test_modeling_common.py 136 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_bitsandbytes,
68
    require_flash_attn,
69
    require_safetensors,
Sylvain Gugger's avatar
Sylvain Gugger committed
70
    require_torch,
71
    require_torch_gpu,
Sylvain Gugger's avatar
Sylvain Gugger committed
72
73
74
75
    require_torch_multi_gpu,
    slow,
    torch_device,
)
76
from transformers.utils import (
77
78
    CONFIG_NAME,
    GENERATION_CONFIG_NAME,
79
    WEIGHTS_NAME,
80
    is_accelerate_available,
81
82
83
84
85
    is_flax_available,
    is_tf_available,
    is_torch_fx_available,
)
from transformers.utils.generic import ModelOutput
86

Aymeric Augustin's avatar
Aymeric Augustin committed
87

88
89
90
91
if is_accelerate_available():
    from accelerate.utils import compute_module_sizes


92
if is_torch_available():
93
    import torch
94
    from torch import nn
thomwolf's avatar
thomwolf committed
95

96
    from transformers import MODEL_MAPPING, AdaptiveEmbedding
Sylvain Gugger's avatar
Sylvain Gugger committed
97
    from transformers.pytorch_utils import id_tensor_storage
thomwolf's avatar
thomwolf committed
98

Sylvain Gugger's avatar
Sylvain Gugger committed
99

100
101
102
if is_tf_available():
    import tensorflow as tf

103
104
if is_flax_available():
    import jax.numpy as jnp
105

106
107
108
109
110
    from transformers.modeling_flax_pytorch_utils import (
        convert_pytorch_state_dict_to_flax,
        load_flax_weights_in_pytorch_model,
    )

111
if is_torch_fx_available():
112
    from transformers.utils.fx import symbolic_trace
113

114

115
116
117
def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
118
        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
119
            setattr(configs_no_init, key, 1e-10)
120
121
122
        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)
123
124
    return configs_no_init

thomwolf's avatar
thomwolf committed
125

126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
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()


151
152
153
154
@require_torch
class ModelTesterMixin:
    model_tester = None
    all_model_classes = ()
155
    all_generative_model_classes = ()
156
    fx_compatible = False
Patrick von Platen's avatar
Patrick von Platen committed
157
158
159
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
160
    test_resize_position_embeddings = False
Patrick von Platen's avatar
Patrick von Platen committed
161
    test_head_masking = True
162
    test_mismatched_shapes = True
163
    test_missing_keys = True
164
    test_model_parallel = False
165
    is_encoder_decoder = False
166
    has_attentions = True
167
    model_split_percents = [0.5, 0.7, 0.9]
168

169
170
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
171
        if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
172
            inputs_dict = {
173
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
174
                if isinstance(v, torch.Tensor) and v.ndim > 1
Sylvain Gugger's avatar
Sylvain Gugger committed
175
                else v
176
177
                for k, v in inputs_dict.items()
            }
178
        elif model_class.__name__ in get_values(MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES):
179
            inputs_dict.pop("attention_mask")
180
181

        if return_labels:
182
            if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
183
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
184
185
186
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
187
            ]:
188
189
190
191
192
193
                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
                )
194
195
196
197
198
199
            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),
200
            ]:
201
202
203
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
204
205
206
207
208
209
            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),
210
211
212
213
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
214
            elif model_class.__name__ in get_values(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
215
216
217
218
                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
                )
219
            elif model_class.__name__ in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
220
221
222
223
                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()
224

225
226
        return inputs_dict

Patrick von Platen's avatar
Patrick von Platen committed
227
    def test_save_load(self):
228
229
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

230
231
232
233
234
235
236
237
238
239
        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)

240
241
242
243
244
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
245
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
246

247
            with tempfile.TemporaryDirectory() as tmpdirname:
248
                model.save_pretrained(tmpdirname)
249
250
251
252
253
254
255

                # 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))
                )

256
                model = model_class.from_pretrained(tmpdirname)
257
                model.to(torch_device)
258
                with torch.no_grad():
259
                    second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
thomwolf's avatar
thomwolf committed
260

261
262
263
264
265
            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)
266

267
268
269
270
271
272
273
274
275
276
277
278
    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))

279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
    def test_keep_in_fp32_modules(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
        for model_class in self.all_model_classes:
            if model_class._keep_in_fp32_modules is None:
                return

            model = model_class(config)
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)

                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16)

                for name, param in model.named_parameters():
                    if any(n in model_class._keep_in_fp32_modules for n in name.split(".")):
                        self.assertTrue(param.dtype == torch.float32)
                    else:
                        self.assertTrue(param.dtype == torch.float16, name)

297
    def test_save_load_keys_to_ignore_on_save(self):
298
299
300
301
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
302
303
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
304
305
306
                continue

            # check the keys are in the original state_dict
307
            for k in _keys_to_ignore_on_save:
308
                self.assertIn(k, model.state_dict().keys(), "\n".join(model.state_dict().keys()))
309
310
311
312
313
314

            # 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)
315
                for k in _keys_to_ignore_on_save:
316
                    self.assertNotIn(k, state_dict_saved.keys(), "\n".join(state_dict_saved.keys()))
317

Sylvain Gugger's avatar
Sylvain Gugger committed
318
319
320
                # 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(
321
322
                    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
323
324
325
                )
                self.assertTrue(len(load_result.unexpected_keys) == 0)

326
327
328
329
330
331
332
333
334
335
336
    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)

337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
    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)

356
357
    def test_save_load_fast_init_from_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
358
359
        if config.__class__ not in MODEL_MAPPING:
            return
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
        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
381
382
            model_class_copy._init_weights = _mock_init_weights
            model_class_copy.init_weights = _mock_all_init_weights
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398

            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)
399
                # Before we test anything
400
401

                for key in model_fast_init.state_dict().keys():
402
403
404
405
406
                    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")
407
408
409

    def test_save_load_fast_init_to_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
410
411
        if config.__class__ not in MODEL_MAPPING:
            return
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
        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
433
434
            base_class_copy._init_weights = _mock_init_weights
            base_class_copy.init_weights = _mock_all_init_weights
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452

            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():
453
454
455
456
457
458
459
460
461
                    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")
462

Patrick von Platen's avatar
Patrick von Platen committed
463
    def test_initialization(self):
464
465
466
467
468
469
470
471
        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
472
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
473
                        [0.0, 1.0],
474
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
475
                    )
thomwolf's avatar
thomwolf committed
476

Patrick von Platen's avatar
Patrick von Platen committed
477
    def test_determinism(self):
478
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
479
480
481
482
483
484
485
486
487

        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)

488
489
490
491
492
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
493
494
                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
495

496
497
498
499
500
            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_determinism(tensor1, tensor2)
            else:
                check_determinism(first, second)
501

502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
    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",
                ]
518
                expected_arg_names.extend(
519
520
                    ["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
521
522
523
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
524
525
526
527
            else:
                expected_arg_names = ["input_ids"]
                self.assertListEqual(arg_names[:1], expected_arg_names)

528
529
530
531
532
    def test_training(self):
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
533
534
535
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.return_dict = True

536
537
538
            if model_class.__name__ in [
                *get_values(MODEL_MAPPING_NAMES),
                *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES),
539
            ]:
540
                continue
541

542
543
544
545
546
547
548
549
            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):
550
        if not self.model_tester.is_training:
551
552
553
            return

        for model_class in self.all_model_classes:
554
555
556
557
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.use_cache = False
            config.return_dict = True

558
            if (
559
560
                model_class.__name__
                in [*get_values(MODEL_MAPPING_NAMES), *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)]
561
562
                or not model_class.supports_gradient_checkpointing
            ):
563
564
565
                continue
            model = model_class(config)
            model.to(torch_device)
566
            model.gradient_checkpointing_enable()
567
568
569
570
571
            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
572
    def test_attention_outputs(self):
573
574
575
        if not self.has_attentions:
            self.skipTest(reason="Model does not output attentions")

576
577
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True
578

579
580
581
582
583
584
585
586
587
588
589
590
        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
591
            config.return_dict = True
592
593
594
595
596
597
598
            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)
599

600
601
602
603
604
605
606
607
608
609
            # 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
610

611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
            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
630
631
632
                if model_class.__name__ in [
                    *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                    *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
633
                ]:
634
635
636
637
638
639
640
641
642
643
644
645
646
647
                    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],
                )
648

649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
                # 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
692

693
    @slow
694
    def test_torchscript_simple(self):
695
696
        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
697

698
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
699
    def test_torchscript_output_attentions(self):
700
701
702
        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
703

704
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
705
    def test_torchscript_output_hidden_state(self):
706
707
708
        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
709

710
711
712
713
    # 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()
714
715
716
        # 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()
717

718
    def _create_and_check_torchscript(self, config, inputs_dict):
Patrick von Platen's avatar
Patrick von Platen committed
719
        if not self.test_torchscript:
720
            return
721

722
723
724
725
726
727
        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()
728
            inputs = self._prepare_for_class(inputs_dict, model_class)
thomwolf's avatar
thomwolf committed
729

730
731
            main_input_name = model_class.main_input_name

732
            try:
733
                if model.config.is_encoder_decoder:
734
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
735
                    main_input = inputs[main_input_name]
736
737
738
                    attention_mask = inputs["attention_mask"]
                    decoder_input_ids = inputs["decoder_input_ids"]
                    decoder_attention_mask = inputs["decoder_attention_mask"]
739
                    model(main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
740
                    traced_model = torch.jit.trace(
741
                        model, (main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
742
                    )
743
744
745
746
                elif "bbox" in inputs and "image" in inputs:  # LayoutLMv2 requires additional inputs
                    input_ids = inputs["input_ids"]
                    bbox = inputs["bbox"]
                    image = inputs["image"].tensor
747
                    model(input_ids, bbox, image)
748
749
750
                    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
Jinho Park's avatar
Jinho Park committed
751
752
753
754
755
756
757
                elif "bbox" in inputs:  # Bros requires additional inputs (bbox)
                    input_ids = inputs["input_ids"]
                    bbox = inputs["bbox"]
                    model(input_ids, bbox)
                    traced_model = torch.jit.trace(
                        model, (input_ids, bbox), check_trace=False
                    )  # when traced model is checked, an error is produced due to name mangling
758
                else:
759
                    main_input = inputs[main_input_name]
760
                    model(main_input)
761
                    traced_model = torch.jit.trace(model, main_input)
762
763
            except RuntimeError:
                self.fail("Couldn't trace module.")
thomwolf's avatar
thomwolf committed
764

765
            with tempfile.TemporaryDirectory() as tmp_dir_name:
766
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
thomwolf's avatar
thomwolf committed
767

768
                try:
769
                    torch.jit.save(traced_model, pt_file_name)
770
771
                except Exception:
                    self.fail("Couldn't save module.")
thomwolf's avatar
thomwolf committed
772

773
774
775
776
                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")
LysandreJik's avatar
LysandreJik committed
777

778
779
            model.to(torch_device)
            model.eval()
thomwolf's avatar
thomwolf committed
780

781
782
            loaded_model.to(torch_device)
            loaded_model.eval()
thomwolf's avatar
thomwolf committed
783

784
785
786
            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

787
788
789
790
791
792
793
794
795
            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
            }

796
            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
thomwolf's avatar
thomwolf committed
797

798
799
800
801
802
803
804
805
806
807
808
            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)

809
            models_equal = True
810
            for layer_name, p1 in model_state_dict.items():
811
812
813
814
                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
815

816
            self.assertTrue(models_equal)
thomwolf's avatar
thomwolf committed
817

818
819
820
821
            # 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()

822
823
824
825
826
827
828
829
    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)

830
831
    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:
832
833
834
835
836
            return

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

837
        for model_class in self.all_model_classes:
838
839
840
841
842
843
844
845
846
            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)
847
848
849
                    input_names = [
                        "attention_mask",
                        "decoder_attention_mask",
850
                        "decoder_input_ids",
851
                        "input_features",
852
853
                        "input_ids",
                        "input_values",
854
                    ]
855
856
                    if labels is not None:
                        input_names.append("labels")
857

858
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
859
                    input_names = list(filtered_inputs.keys())
860

861
                    model_output = model(**filtered_inputs)
862

863
                    traced_model = symbolic_trace(model, input_names)
864
                    traced_output = traced_model(**filtered_inputs)
865
                else:
866
867
868
869
                    input_names = [
                        "attention_mask",
                        "bbox",
                        "input_features",
870
871
872
873
874
875
                        "input_ids",
                        "input_values",
                        "pixel_values",
                        "token_type_ids",
                        "visual_feats",
                        "visual_pos",
876
                    ]
877

878
                    labels = inputs.get("labels", None)
879
880
                    start_positions = inputs.get("start_positions", None)
                    end_positions = inputs.get("end_positions", None)
881
882
                    if labels is not None:
                        input_names.append("labels")
883
884
885
886
                    if start_positions is not None:
                        input_names.append("start_positions")
                    if end_positions is not None:
                        input_names.append("end_positions")
887

888
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
889
                    input_names = list(filtered_inputs.keys())
890

891
                    if model.__class__.__name__ in set(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES.values()) and (
892
                        not hasattr(model.config, "problem_type") or model.config.problem_type is None
893
894
895
                    ):
                        model.config.problem_type = "single_label_classification"

896
                    traced_model = symbolic_trace(model, input_names)
897
                    traced_output = traced_model(**filtered_inputs)
898
                    model_output = model(**filtered_inputs)
899

900
            except Exception as e:
901
                self.fail(f"Couldn't trace module: {e}")
902

903
904
905
906
907
908
909
910
911
912
913
914
915
            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)
916
            num_outputs = len(model_output)
917
918
919
920
921
922

            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}",
                )
923

924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
            # 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}",
                    )

944
945
946
947
            # 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
948
949
    def test_headmasking(self):
        if not self.test_head_masking:
950
            return
951

952
953
954
        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()
LysandreJik's avatar
LysandreJik committed
955

956
        inputs_dict["output_attentions"] = True
957
958
959
960
961
962
        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
963

964
965
966
            # 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
967
968
969
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
970
971
972
973
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
974
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
975
            inputs["head_mask"] = head_mask
976
977
978
979
980
            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
981
982
                if "cross_attn_head_mask" in arg_names:
                    inputs["cross_attn_head_mask"] = head_mask
983
            outputs = model(**inputs, return_dict=True)
984
985
986
987
988
989
990
991
992

            # 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)
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013

            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)
1014
                check_attentions_validity(outputs.cross_attentions)
1015
1016
            else:
                check_attentions_validity(outputs.attentions)
1017

Patrick von Platen's avatar
Patrick von Platen committed
1018
1019
    def test_head_pruning(self):
        if not self.test_pruning:
1020
1021
1022
            return

        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1023
1024
1025
1026
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1027

1028
1029
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1030

1031
            inputs_dict["output_attentions"] = True
1032
1033
1034
1035
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1036
1037
1038
1039
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1040
1041
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
1042
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1043

1044
            attentions = outputs[-1]
1045

1046
            self.assertEqual(attentions[0].shape[-3], 1)
1047
1048
            # 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)
1049
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
LysandreJik's avatar
LysandreJik committed
1050

Patrick von Platen's avatar
Patrick von Platen committed
1051
1052
    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
1053
            return
LysandreJik's avatar
LysandreJik committed
1054

1055
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1056
1057
1058
1059
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1060
1061
1062

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1063

1064
            inputs_dict["output_attentions"] = True
1065
1066
1067
1068
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1069
1070
1071
1072
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1073
            model.prune_heads(heads_to_prune)
1074

1075
            with tempfile.TemporaryDirectory() as temp_dir_name:
1076
1077
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1078
                model.to(torch_device)
1079

1080
            with torch.no_grad():
1081
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1082
1083
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
1084
1085
            # 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)
1086
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1087

Patrick von Platen's avatar
Patrick von Platen committed
1088
1089
    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
1090
            return
1091

1092
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1093
1094
1095
1096
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1097

1098
1099
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1100

1101
            inputs_dict["output_attentions"] = True
1102
            config.output_hidden_states = False
1103

1104
1105
1106
1107
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1108
            config.pruned_heads = heads_to_prune
1109

1110
1111
1112
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1113

1114
            with torch.no_grad():
1115
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1116
            attentions = outputs[-1]
1117

1118
            self.assertEqual(attentions[0].shape[-3], 1)
1119
1120
            # 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)
1121
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1122

Patrick von Platen's avatar
Patrick von Platen committed
1123
1124
    def test_head_pruning_integration(self):
        if not self.test_pruning:
1125
            return
1126

1127
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1128
1129
1130
1131
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1132

1133
1134
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1135

1136
            inputs_dict["output_attentions"] = True
1137
            config.output_hidden_states = False
1138

1139
            heads_to_prune = {1: [1, 2]}
1140
            config.pruned_heads = heads_to_prune
1141

1142
1143
1144
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1145

1146
            with torch.no_grad():
1147
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1148
            attentions = outputs[-1]
1149

1150
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1151
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1152

1153
            with tempfile.TemporaryDirectory() as temp_dir_name:
1154
1155
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1156
                model.to(torch_device)
thomwolf's avatar
thomwolf committed
1157

1158
            with torch.no_grad():
1159
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1160
            attentions = outputs[-1]
LysandreJik's avatar
LysandreJik committed
1161

1162
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1163
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1164

1165
            heads_to_prune = {0: [0], 1: [1, 2]}
1166
            model.prune_heads(heads_to_prune)
1167

1168
            with torch.no_grad():
1169
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1170
            attentions = outputs[-1]
1171

1172
1173
            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)
1174

1175
            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2]})
thomwolf's avatar
thomwolf committed
1176

Patrick von Platen's avatar
Patrick von Platen committed
1177
    def test_hidden_states_output(self):
Joseph Liu's avatar
Joseph Liu committed
1178
        def check_hidden_states_output(inputs_dict, config, model_class):
1179
            model = model_class(config)
1180
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
1181
            model.eval()
Joseph Liu's avatar
Joseph Liu committed
1182

thomwolf's avatar
thomwolf committed
1183
            with torch.no_grad():
1184
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1185
1186

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
1187

Sylvain Gugger's avatar
Sylvain Gugger committed
1188
1189
1190
1191
            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)
1192

Patrick von Platen's avatar
Patrick von Platen committed
1193
1194
1195
1196
1197
1198
1199
            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

1200
            self.assertListEqual(
Lysandre's avatar
Lysandre committed
1201
1202
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
1203
            )
thomwolf's avatar
thomwolf committed
1204

1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
            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
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
        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)

1230
1231
1232
    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
1233
        config.output_attentions = self.has_attentions
1234
1235
1236
1237
1238
1239
1240
1241
1242

        # 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)
1243

1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
        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()

1254
1255
1256
1257
1258
1259
1260
1261
1262
            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()
1263
1264
1265
1266
1267

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

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)
1268
1269
1270
1271
1272

            if self.has_attentions:
                self.assertIsNotNone(encoder_attentions.grad)
                self.assertIsNotNone(decoder_attentions.grad)
                self.assertIsNotNone(cross_attentions.grad)
1273
1274
1275
1276
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            hidden_states.retain_grad()
1277
1278
1279
1280

            if self.has_attentions:
                attentions = outputs.attentions[0]
                attentions.retain_grad()
1281
1282
1283
1284

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

            self.assertIsNotNone(hidden_states.grad)
1285
1286
1287

            if self.has_attentions:
                self.assertIsNotNone(attentions.grad)
1288

Pradhy729's avatar
Pradhy729 committed
1289
    def test_feed_forward_chunking(self):
Lysandre's avatar
Lysandre committed
1290
1291
1292
1293
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Pradhy729's avatar
Pradhy729 committed
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
        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))

1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
    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
1391
    def test_resize_tokens_embeddings(self):
Lysandre's avatar
Lysandre committed
1392
1393
1394
1395
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Patrick von Platen's avatar
Patrick von Platen committed
1396
        if not self.test_resize_embeddings:
1397
1398
1399
1400
1401
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
1402
            model.to(torch_device)
1403

Patrick von Platen's avatar
Patrick von Platen committed
1404
1405
1406
            if self.model_tester.is_training is False:
                model.eval()

1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
            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)
1417
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
1418
            model(**self._prepare_for_class(inputs_dict, model_class))
1419
1420
1421
1422
1423
1424
1425

            # 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)

1426
1427
1428
            # 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)
1429
1430
1431
1432

            # 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)
1433
            model(**self._prepare_for_class(inputs_dict, model_class))
1434

1435
1436
1437
1438
1439
1440
1441
1442
            # 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)

1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
            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)

Arthur's avatar
Arthur committed
1454
1455
1456
            self.assertTrue(model_embed.weight.shape[0], model.config.vocab_size)
            self.assertTrue(model.config.vocab_size, model.vocab_size)

1457
1458
1459
            model_embed = model.resize_token_embeddings(model_vocab_size + 13, pad_to_multiple_of=64)
            self.assertTrue(model_embed.weight.shape[0] // 64, 0)

1460
1461
1462
1463
1464
            # Check that resizing a model to a multiple of pad_to_multiple leads to a model of exactly that size
            target_dimension = 128
            model_embed = model.resize_token_embeddings(target_dimension, pad_to_multiple_of=64)
            self.assertTrue(model_embed.weight.shape[0], target_dimension)

1465
1466
1467
1468
1469
1470
            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)

1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
    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
1522
    def test_model_common_attributes(self):
1523
1524
1525
1526
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1527
1528
            self.assertIsInstance(model.get_input_embeddings(), (nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(nn.Embedding(10, 10))
1529
            x = model.get_output_embeddings()
1530
            self.assertTrue(x is None or isinstance(x, nn.Linear))
1531

1532
1533
1534
1535
1536
1537
1538
    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)

1539
    def test_correct_missing_keys(self):
1540
1541
        if not self.test_missing_keys:
            return
1542
1543
1544
1545
1546
1547
1548
        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):
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
                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

1562
1563
1564
                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)
1565
                    self.assertGreater(len(loading_info["missing_keys"]), 0, model.__class__.__name__)
1566

1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
    def test_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))

1615
1616
    @require_safetensors
    def test_can_use_safetensors(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
1617
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
        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
1634
1635
                # Checking there was no complain of missing weights
                self.assertEqual(infos["missing_keys"], [])
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651

                # 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
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
    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
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
    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
1696
1697
1698
1699
1700
            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
1701

Sylvain Gugger's avatar
Sylvain Gugger committed
1702
1703
    def test_model_weights_reload_no_missing_tied_weights(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1704
        for model_class in self.all_model_classes:
Sylvain Gugger's avatar
Sylvain Gugger committed
1705
1706
1707
            model = model_class(config)
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.save_pretrained(tmp_dir)
1708
1709
1710

                # 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
1711
                with open(os.path.join(tmp_dir, "pytorch_model.bin"), "wb") as f:
1712
                    torch.save({}, f)
Sylvain Gugger's avatar
Sylvain Gugger committed
1713
                model_reloaded, infos = model_class.from_pretrained(tmp_dir, output_loading_info=True)
1714
1715
1716
1717

                prefix = f"{model_reloaded.base_model_prefix}."
                params = dict(model_reloaded.named_parameters())
                params.update(dict(model_reloaded.named_buffers()))
1718
                param_names = {k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys()}
1719
1720
1721
1722

                missing_keys = set(infos["missing_keys"])

                extra_missing = missing_keys - param_names
Sylvain Gugger's avatar
Sylvain Gugger committed
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
                # 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)
1734
1735
1736
1737

                self.assertEqual(
                    extra_missing,
                    set(),
Sylvain Gugger's avatar
Sylvain Gugger committed
1738
1739
                    f"This model {model_class.__name__} might be missing some `keys_to_ignore`: {extra_missing}. "
                    f"For debugging, tied parameters are {tied_params}",
1740
1741
                )

Sylvain Gugger's avatar
Sylvain Gugger committed
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
                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`",
                )
1762

1763
1764
1765
    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
1766
1767
1768
1769
        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

1770
1771
1772
1773
1774
1775
1776
1777
1778
        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
1779
1780
1781
1782
1783
                    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)
1784
1785
1786
1787
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
Sam Shleifer's avatar
Sam Shleifer committed
1788
1789
1790
                            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
1791
1792
1793
1794
1795
1796
                            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)}."
                            ),
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
                        )

                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})

1822
1823
1824
1825
            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})
1826

1827
1828
1829
1830
1831
1832
1833
1834
1835
                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}
                )
1836

1837
1838
1839
1840
    # 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"""
1841

1842
1843
1844
        for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
            if k in inputs_dict:
                attention_mask = inputs_dict[k]
1845

1846
1847
1848
1849
1850
1851
                # 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
                )
1852

1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
                # 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)"""

1870
1871
        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}
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898

        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):
1899
        """Check the outputs from PyTorch and TensorFlow models are close enough. Checks are done in a recursive way.
1900

1901
1902
1903
1904
1905
1906
1907
1908
        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.
        """
1909

1910
1911
1912
        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")
1913

1914
1915
1916
1917
1918
1919
        # 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",
            )
1920

1921
1922
1923
            # 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)
1924

1925
1926
            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]
1927

1928
            self.assertEqual(tf_keys, pt_keys, f"{name}: Output keys differ between TF and PyTorch")
1929

1930
            # convert to the case of `tuple`
1931
            # appending each key to the current (string) `name`
1932
1933
1934
1935
            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
            )
1936

1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
        # 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),
1947
                    f"{name}: The tuple `attributes` should have the same length as `tf_outputs`",
1948
                )
1949
            else:
1950
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
1951
                attributes = tuple([f"{name}_{idx}" for idx in range(len(tf_outputs))])
1952

1953
1954
            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)
1955

1956
1957
1958
1959
        elif isinstance(tf_outputs, tf.Tensor):
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `tf_outputs` is"
            )
1960

1961
1962
            tf_outputs = tf_outputs.numpy()
            pt_outputs = pt_outputs.detach().to("cpu").numpy()
1963

1964
1965
1966
            self.assertEqual(
                tf_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between TF and PyTorch"
            )
1967

1968
1969
1970
1971
            # 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])
1972

1973
1974
            tf_nans = np.isnan(tf_outputs)
            pt_nans = np.isnan(pt_outputs)
1975

1976
1977
1978
1979
            pt_outputs[tf_nans] = 0
            tf_outputs[tf_nans] = 0
            pt_outputs[pt_nans] = 0
            tf_outputs[pt_nans] = 0
1980

1981
            max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
1982
            self.assertLessEqual(max_diff, tol, f"{name}: Difference between PyTorch and TF is {max_diff} (>= {tol}).")
1983
1984
        else:
            raise ValueError(
1985
                "`tf_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `tf.Tensor`. Got"
Sylvain Gugger's avatar
Sylvain Gugger committed
1986
                f" {type(tf_outputs)} instead."
1987
1988
            )

1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
    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)
2006

2007
        return tf_inputs_dict
2008

2009
2010
    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)
2011

2012
2013
2014
2015
        # 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()
        }
2016

2017
2018
        # send pytorch model to the correct device
        pt_model.to(torch_device)
2019

2020
2021
        # Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
        pt_model.eval()
2022

2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
        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
2037
    def test_pt_tf_model_equivalence(self, allow_missing_keys=False):
2038
        import transformers
2039
2040

        for model_class in self.all_model_classes:
2041
2042
2043
            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
2044
            if not hasattr(transformers, tf_model_class_name):
2045
                # transformers does not have this model in TF version yet
2046
2047
                return

2048
2049
2050
            # Output all for aggressive testing
            config.output_hidden_states = True
            config.output_attentions = self.has_attentions
2051

2052
2053
2054
2055
            # 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)
2056
2057

            tf_model_class = getattr(transformers, tf_model_class_name)
2058
2059

            pt_model = model_class(config)
2060
2061
2062
2063
2064
2065
2066
2067
2068
            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,
            )
2069
2070
2071
2072
2073
2074
2075
2076
2077

            # 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")

2078
            pt_inputs_dict = {k: v for k, v in pt_inputs_dict.items() if k in tf_input_keys}
2079
2080
2081
2082
            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.
2083
            if not set(pt_inputs_dict_with_labels.keys()).symmetric_difference(pt_inputs_dict.keys()):
2084
                pt_inputs_dict_with_labels = None
2085
2086

            # Check we can load pt model in tf and vice-versa with model => model functions
2087
2088
            # 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
2089
2090
2091
2092
2093
2094
            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
            )
2095

2096
2097
2098
2099
2100
            # 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)
2101
2102
2103
2104
2105

            # 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
2106
2107
2108
                tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(
                    tf_model, pt_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2109
2110
2111

                tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
                tf_model.save_weights(tf_checkpoint_path)
Matt's avatar
Matt committed
2112
2113
2114
                pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(
                    pt_model, tf_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2115

2116
2117
2118
2119
2120
            # 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)
2121
2122
2123
2124
2125

    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}).")

2126
    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
2127
2128
2129
2130
2131
2132
2133
2134
2135
        """
        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.
        """
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175

        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`",
                )
2176
            else:
2177
2178
2179
2180
2181
2182
                # 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)

2183
        elif isinstance(fx_outputs, jnp.ndarray):
2184
2185
2186
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `fx_outputs` is"
            )
2187
2188
2189
2190
2191

            # 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()

2192
2193
2194
2195
2196
2197
2198
2199
2200
            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])

2201
2202
2203
2204
2205
2206
2207
2208
            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

2209
2210
2211
2212
            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})."
            )
2213
2214
        else:
            raise ValueError(
2215
2216
                "`fx_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `jnp.ndarray`. Got"
                f" {type(fx_outputs)} instead."
2217
2218
            )

2219
2220
2221
2222
2223
2224
2225
2226
2227
    @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):
2228
                    # no flax model exists for this class
2229
2230
                    return

2231
2232
2233
2234
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2235
2236
                fx_model_class = getattr(transformers, fx_model_class_name)

2237
2238
2239
2240
2241
2242
                # 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

2243
2244
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2245

2246
2247
2248
2249
2250
2251
2252
2253
2254
                # 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}

2255
2256
2257
2258
2259
2260
                # 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
2261
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2262

2263
2264
2265
                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
                fx_model.params = fx_state

2266
2267
2268
                # send pytorch model to the correct device
                pt_model.to(torch_device)

2269
                with torch.no_grad():
2270
2271
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)
2272

2273
2274
2275
2276
                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)
2277
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2278
2279
2280
2281
2282

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

2283
2284
2285
2286
2287
2288
                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)
2289
                self.check_pt_flax_outputs(fx_outputs_loaded, pt_outputs, model_class)
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302

    @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

2303
2304
2305
2306
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2307
2308
                fx_model_class = getattr(transformers, fx_model_class_name)

2309
2310
2311
2312
2313
2314
                # 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

2315
2316
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2317

2318
2319
2320
2321
2322
2323
2324
2325
2326
                # 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}

2327
2328
2329
2330
                # 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()
                }
2331

2332
                # convert inputs to Flax
2333
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2334

2335
2336
2337
2338
2339
2340
2341
                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)
2342

2343
2344
2345
2346
2347
2348
2349
2350
                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)
2351
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2352
2353
2354
2355
2356

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

2357
2358
2359
2360
                # send pytorch model to the correct device
                pt_model_loaded.to(torch_device)
                pt_model_loaded.eval()

2361
                with torch.no_grad():
2362
                    pt_outputs_loaded = pt_model_loaded(**pt_inputs)
2363

2364
2365
2366
2367
                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)
2368
                self.check_pt_flax_outputs(fx_outputs, pt_outputs_loaded, model_class)
2369

Patrick von Platen's avatar
Patrick von Platen committed
2370
    def test_inputs_embeds(self):
2371
2372
2373
2374
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
2375
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
2376
            model.eval()
2377

2378
            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
2379

2380
2381
2382
2383
2384
2385
2386
2387
2388
            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)

2389
2390
            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
2391
                inputs["inputs_embeds"] = wte(input_ids)
2392
            else:
2393
2394
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
2395

thomwolf's avatar
thomwolf committed
2396
            with torch.no_grad():
Weizhen's avatar
Weizhen committed
2397
                model(**inputs)[0]
2398

2399
2400
    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
2401
2402
2403
2404
        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.
2405
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask", "cross_attn_head_mask"]
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
        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
2420
            model = nn.DataParallel(model)
2421
            with torch.no_grad():
2422
                _ = model(**self._prepare_for_class(inputs_dict, model_class))
2423

2424
2425
2426
    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
2427
            return
2428

2429
        # a candidate for testing_utils
2430
        def get_current_gpu_memory_use():
Patrick von Platen's avatar
Patrick von Platen committed
2431
            """returns a list of cuda memory allocations per GPU in MBs"""
2432
2433
2434
2435
2436

            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)
2437
2438
2439
2440
2441
2442
2443
2444
2445

            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()

2446
2447
2448
            # 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()
2449

2450
2451
            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
2452
2453
2454
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

2455
2456
2457
            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

2458
            del model
2459
            gc.collect()
2460
2461
            torch.cuda.empty_cache()

2462
2463
2464
            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()
2465
2466

            # Spread model layers over multiple devices
2467
            model = model_class(config)
2468
2469
2470
2471
            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
2472
            for n in range(len(model.device_map.keys())):
2473
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])
2474

2475
            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
2476
2477
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

2478
2479
            # 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
2480
2481
2482
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
2483
            gc.collect()
2484
2485
2486
2487
2488
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
2489
            return
2490
2491
2492
2493
2494
2495

        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)

2496
            def cast_to_device(dictionary, device):
2497
2498
2499
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
2500
                        output[k] = v.to(device)
2501
2502
2503
2504
2505
                    else:
                        output[k] = v

                return output

2506
2507
2508
2509
2510
2511
            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))
2512
2513
2514
2515
2516
2517
2518
2519

            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))

2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
    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
2534
    @require_accelerate
2535
    @mark.accelerate_tests
Sylvain Gugger's avatar
Sylvain Gugger committed
2536
2537
2538
2539
2540
2541
2542
2543
    @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

2544
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2545
2546
            model = model_class(config).eval()
            model = model.to(torch_device)
2547
            torch.manual_seed(0)
2548
            base_output = model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2549
2550
2551
2552
2553
2554

            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
2555
2556
                    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
2557
2558
2559
                    # 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
2560
2561
                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
2562
2563
2564
2565
2566
                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)
2567
                torch.manual_seed(0)
2568
                new_output = new_model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2569

2570
                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
Sylvain Gugger's avatar
Sylvain Gugger committed
2571

2572
    @require_accelerate
2573
    @mark.accelerate_tests
2574
2575
2576
2577
2578
2579
2580
2581
    @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

2582
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2583
2584
            model = model_class(config).eval()
            model = model.to(torch_device)
2585
2586

            torch.manual_seed(0)
2587
            base_output = model(**inputs_dict_class)
2588
2589
2590

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
Yih-Dar's avatar
Yih-Dar committed
2591
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
            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)
2602
2603

                    torch.manual_seed(0)
2604
                    new_output = new_model(**inputs_dict_class)
2605

2606
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2607
2608

    @require_accelerate
2609
    @mark.accelerate_tests
2610
2611
2612
2613
2614
2615
2616
2617
    @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

2618
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2619
2620
            model = model_class(config).eval()
            model = model.to(torch_device)
2621
2622

            torch.manual_seed(0)
2623
            base_output = model(**inputs_dict_class)
2624
2625
2626

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
2627
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
            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)
2638
2639

                    torch.manual_seed(0)
2640
                    new_output = new_model(**inputs_dict_class)
2641

2642
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2643

2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
    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:
2654
2655
2656
            if model_class.__name__ not in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES),
2657
            ]:
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
                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"])

2676
2677
2678
2679
2680
2681
                    # 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
2682
2683
2684
2685
2686
                    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}"
                            )
2687

2688
2689
                    loss.backward()

2690
    def test_load_with_mismatched_shapes(self):
2691
2692
        if not self.test_mismatched_shapes:
            return
2693
2694
2695
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
2696
            if model_class.__name__ not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES):
2697
2698
2699
2700
2701
2702
2703
2704
                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
2705
                    with self.assertRaises(RuntimeError):
2706
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
2707
2708
                    with self.assertRaises(RuntimeError):
                        new_model_without_prefix = AutoModel.from_pretrained(tmp_dir, vocab_size=10)
2709
2710

                    logger = logging.get_logger("transformers.modeling_utils")
2711

2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
                    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)

2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
                    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)

2734
2735
2736
2737
2738
2739
2740
2741
2742
    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
2743
            ), f"{model_class} is too big for the common tests ({num_params})! It should have 1M max."
2744

2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_conversion(self):
        import torch

        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
                return

            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.float16, use_flash_attention_2=True
                ).to(torch_device)

                for _, module in model.named_modules():
                    if "FlashAttention" in module.__class__.__name__:
                        return

                self.assertTrue(False, "FlashAttention2 modules not found in model")

    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_inference(self):
        import torch

        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
                return

            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model_fa = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=True
                )
                model_fa.to(torch_device)

                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=False
                )
                model.to(torch_device)

                dummy_input = torch.LongTensor([[1, 2, 3, 4, 5]]).to(torch_device)
                dummy_attention_mask = torch.LongTensor([[0, 1, 1, 1, 1]]).to(torch_device)

                logits = model(dummy_input, output_hidden_states=True).hidden_states[-1]
                logits_fa = model_fa(dummy_input, output_hidden_states=True).hidden_states[-1]

                self.assertTrue(torch.allclose(logits_fa, logits, atol=4e-2, rtol=4e-2))

                output_fa = model_fa(dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True)
                logits_fa = output_fa.hidden_states[-1]

                output = model(dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True)
                logits = output.hidden_states[-1]

                self.assertTrue(torch.allclose(logits_fa[1:], logits[1:], atol=4e-2, rtol=4e-2))

2814
2815
2816
2817
                # check with inference + dropout
                model.train()
                _ = model_fa(dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True)

2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_inference_padding_right(self):
        import torch

        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
                return

            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model_fa = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=True
                )
                model_fa.to(torch_device)

                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=False
                )
                model.to(torch_device)

                dummy_input = torch.LongTensor([[1, 2, 3, 4, 5]]).to(torch_device)
                dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1, 0]]).to(torch_device)

                logits = model(dummy_input, output_hidden_states=True).hidden_states[-1]
                logits_fa = model_fa(dummy_input, output_hidden_states=True).hidden_states[-1]

                self.assertTrue(torch.allclose(logits_fa, logits, atol=4e-2, rtol=4e-2))

                output_fa = model_fa(dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True)
                logits_fa = output_fa.hidden_states[-1]

                output = model(dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True)
                logits = output.hidden_states[-1]

                self.assertTrue(torch.allclose(logits_fa[:-1], logits[:-1], atol=4e-2, rtol=4e-2))

    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_generate_left_padding(self):
        import torch

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
                return

            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.float16, use_flash_attention_2=False, low_cpu_mem_usage=True
                ).to(torch_device)

                dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
                dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [0, 1, 1, 1]]).to(torch_device)

                out = model.generate(
                    dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
                )

                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.float16, use_flash_attention_2=True, low_cpu_mem_usage=True
                ).to(torch_device)

                out_fa = model.generate(
                    dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
                )

                self.assertTrue(torch.equal(out, out_fa))

    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_generate_padding_right(self):
        import torch

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
                return

            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.float16, use_flash_attention_2=False, low_cpu_mem_usage=True
                ).to(torch_device)

                dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
                dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [1, 1, 1, 0]]).to(torch_device)

                out = model.generate(
                    dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
                )

                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.float16, use_flash_attention_2=True, low_cpu_mem_usage=True
                ).to(torch_device)

                out_fa = model.generate(
                    dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
                )

                self.assertTrue(torch.equal(out, out_fa))

2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_generate_use_cache(self):
        import torch

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
                return

            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
            model = model_class(config)

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

                dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
2952
                dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [0, 1, 1, 1]]).to(torch_device)
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962

                model = model_class.from_pretrained(
                    tmpdirname, torch_dtype=torch.float16, use_flash_attention_2=True, low_cpu_mem_usage=True
                ).to(torch_device)

                # Just test that a large cache works as expected
                _ = model.generate(
                    dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=30, do_sample=False
                )

2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
    @require_flash_attn
    @require_torch_gpu
    @require_bitsandbytes
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_fp32_ln(self):
        import torch

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
                return

            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
            model = model_class(config)

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

                dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
                dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [0, 1, 1, 1]]).to(torch_device)

                model = model_class.from_pretrained(
                    tmpdirname,
                    torch_dtype=torch.float16,
                    use_flash_attention_2=True,
                    low_cpu_mem_usage=True,
                    load_in_4bit=True,
                )

                for _, param in model.named_parameters():
                    # upcast only layer norms
                    if (param.dtype == torch.float16) or (param.dtype == torch.bfloat16):
                        param.data = param.data.to(torch.float32)

                _ = model(input_ids=dummy_input)

                # with attention mask
                _ = model(input_ids=dummy_input, attention_mask=dummy_attention_mask)

3002

3003
global_rng = random.Random()
thomwolf's avatar
thomwolf committed
3004
3005


thomwolf's avatar
thomwolf committed
3006
def ids_tensor(shape, vocab_size, rng=None, name=None):
3007
    #  Creates a random int32 tensor of the shape within the vocab size
thomwolf's avatar
thomwolf committed
3008
    if rng is None:
3009
        rng = global_rng
thomwolf's avatar
thomwolf committed
3010

thomwolf's avatar
thomwolf committed
3011
3012
3013
    total_dims = 1
    for dim in shape:
        total_dims *= dim
thomwolf's avatar
thomwolf committed
3014

thomwolf's avatar
thomwolf committed
3015
3016
3017
    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))
thomwolf's avatar
thomwolf committed
3018

3019
    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
thomwolf's avatar
thomwolf committed
3020
3021


3022
3023
3024
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
3025
3026
    # we choose the 1st token so this property of `at least one being non-zero` still holds after applying causal mask
    attn_mask[:, 0] = 1
3027
3028
3029
    return attn_mask


3030
def floats_tensor(shape, scale=1.0, rng=None, name=None):
Patrick von Platen's avatar
Patrick von Platen committed
3031
    """Creates a random float32 tensor"""
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
    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)

3043
    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()