test_modeling_common.py 179 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.
Sylvain Gugger's avatar
Sylvain Gugger committed
15
import collections
16
import copy
17
import gc
18
import inspect
19
import os
20
import os.path
21
import pickle
Aymeric Augustin's avatar
Aymeric Augustin committed
22
import random
Sylvain Gugger's avatar
Sylvain Gugger committed
23
import re
24
import tempfile
25
import warnings
26
from collections import defaultdict
NielsRogge's avatar
NielsRogge committed
27
from typing import Dict, List, Tuple
thomwolf's avatar
thomwolf committed
28

29
import numpy as np
30
from parameterized import parameterized
31
from pytest import mark
32
33

import transformers
34
35
from transformers import (
    AutoModel,
36
    AutoModelForCausalLM,
37
38
    AutoModelForSequenceClassification,
    PretrainedConfig,
39
    PreTrainedModel,
40
41
    is_torch_available,
    logging,
42
    set_seed,
43
)
44
from transformers.models.auto import get_values
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
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
65
66
from transformers.testing_utils import (
    CaptureLogger,
67
68
    is_pt_flax_cross_test,
    is_pt_tf_cross_test,
69
    require_accelerate,
70
    require_bitsandbytes,
71
    require_flash_attn,
72
    require_safetensors,
Sylvain Gugger's avatar
Sylvain Gugger committed
73
    require_torch,
74
    require_torch_gpu,
Sylvain Gugger's avatar
Sylvain Gugger committed
75
    require_torch_multi_gpu,
76
    require_torch_sdpa,
Sylvain Gugger's avatar
Sylvain Gugger committed
77
78
79
    slow,
    torch_device,
)
80
from transformers.utils import (
81
82
    CONFIG_NAME,
    GENERATION_CONFIG_NAME,
83
    SAFE_WEIGHTS_NAME,
84
    is_accelerate_available,
85
86
87
    is_flax_available,
    is_tf_available,
    is_torch_fx_available,
88
    is_torch_sdpa_available,
89
)
90
from transformers.utils.generic import ContextManagers, ModelOutput
91

Aymeric Augustin's avatar
Aymeric Augustin committed
92

93
94
95
96
if is_accelerate_available():
    from accelerate.utils import compute_module_sizes


97
if is_torch_available():
98
    import torch
99
    from safetensors.torch import load_file as safe_load_file
100
    from safetensors.torch import save_file as safe_save_file
101
    from torch import nn
thomwolf's avatar
thomwolf committed
102

103
    from transformers import MODEL_MAPPING, AdaptiveEmbedding
104
    from transformers.modeling_utils import load_state_dict, no_init_weights
Sylvain Gugger's avatar
Sylvain Gugger committed
105
    from transformers.pytorch_utils import id_tensor_storage
thomwolf's avatar
thomwolf committed
106

Sylvain Gugger's avatar
Sylvain Gugger committed
107

108
109
110
if is_tf_available():
    import tensorflow as tf

111
112
if is_flax_available():
    import jax.numpy as jnp
113

114
    from tests.test_modeling_flax_utils import check_models_equal
115
116
117
118
119
    from transformers.modeling_flax_pytorch_utils import (
        convert_pytorch_state_dict_to_flax,
        load_flax_weights_in_pytorch_model,
    )

120
if is_torch_fx_available():
121
    from transformers.utils.fx import symbolic_trace
122

123

124
125
126
def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
127
        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
128
            setattr(configs_no_init, key, 1e-10)
129
130
131
        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)
132
133
    return configs_no_init

thomwolf's avatar
thomwolf committed
134

135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
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()


160
161
162
163
@require_torch
class ModelTesterMixin:
    model_tester = None
    all_model_classes = ()
164
    all_generative_model_classes = ()
165
    fx_compatible = False
Patrick von Platen's avatar
Patrick von Platen committed
166
167
168
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
169
    test_resize_position_embeddings = False
Patrick von Platen's avatar
Patrick von Platen committed
170
    test_head_masking = True
171
    test_mismatched_shapes = True
172
    test_missing_keys = True
173
    test_model_parallel = False
174
    is_encoder_decoder = False
175
    has_attentions = True
176
    model_split_percents = [0.5, 0.7, 0.9]
177

178
179
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
180
        if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
181
            inputs_dict = {
182
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
183
                if isinstance(v, torch.Tensor) and v.ndim > 1
Sylvain Gugger's avatar
Sylvain Gugger committed
184
                else v
185
186
                for k, v in inputs_dict.items()
            }
187
        elif model_class.__name__ in get_values(MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES):
188
            inputs_dict.pop("attention_mask")
189
190

        if return_labels:
191
            if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
192
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
193
194
195
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
196
            ]:
197
198
199
200
201
202
                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
                )
203
204
205
206
207
208
            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),
209
            ]:
210
211
212
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
213
214
215
216
217
218
            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),
219
220
221
222
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
223
            elif model_class.__name__ in get_values(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
224
225
226
227
                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
                )
228
            elif model_class.__name__ in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
229
230
231
232
                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()
233

234
235
        return inputs_dict

Patrick von Platen's avatar
Patrick von Platen committed
236
    def test_save_load(self):
237
238
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

239
240
241
242
243
244
245
246
247
248
        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)

249
250
251
252
253
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
254
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
255

256
            with tempfile.TemporaryDirectory() as tmpdirname:
257
                model.save_pretrained(tmpdirname)
258
259
260
261
262
263
264

                # 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))
                )

265
                model = model_class.from_pretrained(tmpdirname)
266
                model.to(torch_device)
267
                with torch.no_grad():
268
                    second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
thomwolf's avatar
thomwolf committed
269

270
271
272
273
274
            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)
275

276
277
278
279
280
281
282
283
284
285
286
287
    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))

288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
    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)

306
    def test_save_load_keys_to_ignore_on_save(self):
307
308
309
310
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
311
312
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
313
314
315
                continue

            # check the keys are in the original state_dict
316
            for k in _keys_to_ignore_on_save:
317
                self.assertIn(k, model.state_dict().keys(), "\n".join(model.state_dict().keys()))
318
319
320
321

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
322
323
324
                output_model_file = os.path.join(tmpdirname, SAFE_WEIGHTS_NAME)
                state_dict_saved = safe_load_file(output_model_file)

325
                for k in _keys_to_ignore_on_save:
326
                    self.assertNotIn(k, state_dict_saved.keys(), "\n".join(state_dict_saved.keys()))
327

Sylvain Gugger's avatar
Sylvain Gugger committed
328
329
                # 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)
330
331
332
333
334
335
                keys_to_ignore = set(model._keys_to_ignore_on_save)

                if hasattr(model, "_tied_weights_keys"):
                    keys_to_ignore.update(set(model._tied_weights_keys))

                self.assertTrue(len(load_result.missing_keys) == 0 or set(load_result.missing_keys) == keys_to_ignore)
Sylvain Gugger's avatar
Sylvain Gugger committed
336
337
                self.assertTrue(len(load_result.unexpected_keys) == 0)

338
339
340
341
342
343
344
345
346
347
348
    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)

349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
    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)

364
365
366
367
368
369
370
            # Loop over all modules and check that relevant modules have gradient_checkpointing set to True
            for n, m in model.named_modules():
                if hasattr(m, "gradient_checkpointing"):
                    self.assertTrue(
                        m.gradient_checkpointing, f"Module {n} does not have gradient_checkpointing set to True"
                    )

371
372
373
374
            # check disable works
            model.gradient_checkpointing_disable()
            self.assertFalse(model.is_gradient_checkpointing)

375
376
377
378
379
380
381
            # Loop over all modules and check that relevant modules have gradient_checkpointing set to False
            for n, m in model.named_modules():
                if hasattr(m, "gradient_checkpointing"):
                    self.assertFalse(
                        m.gradient_checkpointing, f"Module {n} does not have gradient_checkpointing set to False"
                    )

382
383
    def test_save_load_fast_init_from_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
384
385
        if config.__class__ not in MODEL_MAPPING:
            return
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
        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
407
408
            model_class_copy._init_weights = _mock_init_weights
            model_class_copy.init_weights = _mock_all_init_weights
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424

            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)
425
                # Before we test anything
426
427

                for key in model_fast_init.state_dict().keys():
428
429
430
431
432
                    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")
433

434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
    def test_fast_init_context_manager(self):
        # 1. Create a dummy class. Should have buffers as well? To make sure we test __init__
        class MyClass(PreTrainedModel):
            config_class = PretrainedConfig

            def __init__(self, config=None):
                super().__init__(config if config is not None else PretrainedConfig())
                self.linear = nn.Linear(10, 10, bias=True)
                self.embedding = nn.Embedding(10, 10)
                self.std = 1

            def _init_weights(self, module):
                if isinstance(module, nn.Linear):
                    module.weight.data = nn.init.kaiming_uniform_(module.weight.data, np.sqrt(5))
                    if module.bias is not None:
                        module.bias.data.normal_(mean=0.0, std=self.std)

        # 2. Make sure a linear layer's reset params is properly skipped:
        with ContextManagers([no_init_weights(True)]):
            no_init_instance = MyClass()

        set_seed(0)
        expected_bias = torch.tensor(
            ([0.2975, 0.2131, -0.1379, -0.0796, -0.3012, -0.0057, -0.2381, -0.2439, -0.0174, 0.0475])
        )
        init_instance = MyClass()
        torch.testing.assert_allclose(init_instance.linear.bias, expected_bias, rtol=1e-3, atol=1e-4)

        set_seed(0)
        torch.testing.assert_allclose(
            init_instance.linear.weight, nn.init.kaiming_uniform_(no_init_instance.linear.weight, np.sqrt(5))
        )

        # 3. Make sure weights that are not present use init_weight_ and get expected values
        with tempfile.TemporaryDirectory() as tmpdirname:
            state_dict = init_instance.state_dict()
            del state_dict["linear.weight"]

            init_instance.config.save_pretrained(tmpdirname)
            torch.save(state_dict, os.path.join(tmpdirname, "pytorch_model.bin"))
            set_seed(0)
            model_fast_init = MyClass.from_pretrained(tmpdirname)

            set_seed(0)
            model_slow_init = MyClass.from_pretrained(tmpdirname, _fast_init=False)

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

484
485
    def test_save_load_fast_init_to_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
486
487
        if config.__class__ not in MODEL_MAPPING:
            return
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
        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
509
510
            base_class_copy._init_weights = _mock_init_weights
            base_class_copy.init_weights = _mock_all_init_weights
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528

            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():
529
530
531
532
533
534
535
536
537
                    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")
538

539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
    def test_torch_save_load(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        if config.__class__ not in MODEL_MAPPING:
            return
        base_class = MODEL_MAPPING[config.__class__]

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

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

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

            base_class_copy = CopyClass

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

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

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

            def check_equal(loaded):
                for key in state_dict.keys():
                    max_diff = torch.max(
                        state_dict()[key] ^ loaded[key]
                        if isinstance(state_dict[key], torch.BoolTensor)
                        else torch.abs(state_dict[key] - loaded[key])
                    ).item()
                    self.assertLessEqual(max_diff, 1e-6, msg=f"{key} not identical")

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                pt_checkpoint_path = os.path.join(tmpdirname, "pytorch_model.bin")
                torch.save(state_dict, pt_checkpoint_path, _use_new_zipfile_serialization=True)
                check_equal(load_state_dict(pt_checkpoint_path))
                torch.save(state_dict, pt_checkpoint_path, _use_new_zipfile_serialization=False)
                check_equal(load_state_dict(pt_checkpoint_path))

Patrick von Platen's avatar
Patrick von Platen committed
587
    def test_initialization(self):
588
589
590
591
592
593
594
595
        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
596
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
597
                        [0.0, 1.0],
598
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
599
                    )
thomwolf's avatar
thomwolf committed
600

Patrick von Platen's avatar
Patrick von Platen committed
601
    def test_determinism(self):
602
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
603
604
605
606
607
608
609
610
611

        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)

612
613
614
615
616
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
617
618
                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
619

620
621
622
623
624
            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_determinism(tensor1, tensor2)
            else:
                check_determinism(first, second)
625

626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
    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",
                ]
642
                expected_arg_names.extend(
643
644
                    ["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
645
646
647
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
648
649
650
651
652
653
            elif model_class.__name__ in [*get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)] and self.has_attentions:
                expected_arg_names = ["pixel_values", "output_hidden_states", "output_attentions", "return_dict"]
                self.assertListEqual(arg_names, expected_arg_names)
            elif model_class.__name__ in [*get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)] and not self.has_attentions:
                expected_arg_names = ["pixel_values", "output_hidden_states", "return_dict"]
                self.assertListEqual(arg_names, expected_arg_names)
654
            else:
655
                expected_arg_names = [model.main_input_name]
656
657
                self.assertListEqual(arg_names[:1], expected_arg_names)

658
    def check_training_gradient_checkpointing(self, gradient_checkpointing_kwargs=None):
659
660
661
662
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
663
664
665
666
667
            if (
                model_class.__name__
                in [*get_values(MODEL_MAPPING_NAMES), *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)]
                or not model_class.supports_gradient_checkpointing
            ):
668
                continue
669

670
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
671
672
            config.use_cache = False
            config.return_dict = True
673
            model = model_class(config)
674

675
            model.to(torch_device)
676
            model.gradient_checkpointing_enable(gradient_checkpointing_kwargs=gradient_checkpointing_kwargs)
677
            model.train()
678
679
680
681
682
683
684

            # unfreeze additional layers
            for p in model.parameters():
                p.requires_grad_(True)

            optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

685
686
687
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()
688
            optimizer.step()
689

690
691
692
693
694
            for k, v in model.named_parameters():
                if v.requires_grad:
                    self.assertTrue(v.grad is not None, f"{k} in {model_class.__name__} has no gradient!")

    def test_training(self):
695
        if not self.model_tester.is_training:
696
697
698
            return

        for model_class in self.all_model_classes:
699
700
701
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.return_dict = True

702
703
704
705
            if model_class.__name__ in [
                *get_values(MODEL_MAPPING_NAMES),
                *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES),
            ]:
706
                continue
707

708
709
710
711
712
713
714
            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()

715
716
717
718
719
720
721
722
723
724
725
726
727
    def test_training_gradient_checkpointing(self):
        # Scenario - 1 default behaviour
        self.check_training_gradient_checkpointing()

    def test_training_gradient_checkpointing_use_reentrant(self):
        # Scenario - 2 with `use_reentrant=True` - this is the default value that is used in pytorch's
        # torch.utils.checkpoint.checkpoint
        self.check_training_gradient_checkpointing(gradient_checkpointing_kwargs={"use_reentrant": True})

    def test_training_gradient_checkpointing_use_reentrant_false(self):
        # Scenario - 3 with `use_reentrant=False` pytorch suggests users to use this value for
        # future releases: https://pytorch.org/docs/stable/checkpoint.html
        self.check_training_gradient_checkpointing(gradient_checkpointing_kwargs={"use_reentrant": False})
728

Patrick von Platen's avatar
Patrick von Platen committed
729
    def test_attention_outputs(self):
730
731
732
        if not self.has_attentions:
            self.skipTest(reason="Model does not output attentions")

733
734
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True
735

736
737
738
739
740
741
742
743
744
745
746
747
        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
748
            config.return_dict = True
749
750
751
752
753
754
755
            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)
756

757
758
759
760
761
762
763
764
765
766
            # 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
767

768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
            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
787
788
789
                if model_class.__name__ in [
                    *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                    *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
790
                ]:
791
792
793
794
795
796
797
798
799
800
801
802
803
804
                    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],
                )
805

806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
                # 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
849

850
    @slow
851
    def test_torchscript_simple(self):
852
853
        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
854

855
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
856
    def test_torchscript_output_attentions(self):
857
858
859
        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
860

861
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
862
    def test_torchscript_output_hidden_state(self):
863
864
865
        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
866

867
868
869
870
    # 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()
871
872
873
        # 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()
874

875
    def _create_and_check_torchscript(self, config, inputs_dict):
Patrick von Platen's avatar
Patrick von Platen committed
876
        if not self.test_torchscript:
877
            return
878

879
880
881
        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:
882
            for attn_implementation in ["eager", "sdpa"]:
883
                if attn_implementation == "sdpa" and (not model_class._supports_sdpa or not is_torch_sdpa_available()):
884
                    continue
885

886
887
888
889
890
                configs_no_init._attn_implementation = attn_implementation
                model = model_class(config=configs_no_init)
                model.to(torch_device)
                model.eval()
                inputs = self._prepare_for_class(inputs_dict, model_class)
thomwolf's avatar
thomwolf committed
891

892
                main_input_name = model_class.main_input_name
thomwolf's avatar
thomwolf committed
893

894
                try:
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
                    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
                        main_input = inputs[main_input_name]
                        attention_mask = inputs["attention_mask"]
                        decoder_input_ids = inputs["decoder_input_ids"]
                        decoder_attention_mask = inputs["decoder_attention_mask"]
                        model(main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
                        traced_model = torch.jit.trace(
                            model, (main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
                        )
                    elif "bbox" in inputs and "image" in inputs:  # LayoutLMv2 requires additional inputs
                        input_ids = inputs["input_ids"]
                        bbox = inputs["bbox"]
                        image = inputs["image"].tensor
                        model(input_ids, bbox, image)
                        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
                    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
                    else:
                        main_input = inputs[main_input_name]

                        if model.config._attn_implementation == "sdpa":
                            trace_input = {main_input_name: main_input}

                            if "attention_mask" in inputs:
                                trace_input["attention_mask"] = inputs["attention_mask"]
                            else:
                                self.skipTest("testing SDPA without attention_mask is not supported")

                            model(main_input, attention_mask=inputs["attention_mask"])
                            # example_kwarg_inputs was introduced in torch==2.0, but it is fine here since SDPA has a requirement on torch>=2.1.
                            traced_model = torch.jit.trace(model, example_kwarg_inputs=trace_input)
                        else:
                            model(main_input)
                            traced_model = torch.jit.trace(model, (main_input,))
                except RuntimeError:
                    self.fail("Couldn't trace module.")

                with tempfile.TemporaryDirectory() as tmp_dir_name:
                    pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")

                    try:
                        torch.jit.save(traced_model, pt_file_name)
                    except Exception:
                        self.fail("Couldn't save module.")

                    try:
                        loaded_model = torch.jit.load(pt_file_name)
                    except Exception:
                        self.fail("Couldn't load module.")
LysandreJik's avatar
LysandreJik committed
952

953
954
                model.to(torch_device)
                model.eval()
thomwolf's avatar
thomwolf committed
955

956
957
                loaded_model.to(torch_device)
                loaded_model.eval()
thomwolf's avatar
thomwolf committed
958

959
960
                model_state_dict = model.state_dict()
                loaded_model_state_dict = loaded_model.state_dict()
961

962
963
964
965
                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]
966

967
968
969
                loaded_model_state_dict = {
                    key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
                }
970

971
                self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
thomwolf's avatar
thomwolf committed
972

973
974
975
976
977
978
979
                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
980

981
982
                    self.assertTrue(found_buffer)
                    model_buffers.pop(i)
983

984
985
986
987
988
989
                models_equal = True
                for layer_name, p1 in model_state_dict.items():
                    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
990

991
                self.assertTrue(models_equal)
thomwolf's avatar
thomwolf committed
992

993
994
995
                # 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()
996

997
998
999
1000
1001
1002
1003
1004
    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)

1005
1006
    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:
1007
1008
1009
1010
1011
            return

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

1012
        for model_class in self.all_model_classes:
1013
1014
1015
1016
1017
1018
1019
1020
1021
            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)
1022
1023
1024
                    input_names = [
                        "attention_mask",
                        "decoder_attention_mask",
1025
                        "decoder_input_ids",
1026
                        "input_features",
1027
1028
                        "input_ids",
                        "input_values",
1029
                    ]
1030
1031
                    if labels is not None:
                        input_names.append("labels")
1032

1033
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
1034
                    input_names = list(filtered_inputs.keys())
1035

1036
                    model_output = model(**filtered_inputs)
1037

1038
                    traced_model = symbolic_trace(model, input_names)
1039
                    traced_output = traced_model(**filtered_inputs)
1040
                else:
1041
1042
1043
1044
                    input_names = [
                        "attention_mask",
                        "bbox",
                        "input_features",
1045
1046
1047
1048
1049
1050
                        "input_ids",
                        "input_values",
                        "pixel_values",
                        "token_type_ids",
                        "visual_feats",
                        "visual_pos",
1051
                    ]
1052

1053
                    labels = inputs.get("labels", None)
1054
1055
                    start_positions = inputs.get("start_positions", None)
                    end_positions = inputs.get("end_positions", None)
1056
1057
                    if labels is not None:
                        input_names.append("labels")
1058
1059
1060
1061
                    if start_positions is not None:
                        input_names.append("start_positions")
                    if end_positions is not None:
                        input_names.append("end_positions")
1062

1063
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
1064
                    input_names = list(filtered_inputs.keys())
1065

1066
                    if model.__class__.__name__ in set(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES.values()) and (
1067
                        not hasattr(model.config, "problem_type") or model.config.problem_type is None
1068
1069
1070
                    ):
                        model.config.problem_type = "single_label_classification"

1071
                    traced_model = symbolic_trace(model, input_names)
1072
                    traced_output = traced_model(**filtered_inputs)
1073
                    model_output = model(**filtered_inputs)
1074

1075
            except Exception as e:
1076
                self.fail(f"Couldn't trace module: {e}")
1077

1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
            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)
1091
            num_outputs = len(model_output)
1092
1093
1094
1095
1096
1097

            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}",
                )
1098

1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
            # 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}",
                    )

1119
1120
1121
1122
            # 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
1123
1124
    def test_headmasking(self):
        if not self.test_head_masking:
1125
            return
1126

1127
1128
1129
        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()
LysandreJik's avatar
LysandreJik committed
1130

1131
        inputs_dict["output_attentions"] = True
1132
1133
1134
1135
1136
1137
        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
1138

1139
1140
1141
            # 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
1142
1143
1144
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
1145
1146
1147
1148
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
1149
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
1150
            inputs["head_mask"] = head_mask
1151
1152
1153
1154
1155
            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
1156
1157
                if "cross_attn_head_mask" in arg_names:
                    inputs["cross_attn_head_mask"] = head_mask
1158
            outputs = model(**inputs, return_dict=True)
1159
1160
1161
1162
1163
1164
1165
1166
1167

            # 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)
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188

            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)
1189
                check_attentions_validity(outputs.cross_attentions)
1190
1191
            else:
                check_attentions_validity(outputs.attentions)
1192

Patrick von Platen's avatar
Patrick von Platen committed
1193
1194
    def test_head_pruning(self):
        if not self.test_pruning:
1195
1196
1197
            return

        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1198
1199
1200
1201
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1202

1203
1204
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1205

1206
            inputs_dict["output_attentions"] = True
1207
1208
1209
1210
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1211
1212
1213
1214
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1215
1216
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
1217
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1218

1219
            attentions = outputs[-1]
1220

1221
            self.assertEqual(attentions[0].shape[-3], 1)
1222
1223
            # 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)
1224
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
LysandreJik's avatar
LysandreJik committed
1225

Patrick von Platen's avatar
Patrick von Platen committed
1226
1227
    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
1228
            return
LysandreJik's avatar
LysandreJik committed
1229

1230
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1231
1232
1233
1234
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1235
1236
1237

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1238

1239
            inputs_dict["output_attentions"] = True
1240
1241
1242
1243
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1244
1245
1246
1247
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1248
            model.prune_heads(heads_to_prune)
1249

1250
            with tempfile.TemporaryDirectory() as temp_dir_name:
1251
1252
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1253
                model.to(torch_device)
1254

1255
            with torch.no_grad():
1256
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1257
1258
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
1259
1260
            # 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)
1261
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1262

Patrick von Platen's avatar
Patrick von Platen committed
1263
1264
    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
1265
            return
1266

1267
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1268
1269
1270
1271
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1272

1273
1274
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1275

1276
            inputs_dict["output_attentions"] = True
1277
            config.output_hidden_states = False
1278

1279
1280
1281
1282
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1283
            config.pruned_heads = heads_to_prune
1284

1285
1286
1287
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1288

1289
            with torch.no_grad():
1290
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1291
            attentions = outputs[-1]
1292

1293
            self.assertEqual(attentions[0].shape[-3], 1)
1294
1295
            # 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)
1296
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1297

Patrick von Platen's avatar
Patrick von Platen committed
1298
1299
    def test_head_pruning_integration(self):
        if not self.test_pruning:
1300
            return
1301

1302
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1303
1304
1305
1306
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1307

1308
1309
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1310

1311
            inputs_dict["output_attentions"] = True
1312
            config.output_hidden_states = False
1313

1314
            heads_to_prune = {1: [1, 2]}
1315
            config.pruned_heads = heads_to_prune
1316

1317
1318
1319
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1320

1321
            with torch.no_grad():
1322
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1323
            attentions = outputs[-1]
1324

1325
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1326
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1327

1328
            with tempfile.TemporaryDirectory() as temp_dir_name:
1329
1330
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1331
                model.to(torch_device)
thomwolf's avatar
thomwolf committed
1332

1333
            with torch.no_grad():
1334
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1335
            attentions = outputs[-1]
LysandreJik's avatar
LysandreJik committed
1336

1337
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1338
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1339

1340
            heads_to_prune = {0: [0], 1: [1, 2]}
1341
            model.prune_heads(heads_to_prune)
1342

1343
            with torch.no_grad():
1344
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1345
            attentions = outputs[-1]
1346

1347
1348
            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)
1349

1350
            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2]})
thomwolf's avatar
thomwolf committed
1351

Patrick von Platen's avatar
Patrick von Platen committed
1352
    def test_hidden_states_output(self):
Joseph Liu's avatar
Joseph Liu committed
1353
        def check_hidden_states_output(inputs_dict, config, model_class):
1354
            model = model_class(config)
1355
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
1356
            model.eval()
Joseph Liu's avatar
Joseph Liu committed
1357

thomwolf's avatar
thomwolf committed
1358
            with torch.no_grad():
1359
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1360
1361

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
1362

Sylvain Gugger's avatar
Sylvain Gugger committed
1363
1364
1365
1366
            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)
1367

Patrick von Platen's avatar
Patrick von Platen committed
1368
1369
1370
1371
1372
1373
1374
            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

1375
            self.assertListEqual(
Lysandre's avatar
Lysandre committed
1376
1377
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
1378
            )
thomwolf's avatar
thomwolf committed
1379

1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
            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
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
        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)

1405
1406
1407
    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
1408
        config.output_attentions = self.has_attentions
1409
1410
1411
1412
1413
1414
1415
1416
1417

        # 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)
1418

1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
        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()

1429
1430
1431
1432
1433
1434
1435
1436
1437
            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()
1438
1439
1440
1441
1442

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

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)
1443
1444
1445
1446
1447

            if self.has_attentions:
                self.assertIsNotNone(encoder_attentions.grad)
                self.assertIsNotNone(decoder_attentions.grad)
                self.assertIsNotNone(cross_attentions.grad)
1448
1449
1450
1451
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            hidden_states.retain_grad()
1452
1453
1454
1455

            if self.has_attentions:
                attentions = outputs.attentions[0]
                attentions.retain_grad()
1456
1457
1458
1459

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

            self.assertIsNotNone(hidden_states.grad)
1460
1461
1462

            if self.has_attentions:
                self.assertIsNotNone(attentions.grad)
1463

Pradhy729's avatar
Pradhy729 committed
1464
    def test_feed_forward_chunking(self):
Lysandre's avatar
Lysandre committed
1465
1466
1467
1468
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Pradhy729's avatar
Pradhy729 committed
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
        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))

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
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
    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
1566
    def test_resize_tokens_embeddings(self):
Lysandre's avatar
Lysandre committed
1567
1568
1569
1570
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Patrick von Platen's avatar
Patrick von Platen committed
1571
        if not self.test_resize_embeddings:
1572
1573
1574
1575
1576
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
1577
            model.to(torch_device)
1578

Patrick von Platen's avatar
Patrick von Platen committed
1579
1580
1581
            if self.model_tester.is_training is False:
                model.eval()

1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
            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)
1592
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
1593
            model(**self._prepare_for_class(inputs_dict, model_class))
1594
1595
1596
1597
1598
1599
1600

            # 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)

1601
1602
1603
            # 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)
1604
1605
1606
1607

            # 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)
1608
            model(**self._prepare_for_class(inputs_dict, model_class))
1609

1610
1611
1612
1613
1614
1615
1616
1617
            # 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)

1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
            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
1629
1630
1631
            self.assertTrue(model_embed.weight.shape[0], model.config.vocab_size)
            self.assertTrue(model.config.vocab_size, model.vocab_size)

1632
1633
1634
            model_embed = model.resize_token_embeddings(model_vocab_size + 13, pad_to_multiple_of=64)
            self.assertTrue(model_embed.weight.shape[0] // 64, 0)

1635
1636
1637
1638
1639
            # 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)

1640
1641
1642
1643
1644
1645
            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)

1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
    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
1697
    def test_model_common_attributes(self):
1698
1699
1700
1701
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1702
1703
            self.assertIsInstance(model.get_input_embeddings(), (nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(nn.Embedding(10, 10))
1704
            x = model.get_output_embeddings()
1705
            self.assertTrue(x is None or isinstance(x, nn.Linear))
1706

1707
1708
1709
1710
1711
1712
1713
    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)

1714
    def test_correct_missing_keys(self):
1715
1716
        if not self.test_missing_keys:
            return
1717
1718
1719
1720
1721
1722
1723
        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):
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
                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

1737
1738
1739
                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)
1740
                    self.assertGreater(len(loading_info["missing_keys"]), 0, model.__class__.__name__)
1741

1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
    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))

1790
1791
    @require_safetensors
    def test_can_use_safetensors(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
1792
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
        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
1809
1810
                # Checking there was no complain of missing weights
                self.assertEqual(infos["missing_keys"], [])
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826

                # 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
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
    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
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
    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
1871
1872
1873
1874
1875
            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
1876

Sylvain Gugger's avatar
Sylvain Gugger committed
1877
1878
    def test_model_weights_reload_no_missing_tied_weights(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1879
        for model_class in self.all_model_classes:
Sylvain Gugger's avatar
Sylvain Gugger committed
1880
1881
1882
            model = model_class(config)
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.save_pretrained(tmp_dir)
1883
1884
1885

                # 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.
1886
1887
                placeholder_dict = {"tensor": torch.tensor([1, 2])}
                safe_save_file(placeholder_dict, os.path.join(tmp_dir, "model.safetensors"), metadata={"format": "pt"})
Sylvain Gugger's avatar
Sylvain Gugger committed
1888
                model_reloaded, infos = model_class.from_pretrained(tmp_dir, output_loading_info=True)
1889
1890
1891
1892

                prefix = f"{model_reloaded.base_model_prefix}."
                params = dict(model_reloaded.named_parameters())
                params.update(dict(model_reloaded.named_buffers()))
1893
                param_names = {k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys()}
1894
1895
1896
1897

                missing_keys = set(infos["missing_keys"])

                extra_missing = missing_keys - param_names
Sylvain Gugger's avatar
Sylvain Gugger committed
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
                # 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)
1909
1910
1911
1912

                self.assertEqual(
                    extra_missing,
                    set(),
Sylvain Gugger's avatar
Sylvain Gugger committed
1913
1914
                    f"This model {model_class.__name__} might be missing some `keys_to_ignore`: {extra_missing}. "
                    f"For debugging, tied parameters are {tied_params}",
1915
1916
                )

Sylvain Gugger's avatar
Sylvain Gugger committed
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
                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`",
                )
1937

1938
1939
1940
    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
1941
1942
1943
1944
        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

1945
1946
1947
1948
1949
1950
1951
1952
1953
        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
1954
1955
1956
1957
1958
                    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)
1959
1960
1961
1962
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
Sam Shleifer's avatar
Sam Shleifer committed
1963
1964
1965
                            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
1966
1967
1968
1969
1970
1971
                            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)}."
                            ),
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
                        )

                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})

1997
1998
1999
2000
            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})
2001

2002
2003
2004
2005
2006
2007
2008
2009
2010
                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}
                )
2011

2012
2013
2014
2015
    # 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"""
2016

2017
2018
2019
        for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
            if k in inputs_dict:
                attention_mask = inputs_dict[k]
2020

2021
2022
2023
2024
2025
2026
                # 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
                )
2027

2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
                # 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)"""

2045
2046
        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}
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072

        key_differences = tf_keys.symmetric_difference(pt_keys)

        if model_class.__name__ in [
            "FlaubertWithLMHeadModel",
            "FunnelForPreTraining",
            "ElectraForPreTraining",
            "XLMWithLMHeadModel",
        ]:
            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):
2073
        """Check the outputs from PyTorch and TensorFlow models are close enough. Checks are done in a recursive way.
2074

2075
2076
2077
2078
2079
2080
2081
2082
        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.
        """
2083

2084
2085
2086
        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")
2087

2088
2089
2090
2091
2092
2093
        # 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",
            )
2094

2095
2096
2097
            # 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)
2098

2099
2100
            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]
2101

2102
            self.assertEqual(tf_keys, pt_keys, f"{name}: Output keys differ between TF and PyTorch")
2103

2104
            # convert to the case of `tuple`
2105
            # appending each key to the current (string) `name`
2106
2107
2108
2109
            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
            )
2110

2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
        # 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),
2121
                    f"{name}: The tuple `attributes` should have the same length as `tf_outputs`",
2122
                )
2123
            else:
2124
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
2125
                attributes = tuple([f"{name}_{idx}" for idx in range(len(tf_outputs))])
2126

2127
2128
            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)
2129

2130
2131
2132
2133
        elif isinstance(tf_outputs, tf.Tensor):
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `tf_outputs` is"
            )
2134

2135
2136
            tf_outputs = tf_outputs.numpy()
            pt_outputs = pt_outputs.detach().to("cpu").numpy()
2137

2138
2139
2140
            self.assertEqual(
                tf_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between TF and PyTorch"
            )
2141

2142
2143
2144
2145
            # 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])
2146

2147
2148
            tf_nans = np.isnan(tf_outputs)
            pt_nans = np.isnan(pt_outputs)
2149

2150
2151
2152
2153
            pt_outputs[tf_nans] = 0
            tf_outputs[tf_nans] = 0
            pt_outputs[pt_nans] = 0
            tf_outputs[pt_nans] = 0
2154

2155
            max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
2156
            self.assertLessEqual(max_diff, tol, f"{name}: Difference between PyTorch and TF is {max_diff} (>= {tol}).")
2157
2158
        else:
            raise ValueError(
2159
                "`tf_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `tf.Tensor`. Got"
Sylvain Gugger's avatar
Sylvain Gugger committed
2160
                f" {type(tf_outputs)} instead."
2161
2162
            )

2163
2164
2165
2166
    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
2167
            if isinstance(tensor, bool):
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
                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)
2180

2181
        return tf_inputs_dict
2182

2183
2184
    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)
2185

2186
2187
2188
2189
        # 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()
        }
2190

2191
2192
        # send pytorch model to the correct device
        pt_model.to(torch_device)
2193

2194
2195
        # Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
        pt_model.eval()
2196

2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
        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
2211
    def test_pt_tf_model_equivalence(self, allow_missing_keys=False):
2212
        import transformers
2213
2214

        for model_class in self.all_model_classes:
2215
2216
2217
            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
2218
            if not hasattr(transformers, tf_model_class_name):
2219
                # transformers does not have this model in TF version yet
2220
2221
                return

2222
2223
2224
            # Output all for aggressive testing
            config.output_hidden_states = True
            config.output_attentions = self.has_attentions
2225

2226
2227
2228
2229
            # 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)
2230
2231

            tf_model_class = getattr(transformers, tf_model_class_name)
2232
2233

            pt_model = model_class(config)
2234
2235
2236
2237
2238
2239
2240
2241
2242
            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,
            )
2243
2244
2245
2246
2247
2248
2249
2250
2251

            # 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")

2252
            pt_inputs_dict = {k: v for k, v in pt_inputs_dict.items() if k in tf_input_keys}
2253
2254
2255
2256
            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.
2257
            if not set(pt_inputs_dict_with_labels.keys()).symmetric_difference(pt_inputs_dict.keys()):
2258
                pt_inputs_dict_with_labels = None
2259
2260

            # Check we can load pt model in tf and vice-versa with model => model functions
2261
2262
            # 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
2263
2264
2265
2266
2267
2268
            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
            )
2269

2270
2271
2272
2273
2274
            # 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)
2275
2276
2277
2278
2279

            # 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
2280
2281
2282
                tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(
                    tf_model, pt_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2283
2284
2285

                tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
                tf_model.save_weights(tf_checkpoint_path)
Matt's avatar
Matt committed
2286
2287
2288
                pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(
                    pt_model, tf_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2289

2290
2291
2292
2293
2294
            # 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)
2295
2296
2297
2298
2299

    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}).")

2300
    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
2301
2302
2303
2304
2305
2306
2307
2308
2309
        """
        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.
        """
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349

        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`",
                )
2350
            else:
2351
2352
2353
2354
2355
2356
                # 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)

2357
        elif isinstance(fx_outputs, jnp.ndarray):
2358
2359
2360
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `fx_outputs` is"
            )
2361
2362
2363
2364
2365

            # 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()

2366
2367
2368
2369
2370
2371
2372
2373
2374
            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])

2375
2376
2377
2378
2379
2380
2381
2382
            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

2383
2384
2385
2386
            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})."
            )
2387
2388
        else:
            raise ValueError(
2389
2390
                "`fx_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `jnp.ndarray`. Got"
                f" {type(fx_outputs)} instead."
2391
2392
            )

2393
2394
2395
2396
2397
2398
2399
2400
2401
    @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):
2402
                    # no flax model exists for this class
2403
2404
                    return

2405
2406
2407
2408
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2409
2410
                fx_model_class = getattr(transformers, fx_model_class_name)

2411
2412
2413
2414
2415
2416
                # 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

2417
2418
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2419

2420
2421
2422
2423
2424
2425
2426
2427
2428
                # 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}

2429
2430
2431
2432
2433
2434
                # 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
2435
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2436

2437
2438
2439
                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
                fx_model.params = fx_state

2440
2441
2442
                # send pytorch model to the correct device
                pt_model.to(torch_device)

2443
                with torch.no_grad():
2444
2445
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)
2446

2447
2448
2449
2450
                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)
2451
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2452
2453
2454
2455
2456

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

2457
2458
2459
2460
2461
2462
                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)
2463
                self.check_pt_flax_outputs(fx_outputs_loaded, pt_outputs, model_class)
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476

    @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

2477
2478
2479
2480
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2481
2482
                fx_model_class = getattr(transformers, fx_model_class_name)

2483
2484
2485
2486
2487
2488
                # 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

2489
2490
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2491

2492
2493
2494
2495
2496
2497
2498
2499
2500
                # 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}

2501
2502
2503
2504
                # 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()
                }
2505

2506
                # convert inputs to Flax
2507
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2508

2509
2510
2511
2512
2513
2514
2515
                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)
2516

2517
2518
2519
2520
2521
2522
2523
2524
                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)
2525
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2526
2527
2528
2529
2530

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

2531
2532
2533
2534
                # send pytorch model to the correct device
                pt_model_loaded.to(torch_device)
                pt_model_loaded.eval()

2535
                with torch.no_grad():
2536
                    pt_outputs_loaded = pt_model_loaded(**pt_inputs)
2537

2538
2539
2540
2541
                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)
2542
                self.check_pt_flax_outputs(fx_outputs, pt_outputs_loaded, model_class)
2543

Patrick von Platen's avatar
Patrick von Platen committed
2544
    def test_inputs_embeds(self):
2545
2546
2547
2548
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
2549
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
2550
            model.eval()
2551

2552
            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
2553

2554
2555
2556
2557
2558
2559
2560
2561
2562
            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)

2563
2564
            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
2565
                inputs["inputs_embeds"] = wte(input_ids)
2566
            else:
2567
2568
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
2569

thomwolf's avatar
thomwolf committed
2570
            with torch.no_grad():
Weizhen's avatar
Weizhen committed
2571
                model(**inputs)[0]
2572

2573
2574
    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
2575
2576
2577
2578
        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.
2579
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask", "cross_attn_head_mask"]
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
        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
2594
            model = nn.DataParallel(model)
2595
            with torch.no_grad():
2596
                _ = model(**self._prepare_for_class(inputs_dict, model_class))
2597

2598
2599
2600
    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
2601
            return
2602

2603
        # a candidate for testing_utils
2604
        def get_current_gpu_memory_use():
Patrick von Platen's avatar
Patrick von Platen committed
2605
            """returns a list of cuda memory allocations per GPU in MBs"""
2606
2607
2608
2609
2610

            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)
2611
2612
2613
2614
2615
2616
2617
2618
2619

            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()

2620
2621
2622
            # 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()
2623

2624
2625
            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
2626
2627
2628
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

2629
2630
2631
            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

2632
            del model
2633
            gc.collect()
2634
2635
            torch.cuda.empty_cache()

2636
2637
2638
            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()
2639
2640

            # Spread model layers over multiple devices
2641
            model = model_class(config)
2642
2643
2644
2645
            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
2646
            for n in range(len(model.device_map.keys())):
2647
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])
2648

2649
            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
2650
2651
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

2652
2653
            # 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
2654
2655
2656
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
2657
            gc.collect()
2658
2659
2660
2661
2662
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
2663
            return
2664
2665
2666
2667
2668
2669

        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)

2670
            def cast_to_device(dictionary, device):
2671
2672
2673
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
2674
                        output[k] = v.to(device)
2675
2676
2677
2678
2679
                    else:
                        output[k] = v

                return output

2680
2681
2682
2683
2684
2685
            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))
2686
2687
2688
2689
2690
2691
2692
2693

            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))

2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
    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
2708
    @require_accelerate
2709
    @mark.accelerate_tests
Sylvain Gugger's avatar
Sylvain Gugger committed
2710
    @require_torch_gpu
2711
    def test_disk_offload_bin(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
2712
2713
2714
2715
2716
2717
        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

2718
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2719
2720
            model = model_class(config).eval()
            model = model.to(torch_device)
2721
            torch.manual_seed(0)
2722
            base_output = model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2723
2724
2725

            model_size = compute_module_sizes(model)[""]
            with tempfile.TemporaryDirectory() as tmp_dir:
2726
                model.cpu().save_pretrained(tmp_dir, safe_serialization=False)
Sylvain Gugger's avatar
Sylvain Gugger committed
2727
2728

                with self.assertRaises(ValueError):
Yih-Dar's avatar
Yih-Dar committed
2729
2730
                    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
2731
2732
2733
                    # 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
2734
2735
                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
2736
2737
2738
2739
2740
                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)
2741
                torch.manual_seed(0)
2742
                new_output = new_model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2743

2744
                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
Sylvain Gugger's avatar
Sylvain Gugger committed
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
    @require_accelerate
    @mark.accelerate_tests
    @require_torch_gpu
    def test_disk_offload_safetensors(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

            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config).eval()
            model = model.to(torch_device)
            torch.manual_seed(0)
            base_output = model(**inputs_dict_class)

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

                max_size = int(self.model_split_percents[1] * model_size)
                max_memory = {0: max_size, "cpu": max_size}

                # This doesn't error out as it's in safetensors and doesn't need an offload folder
                new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)

                self.check_device_map_is_respected(new_model, new_model.hf_device_map)
                torch.manual_seed(0)
                new_output = new_model(**inputs_dict_class)

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

2778
    @require_accelerate
2779
    @mark.accelerate_tests
2780
2781
2782
2783
2784
2785
2786
2787
    @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

2788
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2789
2790
            model = model_class(config).eval()
            model = model.to(torch_device)
2791
2792

            torch.manual_seed(0)
2793
            base_output = model(**inputs_dict_class)
2794
2795
2796

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
Yih-Dar's avatar
Yih-Dar committed
2797
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
            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)
2808
2809

                    torch.manual_seed(0)
2810
                    new_output = new_model(**inputs_dict_class)
2811

2812
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2813
2814

    @require_accelerate
2815
    @mark.accelerate_tests
2816
2817
2818
2819
2820
2821
2822
2823
    @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

2824
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2825
2826
            model = model_class(config).eval()
            model = model.to(torch_device)
2827
2828

            torch.manual_seed(0)
2829
            base_output = model(**inputs_dict_class)
2830
2831
2832

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
2833
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
            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)
2844
2845

                    torch.manual_seed(0)
2846
                    new_output = new_model(**inputs_dict_class)
2847

2848
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2849

2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
    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:
2860
2861
2862
            if model_class.__name__ not in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES),
2863
            ]:
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
                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"])

2882
2883
2884
2885
2886
2887
                    # 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
2888
2889
2890
2891
2892
                    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}"
                            )
2893

2894
2895
                    loss.backward()

2896
    def test_load_with_mismatched_shapes(self):
2897
2898
        if not self.test_mismatched_shapes:
            return
2899
2900
2901
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
2902
            if model_class.__name__ not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES):
2903
2904
2905
2906
2907
2908
2909
2910
                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
2911
                    with self.assertRaises(RuntimeError):
2912
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
2913
2914
                    with self.assertRaises(RuntimeError):
                        new_model_without_prefix = AutoModel.from_pretrained(tmp_dir, vocab_size=10)
2915
2916

                    logger = logging.get_logger("transformers.modeling_utils")
2917

2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
                    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)

2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
                    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)

2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
    def test_mismatched_shapes_have_properly_initialized_weights(self):
        if not self.test_mismatched_shapes:
            return
        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:
            if model_class.__name__ not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES):
                continue

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

                    # Fails when we don't set ignore_mismatched_sizes=True
                    with self.assertRaises(RuntimeError):
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)

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

                    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)

                    for name, param in new_model.named_parameters():
                        if param.requires_grad:
                            self.assertIn(
                                ((param.data.mean() * 1e9).round() / 1e9).item(),
                                [0.0, 1.0],
                                msg=f"Parameter {name} of model {model_class} seems not properly initialized",
                            )

    def test_matched_shapes_have_loaded_weights_when_some_mismatched_shapes_exist(self):
        # 1. Create a dummy class. Should have buffers as well? To make sure we test __init__
        class MyClass(PreTrainedModel):
            config_class = PretrainedConfig

            def __init__(self, config=None):
                super().__init__(config if config is not None else PretrainedConfig())
                self.linear = nn.Linear(10, config.num_labels, bias=True)
                self.embedding = nn.Embedding(10, 10)
                self.std = 1

            def _init_weights(self, module):
                if isinstance(module, nn.Linear):
                    module.weight.data = nn.init.kaiming_uniform_(module.weight.data, np.sqrt(5))
                    if module.bias is not None:
                        module.bias.data = module.bias.data.normal_(mean=0.0, std=self.std)

        # Used to make sure the weights with matched shape are loaded correctly
        config = PretrainedConfig()
        config.num_labels = 3
        model = MyClass(config=config)

        # Used to make sure the weights with mismatched shape are properly initialized
        set_seed(0)
        config = PretrainedConfig()
        config.num_labels = 4
        # not to init. the weights during the creation: to match the logic in `from_pretrained`, so we can keep the
        # same sequence of random ops in the execution path to allow us to compare `target_model` and `new_model` below
        # for `linear` part.
        with ContextManagers([no_init_weights(True)]):
            target_model = MyClass(config=config)
        target_model.apply(target_model._initialize_weights)

        with tempfile.TemporaryDirectory() as tmpdirname:
            state_dict = model.state_dict()
            del state_dict["linear.weight"]

            model.config.save_pretrained(tmpdirname)
            torch.save(state_dict, os.path.join(tmpdirname, "pytorch_model.bin"))

            set_seed(0)
            new_model = MyClass.from_pretrained(tmpdirname, num_labels=4, ignore_mismatched_sizes=True)

            for key in new_model.state_dict().keys():
                # check weight values for weights with matched shapes are identical
                # (i.e. correctly loaded from the checkpoint)
                if key not in ["linear.weight", "linear.bias"]:
                    max_diff = torch.max(torch.abs(model.state_dict()[key] - new_model.state_dict()[key]))
                    self.assertLessEqual(
                        max_diff.item(),
                        1e-6,
                        msg=f"the weight values for `{key}` in `new_model` and `model` are  not identical",
                    )
                else:
                    # check we have some mismatched shapes
                    self.assertNotEqual(
                        model.state_dict()[key].shape,
                        new_model.state_dict()[key].shape,
                        msg=f"the weight shapes for {key} in `model` and `new_model` should differ",
                    )
                    # check the weights with mismatched shape are properly initialized
                    max_diff = torch.max(torch.abs(new_model.state_dict()[key] - target_model.state_dict()[key]))
                    self.assertLessEqual(
                        max_diff.item(),
                        1e-6,
                        msg=f"the weight values for `{key}` in `new_model` and `target_model` are not identical",
                    )

3044
3045
3046
3047
3048
3049
3050
3051
3052
    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
3053
            ), f"{model_class} is too big for the common tests ({num_params})! It should have 1M max."
3054

3055
3056
3057
3058
3059
3060
3061
3062
3063
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_conversion(self):
        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:
3064
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3065
3066
3067
3068
3069
3070

            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model = model_class.from_pretrained(
3071
                    tmpdirname, torch_dtype=torch.float16, attn_implementation="flash_attention_2"
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
                ).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):
        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
3087
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3088

3089
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3090
3091
3092
3093
3094
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model_fa = model_class.from_pretrained(
3095
                    tmpdirname, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2"
3096
3097
3098
3099
                )
                model_fa.to(torch_device)

                model = model_class.from_pretrained(
3100
                    tmpdirname, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2"
3101
3102
3103
                )
                model.to(torch_device)

3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
                dummy_input = inputs_dict[model.main_input_name][:1]
                if dummy_input.dtype in [torch.float32, torch.float16]:
                    dummy_input = dummy_input.to(torch.bfloat16)

                dummy_attention_mask = inputs_dict.get("attention_mask", None)

                if dummy_attention_mask is not None:
                    dummy_attention_mask = dummy_attention_mask[:1]
                    dummy_attention_mask[:, 1:] = 1
                    dummy_attention_mask[:, :1] = 0
3114

3115
3116
3117
3118
3119
3120
3121
3122
                if model.config.is_encoder_decoder:
                    decoder_input_ids = inputs_dict.get("decoder_input_ids", dummy_input)[:1]

                    outputs = model(dummy_input, decoder_input_ids=decoder_input_ids, output_hidden_states=True)
                    outputs_fa = model_fa(dummy_input, decoder_input_ids=decoder_input_ids, output_hidden_states=True)
                else:
                    outputs = model(dummy_input, output_hidden_states=True)
                    outputs_fa = model_fa(dummy_input, output_hidden_states=True)
3123

3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
                logits = (
                    outputs.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs.decoder_hidden_states[-1]
                )
                logits_fa = (
                    outputs_fa.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs_fa.decoder_hidden_states[-1]
                )
3134

3135
                assert torch.allclose(logits_fa, logits, atol=4e-2, rtol=4e-2)
3136

3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
                if model.config.is_encoder_decoder:
                    other_inputs = {
                        "decoder_input_ids": decoder_input_ids,
                        "decoder_attention_mask": dummy_attention_mask,
                        "output_hidden_states": True,
                    }
                    if dummy_attention_mask is not None:
                        other_inputs["attention_mask"] = dummy_attention_mask

                    outputs = model(dummy_input, **other_inputs)
                    outputs_fa = model_fa(dummy_input, **other_inputs)
                else:
                    other_inputs = {
                        "output_hidden_states": True,
                    }
                    if dummy_attention_mask is not None:
                        other_inputs["attention_mask"] = dummy_attention_mask

                    outputs = model(dummy_input, **other_inputs)
                    outputs_fa = model_fa(dummy_input, **other_inputs)

                logits = (
                    outputs.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs.decoder_hidden_states[-1]
                )
                logits_fa = (
                    outputs_fa.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs_fa.decoder_hidden_states[-1]
                )
3168

3169
                assert torch.allclose(logits_fa[1:], logits[1:], atol=4e-2, rtol=4e-2)
3170

3171
3172
                # check with inference + dropout
                model.train()
3173
                _ = model_fa(dummy_input, **other_inputs)
3174

3175
3176
3177
3178
3179
3180
3181
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_inference_padding_right(self):
        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
3182
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3183

3184
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3185
3186
3187
3188
3189
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model_fa = model_class.from_pretrained(
3190
                    tmpdirname, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2"
3191
3192
3193
3194
                )
                model_fa.to(torch_device)

                model = model_class.from_pretrained(
3195
                    tmpdirname, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2"
3196
3197
3198
                )
                model.to(torch_device)

3199
3200
3201
3202
3203
                dummy_input = inputs_dict[model.main_input_name][:1]
                if dummy_input.dtype in [torch.float32, torch.float16]:
                    dummy_input = dummy_input.to(torch.bfloat16)

                dummy_attention_mask = inputs_dict.get("attention_mask", None)
3204

3205
3206
3207
3208
                if dummy_attention_mask is not None:
                    dummy_attention_mask = dummy_attention_mask[:1]
                    dummy_attention_mask[:, :-1] = 1
                    dummy_attention_mask[:, -1:] = 0
3209

3210
3211
                if model.config.is_encoder_decoder:
                    decoder_input_ids = inputs_dict.get("decoder_input_ids", dummy_input)[:1]
3212

3213
3214
3215
3216
3217
                    outputs = model(dummy_input, decoder_input_ids=decoder_input_ids, output_hidden_states=True)
                    outputs_fa = model_fa(dummy_input, decoder_input_ids=decoder_input_ids, output_hidden_states=True)
                else:
                    outputs = model(dummy_input, output_hidden_states=True)
                    outputs_fa = model_fa(dummy_input, output_hidden_states=True)
3218

3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
                logits = (
                    outputs.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs.decoder_hidden_states[-1]
                )
                logits_fa = (
                    outputs_fa.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs_fa.decoder_hidden_states[-1]
                )
3229

3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
                assert torch.allclose(logits_fa, logits, atol=4e-2, rtol=4e-2)

                if model.config.is_encoder_decoder:
                    other_inputs = {
                        "decoder_input_ids": decoder_input_ids,
                        "decoder_attention_mask": dummy_attention_mask,
                        "output_hidden_states": True,
                    }
                    if dummy_attention_mask is not None:
                        other_inputs["attention_mask"] = dummy_attention_mask

                    outputs = model(dummy_input, **other_inputs)
                    outputs_fa = model_fa(dummy_input, **other_inputs)
                else:
                    other_inputs = {
                        "output_hidden_states": True,
                    }
                    if dummy_attention_mask is not None:
                        other_inputs["attention_mask"] = dummy_attention_mask

                    outputs = model(dummy_input, **other_inputs)
                    outputs_fa = model_fa(dummy_input, **other_inputs)

                logits = (
                    outputs.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs.decoder_hidden_states[-1]
                )
                logits_fa = (
                    outputs_fa.hidden_states[-1]
                    if not model.config.is_encoder_decoder
                    else outputs_fa.decoder_hidden_states[-1]
                )

                assert torch.allclose(logits_fa[:-1], logits[:-1], atol=4e-2, rtol=4e-2)
3265
3266
3267
3268
3269
3270
3271
3272

    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_generate_left_padding(self):
        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3273
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3274

3275
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3276
3277
3278
3279
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
3280
3281
3282
                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, low_cpu_mem_usage=True).to(
                    torch_device
                )
3283

3284
3285
3286
3287
3288
3289
3290
3291
                dummy_input = inputs_dict[model.main_input_name]
                if dummy_input.dtype in [torch.float32, torch.bfloat16]:
                    dummy_input = dummy_input.to(torch.float16)

                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
                # make sure we do left padding
                dummy_attention_mask[:, :-1] = 0
                dummy_attention_mask[:, -1:] = 1
3292
3293
3294
3295
3296
3297

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

                model = model_class.from_pretrained(
3298
3299
3300
3301
                    tmpdirname,
                    torch_dtype=torch.float16,
                    attn_implementation="flash_attention_2",
                    low_cpu_mem_usage=True,
3302
3303
3304
3305
3306
3307
                ).to(torch_device)

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

3308
                self.assertTrue(torch.allclose(out, out_fa))
3309
3310
3311
3312
3313
3314
3315
3316

    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_generate_padding_right(self):
        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3317
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3318

3319
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3320
3321
3322
3323
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
3324
3325
3326
                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, low_cpu_mem_usage=True).to(
                    torch_device
                )
3327

3328
3329
3330
3331
3332
                dummy_input = inputs_dict[model.main_input_name]
                if dummy_input.dtype in [torch.float32, torch.bfloat16]:
                    dummy_input = dummy_input.to(torch.float16)

                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
3333
                # make sure we do right padding
3334
3335
                dummy_attention_mask[:, :-1] = 1
                dummy_attention_mask[:, -1:] = 0
3336
3337
3338
3339
3340
3341

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

                model = model_class.from_pretrained(
3342
3343
3344
3345
                    tmpdirname,
                    torch_dtype=torch.float16,
                    attn_implementation="flash_attention_2",
                    low_cpu_mem_usage=True,
3346
3347
3348
3349
3350
3351
                ).to(torch_device)

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

3352
                self.assertTrue(torch.allclose(out, out_fa))
3353

3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
    @require_torch_sdpa
    @slow
    def test_eager_matches_sdpa_inference(self, torch_dtype: str):
        if not self.all_model_classes[0]._supports_sdpa:
            self.skipTest(f"{self.all_model_classes[0].__name__} does not support SDPA")

        if torch_device == "cpu" and torch_dtype == "float16":
            self.skipTest("float16 not supported on cpu")

        # Not sure whether it's fine to put torch.XXX in a decorator if torch is not available so hacking it here instead.
        if torch_dtype == "float16":
            torch_dtype = torch.float16
        elif torch_dtype == "bfloat16":
            torch_dtype = torch.bfloat16
        elif torch_dtype == "float32":
            torch_dtype = torch.float32

        atols = {
            ("cpu", False, torch.float32): 1e-6,
            ("cpu", False, torch.bfloat16): 1e-2,
            ("cpu", True, torch.float32): 1e-6,
            ("cpu", True, torch.bfloat16): 1e-2,
            ("cuda", False, torch.float32): 1e-6,
            ("cuda", False, torch.bfloat16): 1e-2,
            ("cuda", False, torch.float16): 1e-3,
            ("cuda", True, torch.float32): 1e-6,
            ("cuda", True, torch.bfloat16): 1e-2,
            ("cuda", True, torch.float16): 5e-3,
        }
        rtols = {
            ("cpu", False, torch.float32): 1e-4,
            ("cpu", False, torch.bfloat16): 1e-2,
            ("cpu", True, torch.float32): 1e-4,
            ("cpu", True, torch.bfloat16): 1e-2,
            ("cuda", False, torch.float32): 1e-4,
            ("cuda", False, torch.bfloat16): 1e-2,
            ("cuda", False, torch.float16): 1e-3,
            ("cuda", True, torch.float32): 1e-4,
            ("cuda", True, torch.bfloat16): 3e-2,
            ("cuda", True, torch.float16): 5e-3,
        }

        def get_mean_reldiff(failcase, x, ref, atol, rtol):
            return f"{failcase}: mean relative difference: {((x - ref).abs() / (ref.abs() + 1e-12)).mean():.3e}, torch atol = {atol}, torch rtol = {rtol}"

        for model_class in self.all_model_classes:
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            model = model_class(config)

            is_encoder_decoder = model.config.is_encoder_decoder

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model_sdpa = model_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype)
                model_sdpa = model_sdpa.eval().to(torch_device)

                self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")

                model_eager = model_class.from_pretrained(
                    tmpdirname,
                    torch_dtype=torch_dtype,
                    attn_implementation="eager",
                )
                model_eager = model_eager.eval().to(torch_device)

                self.assertTrue(model_eager.config._attn_implementation == "eager")

                for name, submodule in model_eager.named_modules():
                    if "SdpaAttention" in submodule.__class__.__name__:
                        raise ValueError("The eager model should not have SDPA attention layers")

                has_sdpa = False
                for name, submodule in model_sdpa.named_modules():
                    if "SdpaAttention" in submodule.__class__.__name__:
                        has_sdpa = True
                        break
                if not has_sdpa and model_sdpa.config.model_type != "falcon":
                    raise ValueError("The SDPA model should have SDPA attention layers")

                # We use these for loops instead of parameterized.expand just for the interest of avoiding loading/saving 8 times the model,
                # but it would be nicer to have an efficient way to use parameterized.expand
                fail_cases = []
                for padding_side in ["left", "right"]:
                    for use_mask in [False, True]:
                        for batch_size in [1, 5]:
                            dummy_input = inputs_dict[model.main_input_name]

                            if dummy_input.dtype in [torch.float32, torch.bfloat16, torch.float16]:
                                dummy_input = dummy_input.to(torch_dtype)

                            dummy_input = dummy_input[:batch_size]
                            if dummy_input.shape[0] != batch_size:
                                if dummy_input.dtype in [torch.float32, torch.bfloat16, torch.float16]:
                                    extension = torch.rand(
                                        batch_size - dummy_input.shape[0],
                                        *dummy_input.shape[1:],
                                        dtype=torch_dtype,
                                        device=torch_device,
                                    )
                                    dummy_input = torch.cat((dummy_input, extension), dim=0).to(torch_device)
                                else:
                                    extension = torch.randint(
                                        high=5,
                                        size=(batch_size - dummy_input.shape[0], *dummy_input.shape[1:]),
                                        dtype=dummy_input.dtype,
                                        device=torch_device,
                                    )
                                    dummy_input = torch.cat((dummy_input, extension), dim=0).to(torch_device)

                            if not use_mask:
                                dummy_attention_mask = None
                            else:
                                dummy_attention_mask = inputs_dict.get("attention_mask", None)
                                if dummy_attention_mask is None:
                                    if is_encoder_decoder:
                                        seqlen = inputs_dict.get("decoder_input_ids", dummy_input).shape[-1]
                                    else:
                                        seqlen = dummy_input.shape[-1]
                                    dummy_attention_mask = (
                                        torch.ones(batch_size, seqlen).to(torch.int64).to(torch_device)
                                    )

                                dummy_attention_mask = dummy_attention_mask[:batch_size]
                                if dummy_attention_mask.shape[0] != batch_size:
                                    extension = torch.ones(
                                        batch_size - dummy_attention_mask.shape[0],
                                        *dummy_attention_mask.shape[1:],
                                        dtype=dummy_attention_mask.dtype,
                                        device=torch_device,
                                    )
                                    dummy_attention_mask = torch.cat((dummy_attention_mask, extension), dim=0)
                                    dummy_attention_mask = dummy_attention_mask.to(torch_device)

                                dummy_attention_mask[:] = 1
                                if padding_side == "left":
                                    dummy_attention_mask[-1, :-1] = 1
                                    dummy_attention_mask[-1, -4:] = 0
                                elif padding_side == "right":
                                    dummy_attention_mask[-1, 1:] = 1
                                    dummy_attention_mask[-1, :3] = 0

                            for enable_kernels in [False, True]:
                                failcase = f"padding_side={padding_side}, use_mask={use_mask}, batch_size={batch_size}, enable_kernels={enable_kernels}"
                                if is_encoder_decoder:
                                    decoder_input_ids = inputs_dict.get("decoder_input_ids", dummy_input)[:batch_size]
                                    if decoder_input_ids.shape[0] != batch_size:
                                        extension = torch.ones(
                                            batch_size - decoder_input_ids.shape[0],
                                            *decoder_input_ids.shape[1:],
                                            dtype=decoder_input_ids.dtype,
                                            device=torch_device,
                                        )
                                        decoder_input_ids = torch.cat((decoder_input_ids, extension), dim=0)
                                        decoder_input_ids = decoder_input_ids.to(torch_device)

                                    # TODO: never an `attention_mask` arg here?
                                    other_inputs = {
                                        "decoder_input_ids": decoder_input_ids,
                                        "decoder_attention_mask": dummy_attention_mask,
                                        "output_hidden_states": True,
                                    }
                                else:
                                    other_inputs = {
                                        "output_hidden_states": True,
                                    }

                                    # Otherwise fails for e.g. WhisperEncoderModel
                                    if "attention_mask" in inspect.signature(model_eager.forward).parameters:
                                        other_inputs["attention_mask"] = dummy_attention_mask

                                # TODO: test gradients as well (& for FA2 as well!)
                                with torch.no_grad():
                                    with torch.backends.cuda.sdp_kernel(
                                        enable_flash=enable_kernels,
                                        enable_math=True,
                                        enable_mem_efficient=enable_kernels,
                                    ):
                                        outputs_eager = model_eager(dummy_input, **other_inputs)
                                        outputs_sdpa = model_sdpa(dummy_input, **other_inputs)

                                logits_eager = (
                                    outputs_eager.hidden_states[-1]
                                    if not is_encoder_decoder
                                    else outputs_eager.decoder_hidden_states[-1]
                                )
                                logits_sdpa = (
                                    outputs_sdpa.hidden_states[-1]
                                    if not is_encoder_decoder
                                    else outputs_sdpa.decoder_hidden_states[-1]
                                )

                                if torch_device in ["cpu", "cuda"]:
                                    atol = atols[torch_device, enable_kernels, torch_dtype]
                                    rtol = rtols[torch_device, enable_kernels, torch_dtype]
                                else:
                                    atol = 1e-7
                                    rtol = 1e-4

                                # Masked tokens output slightly deviates - we don't mind that.
                                if use_mask:
                                    if padding_side == "left":
                                        sub_sdpa = logits_sdpa[:-1]
                                        sub_eager = logits_eager[:-1]
                                        if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
                                            fail_cases.append(
                                                get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
                                            )

                                        sub_sdpa = logits_sdpa[-1, :-4]
                                        sub_eager = logits_eager[-1, :-4]
                                        if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
                                            fail_cases.append(
                                                get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
                                            )

                                        # Testing the padding tokens is not really meaningful but anyway
                                        # sub_sdpa = logits_sdpa[-1, -4:]
                                        # sub_eager = logits_eager[-1, -4:]
                                        # if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
                                        #     fail_cases.append(get_mean_reldiff(failcase, sub_sdpa, sub_eager, 4e-2, 4e-2))
                                    elif padding_side == "right":
                                        sub_sdpa = logits_sdpa[:-1]
                                        sub_eager = logits_eager[:-1]
                                        if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
                                            fail_cases.append(
                                                get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
                                            )

                                        sub_sdpa = logits_sdpa[-1, 3:]
                                        sub_eager = logits_eager[-1, 3:]
                                        if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
                                            fail_cases.append(
                                                get_mean_reldiff(failcase, sub_sdpa, sub_eager, atol, rtol)
                                            )

                                        # Testing the padding tokens is not really meaningful but anyway
                                        # sub_sdpa = logits_sdpa[-1, :3]
                                        # sub_eager = logits_eager[-1, :3]
                                        # if not torch.allclose(sub_sdpa, sub_eager, atol=atol, rtol=rtol):
                                        #     fail_cases.append(get_mean_reldiff(failcase, sub_sdpa, sub_eager, 4e-2, 4e-2))

                                else:
                                    if not torch.allclose(logits_sdpa, logits_eager, atol=atol, rtol=rtol):
                                        fail_cases.append(
                                            get_mean_reldiff(failcase, logits_sdpa, logits_eager, atol, rtol)
                                        )

                self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))

    @require_torch_sdpa
    @slow
    def test_eager_matches_sdpa_generate(self):
        max_new_tokens = 30

        if len(self.all_generative_model_classes) == 0:
            self.skipTest(f"{self.__class__.__name__} tests a model that does support generate: skipping this test")

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_sdpa:
                self.skipTest(f"{model_class.__name__} does not support SDPA")

            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

            dummy_input = inputs_dict[model_class.main_input_name]
            if dummy_input.dtype in [torch.float32, torch.bfloat16]:
                dummy_input = dummy_input.to(torch.float16)

            # make sure that all models have enough positions for generation
            if hasattr(config, "max_position_embeddings"):
                config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1

            model = model_class(config)

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

                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))

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

                self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")

                model_eager = model_class.from_pretrained(
                    tmpdirname,
                    torch_dtype=torch.float16,
                    low_cpu_mem_usage=True,
                    attn_implementation="eager",
                ).to(torch_device)

                self.assertTrue(model_eager.config._attn_implementation == "eager")

                for name, submodule in model_eager.named_modules():
                    if "SdpaAttention" in submodule.__class__.__name__:
                        raise ValueError("The eager model should not have SDPA attention layers")

                has_sdpa = False
                for name, submodule in model_sdpa.named_modules():
                    if "SdpaAttention" in submodule.__class__.__name__:
                        has_sdpa = True
                        break
                if not has_sdpa:
                    raise ValueError("The SDPA model should have SDPA attention layers")

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

                res_sdpa = model_sdpa.generate(
                    dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
                )

                self.assertTrue(torch.allclose(res_eager, res_sdpa))

3673
3674
3675
3676
3677
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_generate_use_cache(self):
3678
3679
        max_new_tokens = 30

3680
3681
        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3682
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3683

3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

            dummy_input = inputs_dict[model_class.main_input_name]
            if dummy_input.dtype in [torch.float32, torch.bfloat16]:
                dummy_input = dummy_input.to(torch.float16)

            # make sure that all models have enough positions for generation
            if hasattr(config, "max_position_embeddings"):
                config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1

3694
3695
3696
3697
3698
            model = model_class(config)

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

3699
                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
3700
3701

                model = model_class.from_pretrained(
3702
3703
                    tmpdirname,
                    torch_dtype=torch.float16,
3704
                    attn_implementation="flash_attention_2",
3705
                    low_cpu_mem_usage=True,
3706
3707
3708
3709
                ).to(torch_device)

                # Just test that a large cache works as expected
                _ = model.generate(
3710
3711
3712
3713
3714
                    dummy_input,
                    attention_mask=dummy_attention_mask,
                    max_new_tokens=max_new_tokens,
                    do_sample=False,
                    use_cache=True,
3715
3716
                )

3717
3718
3719
3720
3721
3722
3723
3724
    @require_flash_attn
    @require_torch_gpu
    @require_bitsandbytes
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_fp32_ln(self):
        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3725
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3726

3727
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3728
3729
3730
3731
3732
            model = model_class(config)

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

3733
3734
3735
3736
3737
3738
                dummy_input = inputs_dict[model.main_input_name]
                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))

                if model.config.is_encoder_decoder:
                    dummy_decoder_input_ids = inputs_dict["decoder_input_ids"]
                    dummy_decoder_attention_mask = inputs_dict["decoder_attention_mask"]
3739
3740
3741
3742

                model = model_class.from_pretrained(
                    tmpdirname,
                    torch_dtype=torch.float16,
3743
                    attn_implementation="flash_attention_2",
3744
3745
3746
3747
3748
3749
3750
3751
3752
                    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)

3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
                if model.config.is_encoder_decoder:
                    _ = model(dummy_input, decoder_input_ids=dummy_decoder_input_ids)
                    # with attention mask
                    _ = model(
                        dummy_input,
                        attention_mask=dummy_attention_mask,
                        decoder_input_ids=dummy_decoder_input_ids,
                        decoder_attention_mask=dummy_decoder_attention_mask,
                    )
                else:
                    _ = model(dummy_input)
                    # with attention mask
                    _ = model(dummy_input, attention_mask=dummy_attention_mask)
3766

3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
    @is_pt_tf_cross_test
    def test_tf_from_pt_safetensors(self):
        for model_class in self.all_model_classes:
            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
            if not hasattr(transformers, tf_model_class_name):
                # transformers does not have this model in TF version yet
                return

            tf_model_class = getattr(transformers, tf_model_class_name)

            pt_model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                pt_model.save_pretrained(tmpdirname, safe_serialization=True)
                tf_model_1 = tf_model_class.from_pretrained(tmpdirname, from_pt=True)

                pt_model.save_pretrained(tmpdirname, safe_serialization=False)
                tf_model_2 = tf_model_class.from_pretrained(tmpdirname, from_pt=True)

                # Check models are equal
                for p1, p2 in zip(tf_model_1.weights, tf_model_2.weights):
                    self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))

    @is_pt_flax_cross_test
    def test_flax_from_pt_safetensors(self):
        for model_class in self.all_model_classes:
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

            flax_model_class_name = "Flax" + model_class.__name__  # Add the "Flax at the beginning
            if not hasattr(transformers, flax_model_class_name):
                # transformers does not have this model in Flax version yet
                return

            flax_model_class = getattr(transformers, flax_model_class_name)

            pt_model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                pt_model.save_pretrained(tmpdirname, safe_serialization=True)
                flax_model_1 = flax_model_class.from_pretrained(tmpdirname, from_pt=True)

                pt_model.save_pretrained(tmpdirname, safe_serialization=False)
                flax_model_2 = flax_model_class.from_pretrained(tmpdirname, from_pt=True)

                # Check models are equal
                self.assertTrue(check_models_equal(flax_model_1, flax_model_2))

3816
3817
3818
3819
3820
3821
3822
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_from_config(self):
        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3823
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3824
3825
3826
3827

            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
            # TODO: to change it in the future with other relevant auto classes
            fa2_model = AutoModelForCausalLM.from_config(
3828
                config, attn_implementation="flash_attention_2", torch_dtype=torch.bfloat16
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
            ).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)

            fa2_correctly_converted = False

            for _, module in fa2_model.named_modules():
                if "FlashAttention" in module.__class__.__name__:
                    fa2_correctly_converted = True
                    break

            self.assertTrue(fa2_correctly_converted)

            _ = fa2_model(input_ids=dummy_input, attention_mask=dummy_attention_mask)

            with tempfile.TemporaryDirectory() as tmpdirname:
                fa2_model.save_pretrained(tmpdirname)

                model_from_pretrained = AutoModelForCausalLM.from_pretrained(tmpdirname)

3850
                self.assertTrue(model_from_pretrained.config._attn_implementation != "flash_attention_2")
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860

                fa2_correctly_converted = False

                for _, module in model_from_pretrained.named_modules():
                    if "FlashAttention" in module.__class__.__name__:
                        fa2_correctly_converted = True
                        break

                self.assertFalse(fa2_correctly_converted)

3861

3862
global_rng = random.Random()
thomwolf's avatar
thomwolf committed
3863
3864


thomwolf's avatar
thomwolf committed
3865
def ids_tensor(shape, vocab_size, rng=None, name=None):
3866
    #  Creates a random int32 tensor of the shape within the vocab size
thomwolf's avatar
thomwolf committed
3867
    if rng is None:
3868
        rng = global_rng
thomwolf's avatar
thomwolf committed
3869

thomwolf's avatar
thomwolf committed
3870
3871
3872
    total_dims = 1
    for dim in shape:
        total_dims *= dim
thomwolf's avatar
thomwolf committed
3873

thomwolf's avatar
thomwolf committed
3874
3875
3876
    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))
thomwolf's avatar
thomwolf committed
3877

3878
    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
thomwolf's avatar
thomwolf committed
3879
3880


3881
3882
3883
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
3884
3885
    # 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
3886
3887
3888
    return attn_mask


3889
def floats_tensor(shape, scale=1.0, rng=None, name=None):
Patrick von Platen's avatar
Patrick von Platen committed
3890
    """Creates a random float32 tensor"""
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
    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)

3902
    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()