test_modeling_common.py 124 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.
15

Sylvain Gugger's avatar
Sylvain Gugger committed
16
import collections
17
import copy
18
import gc
19
import inspect
20
import os
21
import os.path
22
import pickle
Aymeric Augustin's avatar
Aymeric Augustin committed
23
import random
Sylvain Gugger's avatar
Sylvain Gugger committed
24
import re
25
import tempfile
26
import warnings
27
from collections import defaultdict
NielsRogge's avatar
NielsRogge committed
28
from typing import Dict, List, Tuple
thomwolf's avatar
thomwolf committed
29

30
import numpy as np
31
from pytest import mark
32
33

import transformers
34
35
36
37
38
39
40
from transformers import (
    AutoModel,
    AutoModelForSequenceClassification,
    PretrainedConfig,
    is_torch_available,
    logging,
)
41
from transformers.models.auto import get_values
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES,
    MODEL_FOR_BACKBONE_MAPPING_NAMES,
    MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES,
    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES,
    MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES,
    MODEL_FOR_MASKED_LM_MAPPING_NAMES,
    MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES,
    MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES,
    MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES,
    MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
    MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES,
    MODEL_MAPPING_NAMES,
)
Sylvain Gugger's avatar
Sylvain Gugger committed
62
63
from transformers.testing_utils import (
    CaptureLogger,
64
65
    is_pt_flax_cross_test,
    is_pt_tf_cross_test,
66
    require_accelerate,
67
    require_safetensors,
Sylvain Gugger's avatar
Sylvain Gugger committed
68
    require_torch,
69
    require_torch_gpu,
Sylvain Gugger's avatar
Sylvain Gugger committed
70
71
72
73
    require_torch_multi_gpu,
    slow,
    torch_device,
)
74
from transformers.utils import (
75
76
    CONFIG_NAME,
    GENERATION_CONFIG_NAME,
77
    WEIGHTS_NAME,
78
    is_accelerate_available,
79
80
81
82
83
    is_flax_available,
    is_tf_available,
    is_torch_fx_available,
)
from transformers.utils.generic import ModelOutput
84

Aymeric Augustin's avatar
Aymeric Augustin committed
85

86
87
88
89
if is_accelerate_available():
    from accelerate.utils import compute_module_sizes


90
if is_torch_available():
91
    import torch
92
    from torch import nn
thomwolf's avatar
thomwolf committed
93

94
    from transformers import MODEL_MAPPING, AdaptiveEmbedding
Sylvain Gugger's avatar
Sylvain Gugger committed
95
    from transformers.pytorch_utils import id_tensor_storage
thomwolf's avatar
thomwolf committed
96

Sylvain Gugger's avatar
Sylvain Gugger committed
97

98
99
100
if is_tf_available():
    import tensorflow as tf

101
102
if is_flax_available():
    import jax.numpy as jnp
103

104
105
106
107
108
    from transformers.modeling_flax_pytorch_utils import (
        convert_pytorch_state_dict_to_flax,
        load_flax_weights_in_pytorch_model,
    )

109
if is_torch_fx_available():
110
    from transformers.utils.fx import symbolic_trace
111

112

113
114
115
def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
116
        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
117
            setattr(configs_no_init, key, 1e-10)
118
119
120
        if isinstance(getattr(configs_no_init, key, None), PretrainedConfig):
            no_init_subconfig = _config_zero_init(getattr(configs_no_init, key))
            setattr(configs_no_init, key, no_init_subconfig)
121
122
    return configs_no_init

thomwolf's avatar
thomwolf 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
def _mock_init_weights(self, module):
    for name, param in module.named_parameters(recurse=False):
        # Use the first letter of the name to get a value and go from a <> -13 to z <> 12
        value = ord(name[0].lower()) - 110
        param.data.fill_(value)


def _mock_all_init_weights(self):
    # Prune heads if needed
    if self.config.pruned_heads:
        self.prune_heads(self.config.pruned_heads)

    import transformers.modeling_utils

    if transformers.modeling_utils._init_weights:
        for module in self.modules():
            module._is_hf_initialized = False
        # Initialize weights
        self.apply(self._initialize_weights)

        # Tie weights should be skipped when not initializing all weights
        # since from_pretrained(...) calls tie weights anyways
        self.tie_weights()


149
150
151
152
@require_torch
class ModelTesterMixin:
    model_tester = None
    all_model_classes = ()
153
    all_generative_model_classes = ()
154
    fx_compatible = False
Patrick von Platen's avatar
Patrick von Platen committed
155
156
157
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
158
    test_resize_position_embeddings = False
Patrick von Platen's avatar
Patrick von Platen committed
159
    test_head_masking = True
160
    test_mismatched_shapes = True
161
    test_missing_keys = True
162
    test_model_parallel = False
163
    is_encoder_decoder = False
164
    has_attentions = True
165
    model_split_percents = [0.5, 0.7, 0.9]
166

167
168
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
169
        if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
170
            inputs_dict = {
171
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
172
                if isinstance(v, torch.Tensor) and v.ndim > 1
Sylvain Gugger's avatar
Sylvain Gugger committed
173
                else v
174
175
                for k, v in inputs_dict.items()
            }
176
        elif model_class.__name__ in get_values(MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES):
177
            inputs_dict.pop("attention_mask")
178
179

        if return_labels:
180
            if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
181
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
182
183
184
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
185
            ]:
186
187
188
189
190
191
                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
                )
192
193
194
195
196
197
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES),
198
            ]:
199
200
201
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
202
203
204
205
206
207
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_CAUSAL_LM_MAPPING_NAMES),
                *get_values(MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES),
                *get_values(MODEL_FOR_MASKED_LM_MAPPING_NAMES),
                *get_values(MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES),
208
209
210
211
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
212
            elif model_class.__name__ in get_values(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
213
214
215
216
                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
                )
217
            elif model_class.__name__ in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
218
219
220
221
                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()
222

223
224
        return inputs_dict

Patrick von Platen's avatar
Patrick von Platen committed
225
    def test_save_load(self):
226
227
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

228
229
230
231
232
233
234
235
236
237
        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)

238
239
240
241
242
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
243
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
244

245
            with tempfile.TemporaryDirectory() as tmpdirname:
246
                model.save_pretrained(tmpdirname)
247
248
249
250
251
252
253

                # the config file (and the generation config file, if it can generate) should be saved
                self.assertTrue(os.path.exists(os.path.join(tmpdirname, CONFIG_NAME)))
                self.assertEqual(
                    model.can_generate(), os.path.exists(os.path.join(tmpdirname, GENERATION_CONFIG_NAME))
                )

254
                model = model_class.from_pretrained(tmpdirname)
255
                model.to(torch_device)
256
                with torch.no_grad():
257
                    second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
thomwolf's avatar
thomwolf committed
258

259
260
261
262
263
            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)
264

265
266
267
268
269
270
271
272
273
274
275
276
    def test_from_pretrained_no_checkpoint(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
        for model_class in self.all_model_classes:
            model = model_class(config)
            state_dict = model.state_dict()

            new_model = model_class.from_pretrained(
                pretrained_model_name_or_path=None, config=config, state_dict=state_dict
            )
            for p1, p2 in zip(model.parameters(), new_model.parameters()):
                self.assertTrue(torch.equal(p1, p2))

277
    def test_save_load_keys_to_ignore_on_save(self):
278
279
280
281
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
282
283
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
284
285
286
                continue

            # check the keys are in the original state_dict
287
            for k in _keys_to_ignore_on_save:
288
                self.assertIn(k, model.state_dict().keys(), "\n".join(model.state_dict().keys()))
289
290
291
292
293
294

            # 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)
295
                for k in _keys_to_ignore_on_save:
296
                    self.assertNotIn(k, state_dict_saved.keys(), "\n".join(state_dict_saved.keys()))
297

Sylvain Gugger's avatar
Sylvain Gugger committed
298
299
300
                # 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(
301
302
                    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
303
304
305
                )
                self.assertTrue(len(load_result.unexpected_keys) == 0)

306
307
308
309
310
311
312
313
314
315
316
    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)

317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
    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)

336
337
    def test_save_load_fast_init_from_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
338
339
        if config.__class__ not in MODEL_MAPPING:
            return
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
        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
361
362
            model_class_copy._init_weights = _mock_init_weights
            model_class_copy.init_weights = _mock_all_init_weights
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378

            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)
379
                # Before we test anything
380
381

                for key in model_fast_init.state_dict().keys():
382
383
384
385
386
                    if isinstance(model_slow_init.state_dict()[key], torch.BoolTensor):
                        max_diff = (model_slow_init.state_dict()[key] ^ model_fast_init.state_dict()[key]).sum().item()
                    else:
                        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")
387
388
389

    def test_save_load_fast_init_to_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
390
391
        if config.__class__ not in MODEL_MAPPING:
            return
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
        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
413
414
            base_class_copy._init_weights = _mock_init_weights
            base_class_copy.init_weights = _mock_all_init_weights
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432

            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():
433
434
435
436
437
438
439
440
441
                    if isinstance(model_slow_init.state_dict()[key], torch.BoolTensor):
                        max_diff = torch.max(
                            model_slow_init.state_dict()[key] ^ model_fast_init.state_dict()[key]
                        ).item()
                    else:
                        max_diff = torch.max(
                            torch.abs(model_slow_init.state_dict()[key] - model_fast_init.state_dict()[key])
                        ).item()
                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")
442

Patrick von Platen's avatar
Patrick von Platen committed
443
    def test_initialization(self):
444
445
446
447
448
449
450
451
        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
452
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
453
                        [0.0, 1.0],
454
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
455
                    )
thomwolf's avatar
thomwolf committed
456

Patrick von Platen's avatar
Patrick von Platen committed
457
    def test_determinism(self):
458
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
459
460
461
462
463
464
465
466
467

        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)

468
469
470
471
472
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
473
474
                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
475

476
477
478
479
480
            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_determinism(tensor1, tensor2)
            else:
                check_determinism(first, second)
481

482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
    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",
                ]
498
                expected_arg_names.extend(
499
500
                    ["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
501
502
503
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
504
505
506
507
            else:
                expected_arg_names = ["input_ids"]
                self.assertListEqual(arg_names[:1], expected_arg_names)

508
509
510
511
512
    def test_training(self):
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
513
514
515
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.return_dict = True

516
517
518
            if model_class.__name__ in [
                *get_values(MODEL_MAPPING_NAMES),
                *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES),
519
            ]:
520
                continue
521

522
523
524
525
526
527
528
529
            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):
530
        if not self.model_tester.is_training:
531
532
533
            return

        for model_class in self.all_model_classes:
534
535
536
537
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.use_cache = False
            config.return_dict = True

538
            if (
539
540
                model_class.__name__
                in [*get_values(MODEL_MAPPING_NAMES), *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)]
541
542
                or not model_class.supports_gradient_checkpointing
            ):
543
544
545
                continue
            model = model_class(config)
            model.to(torch_device)
546
            model.gradient_checkpointing_enable()
547
548
549
550
551
            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
552
    def test_attention_outputs(self):
553
554
555
        if not self.has_attentions:
            self.skipTest(reason="Model does not output attentions")

556
557
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True
558

559
560
561
562
563
564
565
566
567
568
569
570
        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
571
            config.return_dict = True
572
573
574
575
576
577
578
            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)
579

580
581
582
583
584
585
586
587
588
589
            # 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
590

591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
            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
610
611
612
                if model_class.__name__ in [
                    *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                    *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
613
                ]:
614
615
616
617
618
619
620
621
622
623
624
625
626
627
                    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],
                )
628

629
630
631
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
                # 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
672

673
    @slow
674
    def test_torchscript_simple(self):
675
676
        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
677

678
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
679
    def test_torchscript_output_attentions(self):
680
681
682
        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
683

684
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
685
    def test_torchscript_output_hidden_state(self):
686
687
688
        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
689

690
691
692
693
    # 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()
694
695
696
        # torch 1.8 has no `_clear_class_state` in `torch.jit._state`
        if hasattr(torch.jit._state, "_clear_class_state"):
            torch.jit._state._clear_class_state()
697

698
    def _create_and_check_torchscript(self, config, inputs_dict):
Patrick von Platen's avatar
Patrick von Platen committed
699
        if not self.test_torchscript:
700
            return
701

702
703
704
705
706
707
        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()
708
            inputs = self._prepare_for_class(inputs_dict, model_class)
thomwolf's avatar
thomwolf committed
709

710
711
            main_input_name = model_class.main_input_name

712
            try:
713
                if model.config.is_encoder_decoder:
714
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
715
                    main_input = inputs[main_input_name]
716
717
718
                    attention_mask = inputs["attention_mask"]
                    decoder_input_ids = inputs["decoder_input_ids"]
                    decoder_attention_mask = inputs["decoder_attention_mask"]
719
                    model(main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
720
                    traced_model = torch.jit.trace(
721
                        model, (main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
722
                    )
723
724
725
726
                elif "bbox" in inputs and "image" in inputs:  # LayoutLMv2 requires additional inputs
                    input_ids = inputs["input_ids"]
                    bbox = inputs["bbox"]
                    image = inputs["image"].tensor
727
                    model(input_ids, bbox, image)
728
729
730
                    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
731
                else:
732
                    main_input = inputs[main_input_name]
733
                    model(main_input)
734
                    traced_model = torch.jit.trace(model, main_input)
735
736
            except RuntimeError:
                self.fail("Couldn't trace module.")
thomwolf's avatar
thomwolf committed
737

738
            with tempfile.TemporaryDirectory() as tmp_dir_name:
739
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
thomwolf's avatar
thomwolf committed
740

741
                try:
742
                    torch.jit.save(traced_model, pt_file_name)
743
744
                except Exception:
                    self.fail("Couldn't save module.")
thomwolf's avatar
thomwolf committed
745

746
747
748
749
                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")
LysandreJik's avatar
LysandreJik committed
750

751
752
            model.to(torch_device)
            model.eval()
thomwolf's avatar
thomwolf committed
753

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

757
758
759
            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

760
761
762
763
764
765
766
767
768
            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
            }

769
            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
thomwolf's avatar
thomwolf committed
770

771
772
773
774
775
776
777
778
779
780
781
            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)

782
            models_equal = True
783
            for layer_name, p1 in model_state_dict.items():
784
785
786
787
                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
788

789
            self.assertTrue(models_equal)
thomwolf's avatar
thomwolf committed
790

791
792
793
794
            # 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()

795
796
797
798
799
800
801
802
    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)

803
804
    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:
805
806
807
808
809
            return

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

810
        for model_class in self.all_model_classes:
811
812
813
814
815
816
817
818
819
            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)
820
821
822
                    input_names = [
                        "attention_mask",
                        "decoder_attention_mask",
823
                        "decoder_input_ids",
824
                        "input_features",
825
826
                        "input_ids",
                        "input_values",
827
                    ]
828
829
                    if labels is not None:
                        input_names.append("labels")
830

831
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
832
                    input_names = list(filtered_inputs.keys())
833

834
                    model_output = model(**filtered_inputs)
835

836
                    traced_model = symbolic_trace(model, input_names)
837
                    traced_output = traced_model(**filtered_inputs)
838
                else:
839
840
841
842
                    input_names = [
                        "attention_mask",
                        "bbox",
                        "input_features",
843
844
845
846
847
848
                        "input_ids",
                        "input_values",
                        "pixel_values",
                        "token_type_ids",
                        "visual_feats",
                        "visual_pos",
849
                    ]
850

851
                    labels = inputs.get("labels", None)
852
853
                    start_positions = inputs.get("start_positions", None)
                    end_positions = inputs.get("end_positions", None)
854
855
                    if labels is not None:
                        input_names.append("labels")
856
857
858
859
                    if start_positions is not None:
                        input_names.append("start_positions")
                    if end_positions is not None:
                        input_names.append("end_positions")
860

861
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
862
                    input_names = list(filtered_inputs.keys())
863

864
                    if model.__class__.__name__ in set(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES.values()) and (
865
                        not hasattr(model.config, "problem_type") or model.config.problem_type is None
866
867
868
                    ):
                        model.config.problem_type = "single_label_classification"

869
                    traced_model = symbolic_trace(model, input_names)
870
                    traced_output = traced_model(**filtered_inputs)
871
                    model_output = model(**filtered_inputs)
872

873
            except Exception as e:
874
                self.fail(f"Couldn't trace module: {e}")
875

876
877
878
879
880
881
882
883
884
885
886
887
888
            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)
889
            num_outputs = len(model_output)
890
891
892
893
894
895

            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}",
                )
896

897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
            # 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}",
                    )

917
918
919
920
            # 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
921
922
    def test_headmasking(self):
        if not self.test_head_masking:
923
            return
924

925
926
927
        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()
LysandreJik's avatar
LysandreJik committed
928

929
        inputs_dict["output_attentions"] = True
930
931
932
933
934
935
        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
936

937
938
939
            # 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
940
941
942
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
943
944
945
946
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
947
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
948
            inputs["head_mask"] = head_mask
949
950
951
952
953
            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
954
955
                if "cross_attn_head_mask" in arg_names:
                    inputs["cross_attn_head_mask"] = head_mask
956
            outputs = model(**inputs, return_dict=True)
957
958
959
960
961
962
963
964
965

            # 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)
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986

            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)
987
                check_attentions_validity(outputs.cross_attentions)
988
989
            else:
                check_attentions_validity(outputs.attentions)
990

Patrick von Platen's avatar
Patrick von Platen committed
991
992
    def test_head_pruning(self):
        if not self.test_pruning:
993
994
995
            return

        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
996
997
998
999
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1000

1001
1002
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1003

1004
            inputs_dict["output_attentions"] = True
1005
1006
1007
1008
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1009
1010
1011
1012
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1013
1014
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
1015
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1016

1017
            attentions = outputs[-1]
1018

1019
            self.assertEqual(attentions[0].shape[-3], 1)
1020
1021
            # TODO: To have this check, we will need at least 3 layers. Do we really need it?
            # self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
1022
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
LysandreJik's avatar
LysandreJik committed
1023

Patrick von Platen's avatar
Patrick von Platen committed
1024
1025
    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
1026
            return
LysandreJik's avatar
LysandreJik committed
1027

1028
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1029
1030
1031
1032
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1033
1034
1035

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1036

1037
            inputs_dict["output_attentions"] = True
1038
1039
1040
1041
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1042
1043
1044
1045
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1046
            model.prune_heads(heads_to_prune)
1047

1048
            with tempfile.TemporaryDirectory() as temp_dir_name:
1049
1050
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1051
                model.to(torch_device)
1052

1053
            with torch.no_grad():
1054
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1055
1056
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
1057
1058
            # TODO: To have this check, we will need at least 3 layers. Do we really need it?
            # self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
1059
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1060

Patrick von Platen's avatar
Patrick von Platen committed
1061
1062
    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
1063
            return
1064

1065
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1066
1067
1068
1069
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1070

1071
1072
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1073

1074
            inputs_dict["output_attentions"] = True
1075
            config.output_hidden_states = False
1076

1077
1078
1079
1080
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1081
            config.pruned_heads = heads_to_prune
1082

1083
1084
1085
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1086

1087
            with torch.no_grad():
1088
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1089
            attentions = outputs[-1]
1090

1091
            self.assertEqual(attentions[0].shape[-3], 1)
1092
1093
            # TODO: To have this check, we will need at least 3 layers. Do we really need it?
            # self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
1094
            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
            heads_to_prune = {1: [1, 2]}
1113
            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
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1124
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1125

1126
            with tempfile.TemporaryDirectory() as temp_dir_name:
1127
1128
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1129
                model.to(torch_device)
thomwolf's avatar
thomwolf committed
1130

1131
            with torch.no_grad():
1132
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1133
            attentions = outputs[-1]
LysandreJik's avatar
LysandreJik committed
1134

1135
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1136
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1137

1138
            heads_to_prune = {0: [0], 1: [1, 2]}
1139
            model.prune_heads(heads_to_prune)
1140

1141
            with torch.no_grad():
1142
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1143
            attentions = outputs[-1]
1144

1145
1146
            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)
1147

1148
            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2]})
thomwolf's avatar
thomwolf committed
1149

Patrick von Platen's avatar
Patrick von Platen committed
1150
    def test_hidden_states_output(self):
Joseph Liu's avatar
Joseph Liu committed
1151
        def check_hidden_states_output(inputs_dict, config, model_class):
1152
            model = model_class(config)
1153
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
1154
            model.eval()
Joseph Liu's avatar
Joseph Liu committed
1155

thomwolf's avatar
thomwolf committed
1156
            with torch.no_grad():
1157
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1158
1159

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
1160

Sylvain Gugger's avatar
Sylvain Gugger committed
1161
1162
1163
1164
            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)
1165

Patrick von Platen's avatar
Patrick von Platen committed
1166
1167
1168
1169
1170
1171
1172
            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

1173
            self.assertListEqual(
Lysandre's avatar
Lysandre committed
1174
1175
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
1176
            )
thomwolf's avatar
thomwolf committed
1177

1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
            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
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
        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)

1203
1204
1205
    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
1206
        config.output_attentions = self.has_attentions
1207
1208
1209
1210
1211
1212
1213
1214
1215

        # 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)
1216

1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
        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()

1227
1228
1229
1230
1231
1232
1233
1234
1235
            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()
1236
1237
1238
1239
1240

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

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)
1241
1242
1243
1244
1245

            if self.has_attentions:
                self.assertIsNotNone(encoder_attentions.grad)
                self.assertIsNotNone(decoder_attentions.grad)
                self.assertIsNotNone(cross_attentions.grad)
1246
1247
1248
1249
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            hidden_states.retain_grad()
1250
1251
1252
1253

            if self.has_attentions:
                attentions = outputs.attentions[0]
                attentions.retain_grad()
1254
1255
1256
1257

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

            self.assertIsNotNone(hidden_states.grad)
1258
1259
1260

            if self.has_attentions:
                self.assertIsNotNone(attentions.grad)
1261

Pradhy729's avatar
Pradhy729 committed
1262
    def test_feed_forward_chunking(self):
Lysandre's avatar
Lysandre committed
1263
1264
1265
1266
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Pradhy729's avatar
Pradhy729 committed
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
        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))

1285
1286
1287
1288
1289
1290
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
    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
1364
    def test_resize_tokens_embeddings(self):
Lysandre's avatar
Lysandre committed
1365
1366
1367
1368
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Patrick von Platen's avatar
Patrick von Platen committed
1369
        if not self.test_resize_embeddings:
1370
1371
1372
1373
1374
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
1375
            model.to(torch_device)
1376

Patrick von Platen's avatar
Patrick von Platen committed
1377
1378
1379
            if self.model_tester.is_training is False:
                model.eval()

1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
            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)
1390
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
1391
            model(**self._prepare_for_class(inputs_dict, model_class))
1392
1393
1394
1395
1396
1397
1398

            # 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)

1399
1400
1401
            # 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)
1402
1403
1404
1405

            # 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)
1406
            model(**self._prepare_for_class(inputs_dict, model_class))
1407

1408
1409
1410
1411
1412
1413
1414
1415
            # 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)

1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)

            model_vocab_size = config.vocab_size
            model.resize_token_embeddings(model_vocab_size + 10, pad_to_multiple_of=1)
            self.assertTrue(model.config.vocab_size + 10, model_vocab_size)

            model_embed = model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=64)
            self.assertTrue(model_embed.weight.shape[0] // 64, 0)

            model_embed = model.resize_token_embeddings(model_vocab_size + 13, pad_to_multiple_of=64)
            self.assertTrue(model_embed.weight.shape[0] // 64, 0)

            with self.assertRaisesRegex(
                ValueError,
                "Asking to pad the embedding matrix to a multiple of `1.3`, which is not and integer. Please make sure to pass an integer",
            ):
                model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=1.3)

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
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
    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
1487
    def test_model_common_attributes(self):
1488
1489
1490
1491
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1492
1493
            self.assertIsInstance(model.get_input_embeddings(), (nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(nn.Embedding(10, 10))
1494
            x = model.get_output_embeddings()
1495
            self.assertTrue(x is None or isinstance(x, nn.Linear))
1496

1497
1498
1499
1500
1501
1502
1503
    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)

1504
    def test_correct_missing_keys(self):
1505
1506
        if not self.test_missing_keys:
            return
1507
1508
1509
1510
1511
1512
1513
        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):
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
                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

1527
1528
1529
                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)
1530
                    self.assertGreater(len(loading_info["missing_keys"]), 0, model.__class__.__name__)
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
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
    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))

1580
1581
    @require_safetensors
    def test_can_use_safetensors(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
1582
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
        for model_class in self.all_model_classes:
            model_tied = model_class(config)
            with tempfile.TemporaryDirectory() as d:
                try:
                    model_tied.save_pretrained(d, safe_serialization=True)
                except Exception as e:
                    raise Exception(f"Class {model_class.__name__} cannot be saved using safetensors: {e}")

                model_reloaded, infos = model_class.from_pretrained(d, output_loading_info=True)
                # Checking the state dicts are correct
                reloaded_state = model_reloaded.state_dict()
                for k, v in model_tied.state_dict().items():
                    self.assertIn(k, reloaded_state, f"Key {k} is missing from reloaded")
                    torch.testing.assert_close(
                        v, reloaded_state[k], msg=lambda x: f"{model_class.__name__}: Tensor {k}: {x}"
                    )
Sylvain Gugger's avatar
Sylvain Gugger committed
1599
1600
                # Checking there was no complain of missing weights
                self.assertEqual(infos["missing_keys"], [])
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616

                # Checking the tensor sharing are correct
                ptrs = defaultdict(list)
                for k, v in model_tied.state_dict().items():
                    ptrs[v.data_ptr()].append(k)

                shared_ptrs = {k: v for k, v in ptrs.items() if len(v) > 1}

                for _, shared_names in shared_ptrs.items():
                    reloaded_ptrs = {reloaded_state[k].data_ptr() for k in shared_names}
                    self.assertEqual(
                        len(reloaded_ptrs),
                        1,
                        f"The shared pointers are incorrect, found different pointers for keys {shared_names}",
                    )

Sylvain Gugger's avatar
Sylvain Gugger committed
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
    def test_load_save_without_tied_weights(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
        config.tie_word_embeddings = False
        for model_class in self.all_model_classes:
            model = model_class(config)
            with tempfile.TemporaryDirectory() as d:
                model.save_pretrained(d)

                model_reloaded, infos = model_class.from_pretrained(d, output_loading_info=True)
                # Checking the state dicts are correct
                reloaded_state = model_reloaded.state_dict()
                for k, v in model.state_dict().items():
                    self.assertIn(k, reloaded_state, f"Key {k} is missing from reloaded")
                    torch.testing.assert_close(
                        v, reloaded_state[k], msg=lambda x: f"{model_class.__name__}: Tensor {k}: {x}"
                    )
                # Checking there was no complain of missing weights
                self.assertEqual(infos["missing_keys"], [])

Sylvain Gugger's avatar
Sylvain Gugger committed
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
    def test_tied_weights_keys(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
        config.tie_word_embeddings = True
        for model_class in self.all_model_classes:
            model_tied = model_class(config)

            ptrs = collections.defaultdict(list)
            for name, tensor in model_tied.state_dict().items():
                ptrs[id_tensor_storage(tensor)].append(name)

            # These are all the pointers of shared tensors.
            tied_params = [names for _, names in ptrs.items() if len(names) > 1]

            tied_weight_keys = model_tied._tied_weights_keys if model_tied._tied_weights_keys is not None else []
            # Detect we get a hit for each key
            for key in tied_weight_keys:
                if not any(re.search(key, p) for group in tied_params for p in group):
                    raise ValueError(f"{key} is not a tied weight key for {model_class}.")

            # Removed tied weights found from tied params -> there should only be one left after
            for key in tied_weight_keys:
                for i in range(len(tied_params)):
                    tied_params[i] = [p for p in tied_params[i] if re.search(key, p) is None]

            tied_params = [group for group in tied_params if len(group) > 1]
Sylvain Gugger's avatar
Sylvain Gugger committed
1661
1662
1663
1664
1665
            self.assertListEqual(
                tied_params,
                [],
                f"Missing `_tied_weights_keys` for {model_class}: add all of {tied_params} except one.",
            )
Sylvain Gugger's avatar
Sylvain Gugger committed
1666

Sylvain Gugger's avatar
Sylvain Gugger committed
1667
1668
    def test_model_weights_reload_no_missing_tied_weights(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1669
        for model_class in self.all_model_classes:
Sylvain Gugger's avatar
Sylvain Gugger committed
1670
1671
1672
            model = model_class(config)
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.save_pretrained(tmp_dir)
1673
1674
1675

                # 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.
Sylvain Gugger's avatar
Sylvain Gugger committed
1676
                with open(os.path.join(tmp_dir, "pytorch_model.bin"), "wb") as f:
1677
                    torch.save({}, f)
Sylvain Gugger's avatar
Sylvain Gugger committed
1678
                model_reloaded, infos = model_class.from_pretrained(tmp_dir, output_loading_info=True)
1679
1680
1681
1682

                prefix = f"{model_reloaded.base_model_prefix}."
                params = dict(model_reloaded.named_parameters())
                params.update(dict(model_reloaded.named_buffers()))
1683
                param_names = {k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys()}
1684
1685
1686
1687

                missing_keys = set(infos["missing_keys"])

                extra_missing = missing_keys - param_names
Sylvain Gugger's avatar
Sylvain Gugger committed
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
                # Remove tied weights from extra missing: they are normally not warned as missing if their tied
                # counterpart is present but here there are no weights at all so we do get the warning.
                ptrs = collections.defaultdict(list)
                for name, tensor in model_reloaded.state_dict().items():
                    ptrs[id_tensor_storage(tensor)].append(name)
                tied_params = [names for _, names in ptrs.items() if len(names) > 1]
                for group in tied_params:
                    group = {k[len(prefix) :] if k.startswith(prefix) else k for k in group}
                    # We remove the group from extra_missing if not all weights from group are in it
                    if len(group - extra_missing) > 0:
                        extra_missing = extra_missing - set(group)
1699
1700
1701
1702

                self.assertEqual(
                    extra_missing,
                    set(),
Sylvain Gugger's avatar
Sylvain Gugger committed
1703
1704
                    f"This model {model_class.__name__} might be missing some `keys_to_ignore`: {extra_missing}. "
                    f"For debugging, tied parameters are {tied_params}",
1705
1706
                )

Sylvain Gugger's avatar
Sylvain Gugger committed
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
                missed_missing = param_names - missing_keys
                # Remove nonpersistent buffers from missed_missing
                buffers = [n for n, _ in model_reloaded.named_buffers()]
                nonpersistent_buffers = {n for n in buffers if n not in model_reloaded.state_dict()}
                nonpersistent_buffers = {
                    k[len(prefix) :] if k.startswith(prefix) else k for k in nonpersistent_buffers
                }
                missed_missing = missed_missing - nonpersistent_buffers

                if model_reloaded._keys_to_ignore_on_load_missing is None:
                    expected_missing = set()
                else:
                    expected_missing = set(model_reloaded._keys_to_ignore_on_load_missing)
                self.assertEqual(
                    missed_missing,
                    expected_missing,
                    f"This model {model_class.__name__} ignores keys {missed_missing} but they look like real"
                    " parameters. If they are non persistent buffers make sure to instantiate them with"
                    " `persistent=False`",
                )
1727

1728
1729
1730
    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
1731
1732
1733
1734
        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

1735
1736
1737
1738
1739
1740
1741
1742
1743
        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
1744
1745
1746
1747
1748
                    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)
1749
1750
1751
1752
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
Sam Shleifer's avatar
Sam Shleifer committed
1753
1754
1755
                            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
1756
1757
1758
1759
1760
1761
                            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)}."
                            ),
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
                        )

                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})

1787
1788
1789
1790
            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})
1791

1792
1793
1794
1795
1796
1797
1798
1799
1800
                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}
                )
1801

1802
1803
1804
1805
    # 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"""
1806

1807
1808
1809
        for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
            if k in inputs_dict:
                attention_mask = inputs_dict[k]
1810

1811
1812
1813
1814
1815
1816
                # 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
                )
1817

1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
                # 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)"""

1835
1836
        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}
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863

        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):
1864
        """Check the outputs from PyTorch and TensorFlow models are close enough. Checks are done in a recursive way.
1865

1866
1867
1868
1869
1870
1871
1872
1873
        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.
        """
1874

1875
1876
1877
        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")
1878

1879
1880
1881
1882
1883
1884
        # 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",
            )
1885

1886
1887
1888
            # 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)
1889

1890
1891
            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]
1892

1893
            self.assertEqual(tf_keys, pt_keys, f"{name}: Output keys differ between TF and PyTorch")
1894

1895
            # convert to the case of `tuple`
1896
            # appending each key to the current (string) `name`
1897
1898
1899
1900
            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
            )
1901

1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
        # 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),
1912
                    f"{name}: The tuple `attributes` should have the same length as `tf_outputs`",
1913
                )
1914
            else:
1915
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
1916
                attributes = tuple([f"{name}_{idx}" for idx in range(len(tf_outputs))])
1917

1918
1919
            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)
1920

1921
1922
1923
1924
        elif isinstance(tf_outputs, tf.Tensor):
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `tf_outputs` is"
            )
1925

1926
1927
            tf_outputs = tf_outputs.numpy()
            pt_outputs = pt_outputs.detach().to("cpu").numpy()
1928

1929
1930
1931
            self.assertEqual(
                tf_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between TF and PyTorch"
            )
1932

1933
1934
1935
1936
            # 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])
1937

1938
1939
            tf_nans = np.isnan(tf_outputs)
            pt_nans = np.isnan(pt_outputs)
1940

1941
1942
1943
1944
            pt_outputs[tf_nans] = 0
            tf_outputs[tf_nans] = 0
            pt_outputs[pt_nans] = 0
            tf_outputs[pt_nans] = 0
1945

1946
            max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
1947
            self.assertLessEqual(max_diff, tol, f"{name}: Difference between PyTorch and TF is {max_diff} (>= {tol}).")
1948
1949
        else:
            raise ValueError(
1950
                "`tf_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `tf.Tensor`. Got"
Sylvain Gugger's avatar
Sylvain Gugger committed
1951
                f" {type(tf_outputs)} instead."
1952
1953
            )

1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
    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)
1971

1972
        return tf_inputs_dict
1973

1974
1975
    def check_pt_tf_models(self, tf_model, pt_model, pt_inputs_dict):
        tf_inputs_dict = self.prepare_tf_inputs_from_pt_inputs(pt_inputs_dict)
1976

1977
1978
1979
1980
        # 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()
        }
1981

1982
1983
        # send pytorch model to the correct device
        pt_model.to(torch_device)
1984

1985
1986
        # Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
        pt_model.eval()
1987

1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
        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
Matt's avatar
Matt committed
2002
    def test_pt_tf_model_equivalence(self, allow_missing_keys=False):
2003
        import transformers
2004
2005

        for model_class in self.all_model_classes:
2006
2007
2008
            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
2009
            if not hasattr(transformers, tf_model_class_name):
2010
                # transformers does not have this model in TF version yet
2011
2012
                return

2013
2014
2015
            # Output all for aggressive testing
            config.output_hidden_states = True
            config.output_attentions = self.has_attentions
2016

2017
2018
2019
2020
            # 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)
2021
2022

            tf_model_class = getattr(transformers, tf_model_class_name)
2023
2024

            pt_model = model_class(config)
2025
2026
2027
2028
2029
2030
2031
2032
2033
            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,
            )
2034
2035
2036
2037
2038
2039
2040
2041
2042

            # 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")

2043
            pt_inputs_dict = {k: v for k, v in pt_inputs_dict.items() if k in tf_input_keys}
2044
2045
2046
2047
            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.
2048
            if not set(pt_inputs_dict_with_labels.keys()).symmetric_difference(pt_inputs_dict.keys()):
2049
                pt_inputs_dict_with_labels = None
2050
2051

            # Check we can load pt model in tf and vice-versa with model => model functions
2052
2053
            # Here requires `tf_inputs_dict` to build `tf_model`
            tf_inputs_dict = self.prepare_tf_inputs_from_pt_inputs(pt_inputs_dict)
Matt's avatar
Matt committed
2054
2055
2056
2057
2058
2059
            tf_model = transformers.load_pytorch_model_in_tf2_model(
                tf_model, pt_model, tf_inputs=tf_inputs_dict, allow_missing_keys=allow_missing_keys
            )
            pt_model = transformers.load_tf2_model_in_pytorch_model(
                pt_model, tf_model, allow_missing_keys=allow_missing_keys
            )
2060

2061
2062
2063
2064
2065
            # 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)
2066
2067
2068
2069
2070

            # 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)
Matt's avatar
Matt committed
2071
2072
2073
                tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(
                    tf_model, pt_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2074
2075
2076

                tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
                tf_model.save_weights(tf_checkpoint_path)
Matt's avatar
Matt committed
2077
2078
2079
                pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(
                    pt_model, tf_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2080

2081
2082
2083
2084
2085
            # 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)
2086
2087
2088
2089
2090

    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}).")

2091
    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
2092
2093
2094
2095
2096
2097
2098
2099
2100
        """
        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.
        """
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140

        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`",
                )
2141
            else:
2142
2143
2144
2145
2146
2147
                # 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)

2148
        elif isinstance(fx_outputs, jnp.ndarray):
2149
2150
2151
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `fx_outputs` is"
            )
2152
2153
2154
2155
2156

            # 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()

2157
2158
2159
2160
2161
2162
2163
2164
2165
            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])

2166
2167
2168
2169
2170
2171
2172
2173
            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

2174
2175
2176
2177
            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})."
            )
2178
2179
        else:
            raise ValueError(
2180
2181
                "`fx_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `jnp.ndarray`. Got"
                f" {type(fx_outputs)} instead."
2182
2183
            )

2184
2185
2186
2187
2188
2189
2190
2191
2192
    @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):
2193
                    # no flax model exists for this class
2194
2195
                    return

2196
2197
2198
2199
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2200
2201
                fx_model_class = getattr(transformers, fx_model_class_name)

2202
2203
2204
2205
2206
2207
                # 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

2208
2209
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2210

2211
2212
2213
2214
2215
2216
2217
2218
2219
                # 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}

2220
2221
2222
2223
2224
2225
                # 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
2226
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2227

2228
2229
2230
                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
                fx_model.params = fx_state

2231
2232
2233
                # send pytorch model to the correct device
                pt_model.to(torch_device)

2234
                with torch.no_grad():
2235
2236
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)
2237

2238
2239
2240
2241
                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)
2242
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2243
2244
2245
2246
2247

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

2248
2249
2250
2251
2252
2253
                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)
2254
                self.check_pt_flax_outputs(fx_outputs_loaded, pt_outputs, model_class)
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267

    @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

2268
2269
2270
2271
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2272
2273
                fx_model_class = getattr(transformers, fx_model_class_name)

2274
2275
2276
2277
2278
2279
                # 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

2280
2281
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2282

2283
2284
2285
2286
2287
2288
2289
2290
2291
                # 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}

2292
2293
2294
2295
                # 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()
                }
2296

2297
                # convert inputs to Flax
2298
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2299

2300
2301
2302
2303
2304
2305
2306
                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)
2307

2308
2309
2310
2311
2312
2313
2314
2315
                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)
2316
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2317
2318
2319
2320
2321

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

2322
2323
2324
2325
                # send pytorch model to the correct device
                pt_model_loaded.to(torch_device)
                pt_model_loaded.eval()

2326
                with torch.no_grad():
2327
                    pt_outputs_loaded = pt_model_loaded(**pt_inputs)
2328

2329
2330
2331
2332
                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)
2333
                self.check_pt_flax_outputs(fx_outputs, pt_outputs_loaded, model_class)
2334

Patrick von Platen's avatar
Patrick von Platen committed
2335
    def test_inputs_embeds(self):
2336
2337
2338
2339
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
2340
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
2341
            model.eval()
2342

2343
            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
2344

2345
2346
2347
2348
2349
2350
2351
2352
2353
            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)

2354
2355
            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
2356
                inputs["inputs_embeds"] = wte(input_ids)
2357
            else:
2358
2359
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
2360

thomwolf's avatar
thomwolf committed
2361
            with torch.no_grad():
Weizhen's avatar
Weizhen committed
2362
                model(**inputs)[0]
2363

2364
2365
    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
2366
2367
2368
2369
        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.
2370
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask", "cross_attn_head_mask"]
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
        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
2385
            model = nn.DataParallel(model)
2386
            with torch.no_grad():
2387
                _ = model(**self._prepare_for_class(inputs_dict, model_class))
2388

2389
2390
2391
    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
2392
            return
2393

2394
        # a candidate for testing_utils
2395
        def get_current_gpu_memory_use():
Patrick von Platen's avatar
Patrick von Platen committed
2396
            """returns a list of cuda memory allocations per GPU in MBs"""
2397
2398
2399
2400
2401

            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)
2402
2403
2404
2405
2406
2407
2408
2409
2410

            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()

2411
2412
2413
            # 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()
2414

2415
2416
            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
2417
2418
2419
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

2420
2421
2422
            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

2423
            del model
2424
            gc.collect()
2425
2426
            torch.cuda.empty_cache()

2427
2428
2429
            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()
2430
2431

            # Spread model layers over multiple devices
2432
            model = model_class(config)
2433
2434
2435
2436
            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
2437
            for n in range(len(model.device_map.keys())):
2438
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])
2439

2440
            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
2441
2442
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

2443
2444
            # 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
2445
2446
2447
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
2448
            gc.collect()
2449
2450
2451
2452
2453
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
2454
            return
2455
2456
2457
2458
2459
2460

        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)

2461
            def cast_to_device(dictionary, device):
2462
2463
2464
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
2465
                        output[k] = v.to(device)
2466
2467
2468
2469
2470
                    else:
                        output[k] = v

                return output

2471
2472
2473
2474
2475
2476
            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))
2477
2478
2479
2480
2481
2482
2483
2484

            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))

2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
    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
2499
    @require_accelerate
2500
    @mark.accelerate_tests
Sylvain Gugger's avatar
Sylvain Gugger committed
2501
2502
2503
2504
2505
2506
2507
2508
    @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

2509
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2510
2511
            model = model_class(config).eval()
            model = model.to(torch_device)
2512
            torch.manual_seed(0)
2513
            base_output = model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2514
2515
2516
2517
2518
2519

            model_size = compute_module_sizes(model)[""]
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(tmp_dir)

                with self.assertRaises(ValueError):
Yih-Dar's avatar
Yih-Dar committed
2520
2521
                    max_size = int(self.model_split_percents[0] * model_size)
                    max_memory = {0: max_size, "cpu": max_size}
Sylvain Gugger's avatar
Sylvain Gugger committed
2522
2523
2524
                    # 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)

Yih-Dar's avatar
Yih-Dar committed
2525
2526
                max_size = int(self.model_split_percents[1] * model_size)
                max_memory = {0: max_size, "cpu": max_size}
Sylvain Gugger's avatar
Sylvain Gugger committed
2527
2528
2529
2530
2531
                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)
2532
                torch.manual_seed(0)
2533
                new_output = new_model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2534

2535
                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
Sylvain Gugger's avatar
Sylvain Gugger committed
2536

2537
    @require_accelerate
2538
    @mark.accelerate_tests
2539
2540
2541
2542
2543
2544
2545
2546
    @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

2547
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2548
2549
            model = model_class(config).eval()
            model = model.to(torch_device)
2550
2551

            torch.manual_seed(0)
2552
            base_output = model(**inputs_dict_class)
2553
2554
2555

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
Yih-Dar's avatar
Yih-Dar committed
2556
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
            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)
2567
2568

                    torch.manual_seed(0)
2569
                    new_output = new_model(**inputs_dict_class)
2570

2571
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2572
2573

    @require_accelerate
2574
    @mark.accelerate_tests
2575
2576
2577
2578
2579
2580
2581
2582
    @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

2583
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2584
2585
            model = model_class(config).eval()
            model = model.to(torch_device)
2586
2587

            torch.manual_seed(0)
2588
            base_output = model(**inputs_dict_class)
2589
2590
2591

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
2592
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
            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)
2603
2604

                    torch.manual_seed(0)
2605
                    new_output = new_model(**inputs_dict_class)
2606

2607
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2608

2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
    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:
2619
2620
2621
            if model_class.__name__ not in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES),
2622
            ]:
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
                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"])

2641
2642
2643
2644
2645
2646
                    # 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
2647
2648
2649
2650
2651
                    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}"
                            )
2652

2653
2654
                    loss.backward()

2655
    def test_load_with_mismatched_shapes(self):
2656
2657
        if not self.test_mismatched_shapes:
            return
2658
2659
2660
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
2661
            if model_class.__name__ not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES):
2662
2663
2664
2665
2666
2667
2668
2669
                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
2670
                    with self.assertRaises(RuntimeError):
2671
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
2672
2673
                    with self.assertRaises(RuntimeError):
                        new_model_without_prefix = AutoModel.from_pretrained(tmp_dir, vocab_size=10)
2674
2675

                    logger = logging.get_logger("transformers.modeling_utils")
2676

2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
                    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)

2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
                    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)

2699
2700
2701
2702
2703
2704
2705
2706
2707
    def test_model_is_small(self):
        # Just a consistency check to make sure we are not running tests on 80M parameter models.
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            num_params = model.num_parameters()
            assert (
                num_params < 1000000
2708
            ), f"{model_class} is too big for the common tests ({num_params})! It should have 1M max."
2709

2710

2711
global_rng = random.Random()
thomwolf's avatar
thomwolf committed
2712
2713


thomwolf's avatar
thomwolf committed
2714
def ids_tensor(shape, vocab_size, rng=None, name=None):
2715
    #  Creates a random int32 tensor of the shape within the vocab size
thomwolf's avatar
thomwolf committed
2716
    if rng is None:
2717
        rng = global_rng
thomwolf's avatar
thomwolf committed
2718

thomwolf's avatar
thomwolf committed
2719
2720
2721
    total_dims = 1
    for dim in shape:
        total_dims *= dim
thomwolf's avatar
thomwolf committed
2722

thomwolf's avatar
thomwolf committed
2723
2724
2725
    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))
thomwolf's avatar
thomwolf committed
2726

2727
    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
thomwolf's avatar
thomwolf committed
2728
2729


2730
2731
2732
2733
2734
2735
2736
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


2737
def floats_tensor(shape, scale=1.0, rng=None, name=None):
Patrick von Platen's avatar
Patrick von Platen committed
2738
    """Creates a random float32 tensor"""
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
    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)

2750
    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()