        new_model = BertModel.from_pretrained("valid_org/test-model-org")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))
3239
3240

    def test_push_to_hub_dynamic_model(self):
3241
3242
3243
3244
3245
        CustomConfig.register_for_auto_class()
        CustomModel.register_for_auto_class()

        config = CustomConfig(hidden_size=32)
        model = CustomModel(config)
3246

3247
3248
3249
3250
3251
3252
        model.push_to_hub("test-dynamic-model", use_auth_token=self._token)
        # checks
        self.assertDictEqual(
            config.auto_map,
            {"AutoConfig": "custom_configuration.CustomConfig", "AutoModel": "custom_modeling.CustomModel"},
        )
3253
3254

        new_model = AutoModel.from_pretrained(f"{USER}/test-dynamic-model", trust_remote_code=True)
3255
3256
        # Can't make an isinstance check because the new_model is from the CustomModel class of a dynamic module
        self.assertEqual(new_model.__class__.__name__, "CustomModel")
3257
3258
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))
3259

3260
        config = AutoConfig.from_pretrained(f"{USER}/test-dynamic-model", trust_remote_code=True)
3261
        new_model = AutoModel.from_config(config, trust_remote_code=True)
3262
        self.assertEqual(new_model.__class__.__name__, "CustomModel")