test_modeling_common.py 150 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
15

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

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

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

Aymeric Augustin's avatar
Aymeric Augustin committed
88

89
90
91
92
if is_accelerate_available():
    from accelerate.utils import compute_module_sizes


93
if is_torch_available():
94
    import torch
95
    from safetensors.torch import load_file as safe_load_file
96
    from safetensors.torch import save_file as safe_save_file
97
    from torch import nn
thomwolf's avatar
thomwolf committed
98

99
    from transformers import MODEL_MAPPING, AdaptiveEmbedding
Sylvain Gugger's avatar
Sylvain Gugger committed
100
    from transformers.pytorch_utils import id_tensor_storage
thomwolf's avatar
thomwolf committed
101

Sylvain Gugger's avatar
Sylvain Gugger committed
102

103
104
105
if is_tf_available():
    import tensorflow as tf

106
107
if is_flax_available():
    import jax.numpy as jnp
108

109
    from tests.test_modeling_flax_utils import check_models_equal
110
111
112
113
114
    from transformers.modeling_flax_pytorch_utils import (
        convert_pytorch_state_dict_to_flax,
        load_flax_weights_in_pytorch_model,
    )

115
if is_torch_fx_available():
116
    from transformers.utils.fx import symbolic_trace
117

118

119
120
121
def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
122
        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
123
            setattr(configs_no_init, key, 1e-10)
124
125
126
        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)
127
128
    return configs_no_init

thomwolf's avatar
thomwolf committed
129

130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
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()


155
156
157
158
@require_torch
class ModelTesterMixin:
    model_tester = None
    all_model_classes = ()
159
    all_generative_model_classes = ()
160
    fx_compatible = False
Patrick von Platen's avatar
Patrick von Platen committed
161
162
163
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
164
    test_resize_position_embeddings = False
Patrick von Platen's avatar
Patrick von Platen committed
165
    test_head_masking = True
166
    test_mismatched_shapes = True
167
    test_missing_keys = True
168
    test_model_parallel = False
169
    is_encoder_decoder = False
170
    has_attentions = True
171
    model_split_percents = [0.5, 0.7, 0.9]
172

173
174
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
175
        if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
176
            inputs_dict = {
177
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
178
                if isinstance(v, torch.Tensor) and v.ndim > 1
Sylvain Gugger's avatar
Sylvain Gugger committed
179
                else v
180
181
                for k, v in inputs_dict.items()
            }
182
        elif model_class.__name__ in get_values(MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES):
183
            inputs_dict.pop("attention_mask")
184
185

        if return_labels:
186
            if model_class.__name__ in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES):
187
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
188
189
190
            elif model_class.__name__ in [
                *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
191
            ]:
192
193
194
195
196
197
                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
                )
198
199
200
201
202
203
            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),
204
            ]:
205
206
207
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
208
209
210
211
212
213
            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),
214
215
216
217
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
218
            elif model_class.__name__ in get_values(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
219
220
221
222
                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
                )
223
            elif model_class.__name__ in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES):
NielsRogge's avatar
NielsRogge committed
224
225
226
227
                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()
228

229
230
        return inputs_dict

Patrick von Platen's avatar
Patrick von Platen committed
231
    def test_save_load(self):
232
233
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

234
235
236
237
238
239
240
241
242
243
        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)

244
245
246
247
248
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
249
                first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
250

251
            with tempfile.TemporaryDirectory() as tmpdirname:
252
                model.save_pretrained(tmpdirname)
253
254
255
256
257
258
259

                # 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))
                )

260
                model = model_class.from_pretrained(tmpdirname)
261
                model.to(torch_device)
262
                with torch.no_grad():
263
                    second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
thomwolf's avatar
thomwolf committed
264

265
266
267
268
269
            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)
270

271
272
273
274
275
276
277
278
279
280
281
282
    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))

283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
    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)

301
    def test_save_load_keys_to_ignore_on_save(self):
302
303
304
305
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
306
307
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
308
309
310
                continue

            # check the keys are in the original state_dict
311
            for k in _keys_to_ignore_on_save:
312
                self.assertIn(k, model.state_dict().keys(), "\n".join(model.state_dict().keys()))
313
314
315
316

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
317
318
319
                output_model_file = os.path.join(tmpdirname, SAFE_WEIGHTS_NAME)
                state_dict_saved = safe_load_file(output_model_file)

320
                for k in _keys_to_ignore_on_save:
321
                    self.assertNotIn(k, state_dict_saved.keys(), "\n".join(state_dict_saved.keys()))
322

Sylvain Gugger's avatar
Sylvain Gugger committed
323
324
                # 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)
325
326
327
328
329
330
                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
331
332
                self.assertTrue(len(load_result.unexpected_keys) == 0)

333
334
335
336
337
338
339
340
341
342
343
    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)

344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
    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)

359
360
361
362
363
364
365
            # 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"
                    )

366
367
368
369
            # check disable works
            model.gradient_checkpointing_disable()
            self.assertFalse(model.is_gradient_checkpointing)

370
371
372
373
374
375
376
            # 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"
                    )

377
378
    def test_save_load_fast_init_from_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
379
380
        if config.__class__ not in MODEL_MAPPING:
            return
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
        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
402
403
            model_class_copy._init_weights = _mock_init_weights
            model_class_copy.init_weights = _mock_all_init_weights
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419

            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)
420
                # Before we test anything
421
422

                for key in model_fast_init.state_dict().keys():
423
424
425
426
427
                    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")
428
429
430

    def test_save_load_fast_init_to_base(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
431
432
        if config.__class__ not in MODEL_MAPPING:
            return
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
        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
454
455
            base_class_copy._init_weights = _mock_init_weights
            base_class_copy.init_weights = _mock_all_init_weights
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473

            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():
474
475
476
477
478
479
480
481
482
                    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")
483

Patrick von Platen's avatar
Patrick von Platen committed
484
    def test_initialization(self):
485
486
487
488
489
490
491
492
        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
493
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
494
                        [0.0, 1.0],
495
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
496
                    )
thomwolf's avatar
thomwolf committed
497

Patrick von Platen's avatar
Patrick von Platen committed
498
    def test_determinism(self):
499
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
500
501
502
503
504
505
506
507
508

        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)

509
510
511
512
513
        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
514
515
                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
516

517
518
519
520
521
            if isinstance(first, tuple) and isinstance(second, tuple):
                for tensor1, tensor2 in zip(first, second):
                    check_determinism(tensor1, tensor2)
            else:
                check_determinism(first, second)
522

523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
    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",
                ]
539
                expected_arg_names.extend(
540
541
                    ["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
542
543
544
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
545
546
547
548
549
550
            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)
551
            else:
552
                expected_arg_names = [model.main_input_name]
553
554
                self.assertListEqual(arg_names[:1], expected_arg_names)

555
    def check_training_gradient_checkpointing(self, gradient_checkpointing_kwargs=None):
556
557
558
559
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
560
561
562
563
564
            if (
                model_class.__name__
                in [*get_values(MODEL_MAPPING_NAMES), *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)]
                or not model_class.supports_gradient_checkpointing
            ):
565
                continue
566

567
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
568
569
            config.use_cache = False
            config.return_dict = True
570
            model = model_class(config)
571

572
            model.to(torch_device)
573
            model.gradient_checkpointing_enable(gradient_checkpointing_kwargs=gradient_checkpointing_kwargs)
574
            model.train()
575
576
577
578
579
580
581

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

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

582
583
584
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()
585
            optimizer.step()
586

587
588
589
590
591
            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):
592
        if not self.model_tester.is_training:
593
594
595
            return

        for model_class in self.all_model_classes:
596
597
598
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.return_dict = True

599
600
601
602
            if model_class.__name__ in [
                *get_values(MODEL_MAPPING_NAMES),
                *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES),
            ]:
603
                continue
604

605
606
607
608
609
610
611
            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()

612
613
614
615
616
617
618
619
620
621
622
623
624
    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})
625

Patrick von Platen's avatar
Patrick von Platen committed
626
    def test_attention_outputs(self):
627
628
629
        if not self.has_attentions:
            self.skipTest(reason="Model does not output attentions")

630
631
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True
632

633
634
635
636
637
638
639
640
641
642
643
644
        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
645
            config.return_dict = True
646
647
648
649
650
651
652
            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)
653

654
655
656
657
658
659
660
661
662
663
            # 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
664

665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
            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
684
685
686
                if model_class.__name__ in [
                    *get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES),
                    *get_values(MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES),
687
                ]:
688
689
690
691
692
693
694
695
696
697
698
699
700
701
                    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],
                )
702

703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
                # 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
746

747
    @slow
748
    def test_torchscript_simple(self):
749
750
        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
751

752
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
753
    def test_torchscript_output_attentions(self):
754
755
756
        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
757

758
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
759
    def test_torchscript_output_hidden_state(self):
760
761
762
        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
763

764
765
766
767
    # 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()
768
769
770
        # 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()
771

772
    def _create_and_check_torchscript(self, config, inputs_dict):
Patrick von Platen's avatar
Patrick von Platen committed
773
        if not self.test_torchscript:
774
            return
775

776
777
778
779
780
781
        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        configs_no_init.torchscript = True
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
782
            inputs = self._prepare_for_class(inputs_dict, model_class)
thomwolf's avatar
thomwolf committed
783

784
785
            main_input_name = model_class.main_input_name

786
            try:
787
                if model.config.is_encoder_decoder:
788
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
789
                    main_input = inputs[main_input_name]
790
791
792
                    attention_mask = inputs["attention_mask"]
                    decoder_input_ids = inputs["decoder_input_ids"]
                    decoder_attention_mask = inputs["decoder_attention_mask"]
793
                    model(main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
794
                    traced_model = torch.jit.trace(
795
                        model, (main_input, attention_mask, decoder_input_ids, decoder_attention_mask)
796
                    )
797
798
799
800
                elif "bbox" in inputs and "image" in inputs:  # LayoutLMv2 requires additional inputs
                    input_ids = inputs["input_ids"]
                    bbox = inputs["bbox"]
                    image = inputs["image"].tensor
801
                    model(input_ids, bbox, image)
802
803
804
                    traced_model = torch.jit.trace(
                        model, (input_ids, bbox, image), check_trace=False
                    )  # when traced model is checked, an error is produced due to name mangling
Jinho Park's avatar
Jinho Park committed
805
806
807
808
809
810
811
                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
812
                else:
813
                    main_input = inputs[main_input_name]
814
                    model(main_input)
815
                    traced_model = torch.jit.trace(model, main_input)
816
817
            except RuntimeError:
                self.fail("Couldn't trace module.")
thomwolf's avatar
thomwolf committed
818

819
            with tempfile.TemporaryDirectory() as tmp_dir_name:
820
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
thomwolf's avatar
thomwolf committed
821

822
                try:
823
                    torch.jit.save(traced_model, pt_file_name)
824
825
                except Exception:
                    self.fail("Couldn't save module.")
thomwolf's avatar
thomwolf committed
826

827
828
829
830
                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")
LysandreJik's avatar
LysandreJik committed
831

832
833
            model.to(torch_device)
            model.eval()
thomwolf's avatar
thomwolf committed
834

835
836
            loaded_model.to(torch_device)
            loaded_model.eval()
thomwolf's avatar
thomwolf committed
837

838
839
840
            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

841
842
843
844
845
846
847
848
849
            non_persistent_buffers = {}
            for key in loaded_model_state_dict.keys():
                if key not in model_state_dict.keys():
                    non_persistent_buffers[key] = loaded_model_state_dict[key]

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

850
            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
thomwolf's avatar
thomwolf committed
851

852
853
854
855
856
857
858
859
860
861
862
            model_buffers = list(model.buffers())
            for non_persistent_buffer in non_persistent_buffers.values():
                found_buffer = False
                for i, model_buffer in enumerate(model_buffers):
                    if torch.equal(non_persistent_buffer, model_buffer):
                        found_buffer = True
                        break

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

863
            models_equal = True
864
            for layer_name, p1 in model_state_dict.items():
865
866
867
868
                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
869

870
            self.assertTrue(models_equal)
thomwolf's avatar
thomwolf committed
871

872
873
874
875
            # 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()

876
877
878
879
880
881
882
883
    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)

884
885
    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:
886
887
888
889
890
            return

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

891
        for model_class in self.all_model_classes:
892
893
894
895
896
897
898
899
900
            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)
901
902
903
                    input_names = [
                        "attention_mask",
                        "decoder_attention_mask",
904
                        "decoder_input_ids",
905
                        "input_features",
906
907
                        "input_ids",
                        "input_values",
908
                    ]
909
910
                    if labels is not None:
                        input_names.append("labels")
911

912
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
913
                    input_names = list(filtered_inputs.keys())
914

915
                    model_output = model(**filtered_inputs)
916

917
                    traced_model = symbolic_trace(model, input_names)
918
                    traced_output = traced_model(**filtered_inputs)
919
                else:
920
921
922
923
                    input_names = [
                        "attention_mask",
                        "bbox",
                        "input_features",
924
925
926
927
928
929
                        "input_ids",
                        "input_values",
                        "pixel_values",
                        "token_type_ids",
                        "visual_feats",
                        "visual_pos",
930
                    ]
931

932
                    labels = inputs.get("labels", None)
933
934
                    start_positions = inputs.get("start_positions", None)
                    end_positions = inputs.get("end_positions", None)
935
936
                    if labels is not None:
                        input_names.append("labels")
937
938
939
940
                    if start_positions is not None:
                        input_names.append("start_positions")
                    if end_positions is not None:
                        input_names.append("end_positions")
941

942
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
943
                    input_names = list(filtered_inputs.keys())
944

945
                    if model.__class__.__name__ in set(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES.values()) and (
946
                        not hasattr(model.config, "problem_type") or model.config.problem_type is None
947
948
949
                    ):
                        model.config.problem_type = "single_label_classification"

950
                    traced_model = symbolic_trace(model, input_names)
951
                    traced_output = traced_model(**filtered_inputs)
952
                    model_output = model(**filtered_inputs)
953

954
            except Exception as e:
955
                self.fail(f"Couldn't trace module: {e}")
956

957
958
959
960
961
962
963
964
965
966
967
968
969
            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)
970
            num_outputs = len(model_output)
971
972
973
974
975
976

            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}",
                )
977

978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
            # 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}",
                    )

998
999
1000
1001
            # 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
1002
1003
    def test_headmasking(self):
        if not self.test_head_masking:
1004
            return
1005

1006
1007
1008
        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()
LysandreJik's avatar
LysandreJik committed
1009

1010
        inputs_dict["output_attentions"] = True
1011
1012
1013
1014
1015
1016
        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
1017

1018
1019
1020
            # 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
1021
1022
1023
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
1024
1025
1026
1027
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
1028
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
1029
            inputs["head_mask"] = head_mask
1030
1031
1032
1033
1034
            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
1035
1036
                if "cross_attn_head_mask" in arg_names:
                    inputs["cross_attn_head_mask"] = head_mask
1037
            outputs = model(**inputs, return_dict=True)
1038
1039
1040
1041
1042
1043
1044
1045
1046

            # 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)
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067

            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)
1068
                check_attentions_validity(outputs.cross_attentions)
1069
1070
            else:
                check_attentions_validity(outputs.attentions)
1071

Patrick von Platen's avatar
Patrick von Platen committed
1072
1073
    def test_head_pruning(self):
        if not self.test_pruning:
1074
1075
1076
            return

        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1077
1078
1079
1080
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1081

1082
1083
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1084

1085
            inputs_dict["output_attentions"] = True
1086
1087
1088
1089
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1090
1091
1092
1093
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1094
1095
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
1096
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1097

1098
            attentions = outputs[-1]
1099

1100
            self.assertEqual(attentions[0].shape[-3], 1)
1101
1102
            # 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)
1103
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
LysandreJik's avatar
LysandreJik committed
1104

Patrick von Platen's avatar
Patrick von Platen committed
1105
1106
    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
1107
            return
LysandreJik's avatar
LysandreJik committed
1108

1109
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1110
1111
1112
1113
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1114
1115
1116

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1117

1118
            inputs_dict["output_attentions"] = True
1119
1120
1121
1122
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1123
1124
1125
1126
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1127
            model.prune_heads(heads_to_prune)
1128

1129
            with tempfile.TemporaryDirectory() as temp_dir_name:
1130
1131
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1132
                model.to(torch_device)
1133

1134
            with torch.no_grad():
1135
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1136
1137
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
1138
1139
            # 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)
1140
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1141

Patrick von Platen's avatar
Patrick von Platen committed
1142
1143
    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
1144
            return
1145

1146
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1147
1148
1149
1150
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1151

1152
1153
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1154

1155
            inputs_dict["output_attentions"] = True
1156
            config.output_hidden_states = False
1157

1158
1159
1160
1161
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
1162
            config.pruned_heads = heads_to_prune
1163

1164
1165
1166
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1167

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

1172
            self.assertEqual(attentions[0].shape[-3], 1)
1173
1174
            # 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)
1175
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
1176

Patrick von Platen's avatar
Patrick von Platen committed
1177
1178
    def test_head_pruning_integration(self):
        if not self.test_pruning:
1179
            return
1180

1181
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
1182
1183
1184
1185
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
1186

1187
1188
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
1189

1190
            inputs_dict["output_attentions"] = True
1191
            config.output_hidden_states = False
1192

1193
            heads_to_prune = {1: [1, 2]}
1194
            config.pruned_heads = heads_to_prune
1195

1196
1197
1198
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
1199

1200
            with torch.no_grad():
1201
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1202
            attentions = outputs[-1]
1203

1204
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1205
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1206

1207
            with tempfile.TemporaryDirectory() as temp_dir_name:
1208
1209
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
1210
                model.to(torch_device)
thomwolf's avatar
thomwolf committed
1211

1212
            with torch.no_grad():
1213
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1214
            attentions = outputs[-1]
LysandreJik's avatar
LysandreJik committed
1215

1216
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 0)
1217
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
thomwolf's avatar
thomwolf committed
1218

1219
            heads_to_prune = {0: [0], 1: [1, 2]}
1220
            model.prune_heads(heads_to_prune)
1221

1222
            with torch.no_grad():
1223
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1224
            attentions = outputs[-1]
1225

1226
1227
            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)
1228

1229
            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2]})
thomwolf's avatar
thomwolf committed
1230

Patrick von Platen's avatar
Patrick von Platen committed
1231
    def test_hidden_states_output(self):
Joseph Liu's avatar
Joseph Liu committed
1232
        def check_hidden_states_output(inputs_dict, config, model_class):
1233
            model = model_class(config)
1234
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
1235
            model.eval()
Joseph Liu's avatar
Joseph Liu committed
1236

thomwolf's avatar
thomwolf committed
1237
            with torch.no_grad():
1238
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
1239
1240

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
1241

Sylvain Gugger's avatar
Sylvain Gugger committed
1242
1243
1244
1245
            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)
1246

Patrick von Platen's avatar
Patrick von Platen committed
1247
1248
1249
1250
1251
1252
1253
            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

1254
            self.assertListEqual(
Lysandre's avatar
Lysandre committed
1255
1256
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
1257
            )
thomwolf's avatar
thomwolf committed
1258

1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
            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
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
        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)

1284
1285
1286
    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
1287
        config.output_attentions = self.has_attentions
1288
1289
1290
1291
1292
1293
1294
1295
1296

        # 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)
1297

1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
        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()

1308
1309
1310
1311
1312
1313
1314
1315
1316
            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()
1317
1318
1319
1320
1321

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

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)
1322
1323
1324
1325
1326

            if self.has_attentions:
                self.assertIsNotNone(encoder_attentions.grad)
                self.assertIsNotNone(decoder_attentions.grad)
                self.assertIsNotNone(cross_attentions.grad)
1327
1328
1329
1330
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            hidden_states.retain_grad()
1331
1332
1333
1334

            if self.has_attentions:
                attentions = outputs.attentions[0]
                attentions.retain_grad()
1335
1336
1337
1338

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

            self.assertIsNotNone(hidden_states.grad)
1339
1340
1341

            if self.has_attentions:
                self.assertIsNotNone(attentions.grad)
1342

Pradhy729's avatar
Pradhy729 committed
1343
    def test_feed_forward_chunking(self):
Lysandre's avatar
Lysandre committed
1344
1345
1346
1347
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Pradhy729's avatar
Pradhy729 committed
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
        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))

1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
    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
1445
    def test_resize_tokens_embeddings(self):
Lysandre's avatar
Lysandre committed
1446
1447
1448
1449
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Patrick von Platen's avatar
Patrick von Platen committed
1450
        if not self.test_resize_embeddings:
1451
1452
1453
1454
1455
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
1456
            model.to(torch_device)
1457

Patrick von Platen's avatar
Patrick von Platen committed
1458
1459
1460
            if self.model_tester.is_training is False:
                model.eval()

1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
            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)
1471
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
1472
            model(**self._prepare_for_class(inputs_dict, model_class))
1473
1474
1475
1476
1477
1478
1479

            # 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)

1480
1481
1482
            # 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)
1483
1484
1485
1486

            # 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)
1487
            model(**self._prepare_for_class(inputs_dict, model_class))
1488

1489
1490
1491
1492
1493
1494
1495
1496
            # 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)

1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
            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
1508
1509
1510
            self.assertTrue(model_embed.weight.shape[0], model.config.vocab_size)
            self.assertTrue(model.config.vocab_size, model.vocab_size)

1511
1512
1513
            model_embed = model.resize_token_embeddings(model_vocab_size + 13, pad_to_multiple_of=64)
            self.assertTrue(model_embed.weight.shape[0] // 64, 0)

1514
1515
1516
1517
1518
            # 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)

1519
1520
1521
1522
1523
1524
            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)

1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
    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
1576
    def test_model_common_attributes(self):
1577
1578
1579
1580
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1581
1582
            self.assertIsInstance(model.get_input_embeddings(), (nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(nn.Embedding(10, 10))
1583
            x = model.get_output_embeddings()
1584
            self.assertTrue(x is None or isinstance(x, nn.Linear))
1585

1586
1587
1588
1589
1590
1591
1592
    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)

1593
    def test_correct_missing_keys(self):
1594
1595
        if not self.test_missing_keys:
            return
1596
1597
1598
1599
1600
1601
1602
        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):
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
                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

1616
1617
1618
                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)
1619
                    self.assertGreater(len(loading_info["missing_keys"]), 0, model.__class__.__name__)
1620

1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
    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))

1669
1670
    @require_safetensors
    def test_can_use_safetensors(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
1671
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
        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
1688
1689
                # Checking there was no complain of missing weights
                self.assertEqual(infos["missing_keys"], [])
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705

                # 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
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
    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
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
    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
1750
1751
1752
1753
1754
            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
1755

Sylvain Gugger's avatar
Sylvain Gugger committed
1756
1757
    def test_model_weights_reload_no_missing_tied_weights(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
1758
        for model_class in self.all_model_classes:
Sylvain Gugger's avatar
Sylvain Gugger committed
1759
1760
1761
            model = model_class(config)
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.save_pretrained(tmp_dir)
1762
1763
1764

                # 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.
1765
1766
                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
1767
                model_reloaded, infos = model_class.from_pretrained(tmp_dir, output_loading_info=True)
1768
1769
1770
1771

                prefix = f"{model_reloaded.base_model_prefix}."
                params = dict(model_reloaded.named_parameters())
                params.update(dict(model_reloaded.named_buffers()))
1772
                param_names = {k[len(prefix) :] if k.startswith(prefix) else k for k in params.keys()}
1773
1774
1775
1776

                missing_keys = set(infos["missing_keys"])

                extra_missing = missing_keys - param_names
Sylvain Gugger's avatar
Sylvain Gugger committed
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
                # 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)
1788
1789
1790
1791

                self.assertEqual(
                    extra_missing,
                    set(),
Sylvain Gugger's avatar
Sylvain Gugger committed
1792
1793
                    f"This model {model_class.__name__} might be missing some `keys_to_ignore`: {extra_missing}. "
                    f"For debugging, tied parameters are {tied_params}",
1794
1795
                )

Sylvain Gugger's avatar
Sylvain Gugger committed
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
                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`",
                )
1816

1817
1818
1819
    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
1820
1821
1822
1823
        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

1824
1825
1826
1827
1828
1829
1830
1831
1832
        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
1833
1834
1835
1836
1837
                    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)
1838
1839
1840
1841
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
Sam Shleifer's avatar
Sam Shleifer committed
1842
1843
1844
                            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
1845
1846
1847
1848
1849
1850
                            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)}."
                            ),
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
                        )

                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})

1876
1877
1878
1879
            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})
1880

1881
1882
1883
1884
1885
1886
1887
1888
1889
                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}
                )
1890

1891
1892
1893
1894
    # 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"""
1895

1896
1897
1898
        for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
            if k in inputs_dict:
                attention_mask = inputs_dict[k]
1899

1900
1901
1902
1903
1904
1905
                # 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
                )
1906

1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
                # 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)"""

1924
1925
        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}
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951

        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):
1952
        """Check the outputs from PyTorch and TensorFlow models are close enough. Checks are done in a recursive way.
1953

1954
1955
1956
1957
1958
1959
1960
1961
        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.
        """
1962

1963
1964
1965
        self.assertEqual(type(name), str)
        if attributes is not None:
            self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")
1966

1967
1968
1969
1970
1971
1972
        # 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",
            )
1973

1974
1975
1976
            # 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)
1977

1978
1979
            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]
1980

1981
            self.assertEqual(tf_keys, pt_keys, f"{name}: Output keys differ between TF and PyTorch")
1982

1983
            # convert to the case of `tuple`
1984
            # appending each key to the current (string) `name`
1985
1986
1987
1988
            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
            )
1989

1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
        # 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),
2000
                    f"{name}: The tuple `attributes` should have the same length as `tf_outputs`",
2001
                )
2002
            else:
2003
                # case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `name`
2004
                attributes = tuple([f"{name}_{idx}" for idx in range(len(tf_outputs))])
2005

2006
2007
            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)
2008

2009
2010
2011
2012
        elif isinstance(tf_outputs, tf.Tensor):
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `tf_outputs` is"
            )
2013

2014
2015
            tf_outputs = tf_outputs.numpy()
            pt_outputs = pt_outputs.detach().to("cpu").numpy()
2016

2017
2018
2019
            self.assertEqual(
                tf_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between TF and PyTorch"
            )
2020

2021
2022
2023
2024
            # 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])
2025

2026
2027
            tf_nans = np.isnan(tf_outputs)
            pt_nans = np.isnan(pt_outputs)
2028

2029
2030
2031
2032
            pt_outputs[tf_nans] = 0
            tf_outputs[tf_nans] = 0
            pt_outputs[pt_nans] = 0
            tf_outputs[pt_nans] = 0
2033

2034
            max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
2035
            self.assertLessEqual(max_diff, tol, f"{name}: Difference between PyTorch and TF is {max_diff} (>= {tol}).")
2036
2037
        else:
            raise ValueError(
2038
                "`tf_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `tf.Tensor`. Got"
Sylvain Gugger's avatar
Sylvain Gugger committed
2039
                f" {type(tf_outputs)} instead."
2040
2041
            )

2042
2043
2044
2045
    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
2046
            if isinstance(tensor, bool):
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
                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)
2059

2060
        return tf_inputs_dict
2061

2062
2063
    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)
2064

2065
2066
2067
2068
        # 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()
        }
2069

2070
2071
        # send pytorch model to the correct device
        pt_model.to(torch_device)
2072

2073
2074
        # Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
        pt_model.eval()
2075

2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
        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
2090
    def test_pt_tf_model_equivalence(self, allow_missing_keys=False):
2091
        import transformers
2092
2093

        for model_class in self.all_model_classes:
2094
2095
2096
            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
2097
            if not hasattr(transformers, tf_model_class_name):
2098
                # transformers does not have this model in TF version yet
2099
2100
                return

2101
2102
2103
            # Output all for aggressive testing
            config.output_hidden_states = True
            config.output_attentions = self.has_attentions
2104

2105
2106
2107
2108
            # 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)
2109
2110

            tf_model_class = getattr(transformers, tf_model_class_name)
2111
2112

            pt_model = model_class(config)
2113
2114
2115
2116
2117
2118
2119
2120
2121
            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,
            )
2122
2123
2124
2125
2126
2127
2128
2129
2130

            # 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")

2131
            pt_inputs_dict = {k: v for k, v in pt_inputs_dict.items() if k in tf_input_keys}
2132
2133
2134
2135
            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.
2136
            if not set(pt_inputs_dict_with_labels.keys()).symmetric_difference(pt_inputs_dict.keys()):
2137
                pt_inputs_dict_with_labels = None
2138
2139

            # Check we can load pt model in tf and vice-versa with model => model functions
2140
2141
            # 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
2142
2143
2144
2145
2146
2147
            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
            )
2148

2149
2150
2151
2152
2153
            # 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)
2154
2155
2156
2157
2158

            # 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
2159
2160
2161
                tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(
                    tf_model, pt_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2162
2163
2164

                tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
                tf_model.save_weights(tf_checkpoint_path)
Matt's avatar
Matt committed
2165
2166
2167
                pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(
                    pt_model, tf_checkpoint_path, allow_missing_keys=allow_missing_keys
                )
2168

2169
2170
2171
2172
2173
            # 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)
2174
2175
2176
2177
2178

    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}).")

2179
    def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
2180
2181
2182
2183
2184
2185
2186
2187
2188
        """
        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.
        """
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228

        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`",
                )
2229
            else:
2230
2231
2232
2233
2234
2235
                # 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)

2236
        elif isinstance(fx_outputs, jnp.ndarray):
2237
2238
2239
            self.assertTrue(
                isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `fx_outputs` is"
            )
2240
2241
2242
2243
2244

            # 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()

2245
2246
2247
2248
2249
2250
2251
2252
2253
            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])

2254
2255
2256
2257
2258
2259
2260
2261
            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

2262
2263
2264
2265
            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})."
            )
2266
2267
        else:
            raise ValueError(
2268
2269
                "`fx_outputs` should be an instance of `ModelOutput`, a `tuple`, or an instance of `jnp.ndarray`. Got"
                f" {type(fx_outputs)} instead."
2270
2271
            )

2272
2273
2274
2275
2276
2277
2278
2279
2280
    @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):
2281
                    # no flax model exists for this class
2282
2283
                    return

2284
2285
2286
2287
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2288
2289
                fx_model_class = getattr(transformers, fx_model_class_name)

2290
2291
2292
2293
2294
2295
                # 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

2296
2297
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2298

2299
2300
2301
2302
2303
2304
2305
2306
2307
                # 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}

2308
2309
2310
2311
2312
2313
                # 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
2314
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2315

2316
2317
2318
                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
                fx_model.params = fx_state

2319
2320
2321
                # send pytorch model to the correct device
                pt_model.to(torch_device)

2322
                with torch.no_grad():
2323
2324
                    pt_outputs = pt_model(**pt_inputs)
                fx_outputs = fx_model(**fx_inputs)
2325

2326
2327
2328
2329
                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)
2330
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2331
2332
2333
2334
2335

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

2336
2337
2338
2339
2340
2341
                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)
2342
                self.check_pt_flax_outputs(fx_outputs_loaded, pt_outputs, model_class)
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355

    @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

2356
2357
2358
2359
                # Output all for aggressive testing
                config.output_hidden_states = True
                config.output_attentions = self.has_attentions

2360
2361
                fx_model_class = getattr(transformers, fx_model_class_name)

2362
2363
2364
2365
2366
2367
                # 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

2368
2369
                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
2370

2371
2372
2373
2374
2375
2376
2377
2378
2379
                # 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}

2380
2381
2382
2383
                # 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()
                }
2384

2385
                # convert inputs to Flax
2386
                fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)}
2387

2388
2389
2390
2391
2392
2393
2394
                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)
2395

2396
2397
2398
2399
2400
2401
2402
2403
                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)
2404
                self.check_pt_flax_outputs(fx_outputs, pt_outputs, model_class)
2405
2406
2407
2408
2409

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

2410
2411
2412
2413
                # send pytorch model to the correct device
                pt_model_loaded.to(torch_device)
                pt_model_loaded.eval()

2414
                with torch.no_grad():
2415
                    pt_outputs_loaded = pt_model_loaded(**pt_inputs)
2416

2417
2418
2419
2420
                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)
2421
                self.check_pt_flax_outputs(fx_outputs, pt_outputs_loaded, model_class)
2422

Patrick von Platen's avatar
Patrick von Platen committed
2423
    def test_inputs_embeds(self):
2424
2425
2426
2427
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
2428
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
2429
            model.eval()
2430

2431
            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
2432

2433
2434
2435
2436
2437
2438
2439
2440
2441
            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)

2442
2443
            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
2444
                inputs["inputs_embeds"] = wte(input_ids)
2445
            else:
2446
2447
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
2448

thomwolf's avatar
thomwolf committed
2449
            with torch.no_grad():
Weizhen's avatar
Weizhen committed
2450
                model(**inputs)[0]
2451

2452
2453
    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
2454
2455
2456
2457
        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.
2458
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask", "cross_attn_head_mask"]
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
        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
2473
            model = nn.DataParallel(model)
2474
            with torch.no_grad():
2475
                _ = model(**self._prepare_for_class(inputs_dict, model_class))
2476

2477
2478
2479
    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
2480
            return
2481

2482
        # a candidate for testing_utils
2483
        def get_current_gpu_memory_use():
Patrick von Platen's avatar
Patrick von Platen committed
2484
            """returns a list of cuda memory allocations per GPU in MBs"""
2485
2486
2487
2488
2489

            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)
2490
2491
2492
2493
2494
2495
2496
2497
2498

            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()

2499
2500
2501
            # 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()
2502

2503
2504
            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
2505
2506
2507
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

2508
2509
2510
            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

2511
            del model
2512
            gc.collect()
2513
2514
            torch.cuda.empty_cache()

2515
2516
2517
            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()
2518
2519

            # Spread model layers over multiple devices
2520
            model = model_class(config)
2521
2522
2523
2524
            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
2525
            for n in range(len(model.device_map.keys())):
2526
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])
2527

2528
            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
2529
2530
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

2531
2532
            # 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
2533
2534
2535
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
2536
            gc.collect()
2537
2538
2539
2540
2541
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
2542
            return
2543
2544
2545
2546
2547
2548

        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)

2549
            def cast_to_device(dictionary, device):
2550
2551
2552
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
2553
                        output[k] = v.to(device)
2554
2555
2556
2557
2558
                    else:
                        output[k] = v

                return output

2559
2560
2561
2562
2563
2564
            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))
2565
2566
2567
2568
2569
2570
2571
2572

            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))

2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
    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
2587
    @require_accelerate
2588
    @mark.accelerate_tests
Sylvain Gugger's avatar
Sylvain Gugger committed
2589
    @require_torch_gpu
2590
    def test_disk_offload_bin(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
2591
2592
2593
2594
2595
2596
        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

2597
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2598
2599
            model = model_class(config).eval()
            model = model.to(torch_device)
2600
            torch.manual_seed(0)
2601
            base_output = model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2602
2603
2604

            model_size = compute_module_sizes(model)[""]
            with tempfile.TemporaryDirectory() as tmp_dir:
2605
                model.cpu().save_pretrained(tmp_dir, safe_serialization=False)
Sylvain Gugger's avatar
Sylvain Gugger committed
2606
2607

                with self.assertRaises(ValueError):
Yih-Dar's avatar
Yih-Dar committed
2608
2609
                    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
2610
2611
2612
                    # 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
2613
2614
                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
2615
2616
2617
2618
2619
                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)
2620
                torch.manual_seed(0)
2621
                new_output = new_model(**inputs_dict_class)
Sylvain Gugger's avatar
Sylvain Gugger committed
2622

2623
                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
Sylvain Gugger's avatar
Sylvain Gugger committed
2624

2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
    @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))

2657
    @require_accelerate
2658
    @mark.accelerate_tests
2659
2660
2661
2662
2663
2664
2665
2666
    @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

2667
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2668
2669
            model = model_class(config).eval()
            model = model.to(torch_device)
2670
2671

            torch.manual_seed(0)
2672
            base_output = model(**inputs_dict_class)
2673
2674
2675

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
Yih-Dar's avatar
Yih-Dar committed
2676
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
            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)
2687
2688

                    torch.manual_seed(0)
2689
                    new_output = new_model(**inputs_dict_class)
2690

2691
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2692
2693

    @require_accelerate
2694
    @mark.accelerate_tests
2695
2696
2697
2698
2699
2700
2701
2702
    @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

2703
            inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
2704
2705
            model = model_class(config).eval()
            model = model.to(torch_device)
2706
2707

            torch.manual_seed(0)
2708
            base_output = model(**inputs_dict_class)
2709
2710
2711

            model_size = compute_module_sizes(model)[""]
            # We test several splits of sizes to make sure it works.
2712
            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
            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)
2723
2724

                    torch.manual_seed(0)
2725
                    new_output = new_model(**inputs_dict_class)
2726

2727
                    self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
2728

2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
    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:
2739
2740
2741
            if model_class.__name__ not in [
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES),
2742
            ]:
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
                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"])

2761
2762
2763
2764
2765
2766
                    # 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
2767
2768
2769
2770
2771
                    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}"
                            )
2772

2773
2774
                    loss.backward()

2775
    def test_load_with_mismatched_shapes(self):
2776
2777
        if not self.test_mismatched_shapes:
            return
2778
2779
2780
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
2781
            if model_class.__name__ not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES):
2782
2783
2784
2785
2786
2787
2788
2789
                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
2790
                    with self.assertRaises(RuntimeError):
2791
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
2792
2793
                    with self.assertRaises(RuntimeError):
                        new_model_without_prefix = AutoModel.from_pretrained(tmp_dir, vocab_size=10)
2794
2795

                    logger = logging.get_logger("transformers.modeling_utils")
2796

2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
                    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)

2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
                    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)

2819
2820
2821
2822
2823
2824
2825
2826
2827
    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
2828
            ), f"{model_class} is too big for the common tests ({num_params})! It should have 1M max."
2829

2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_conversion(self):
        import torch

        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
2841
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865

            model = model_class(config)

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

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

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

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

        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
2866
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
2867

2868
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
            model = model_class(config)

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

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

2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
                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
2893

2894
2895
2896
2897
2898
2899
2900
2901
                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)
2902

2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
                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]
                )
2913

2914
                assert torch.allclose(logits_fa, logits, atol=4e-2, rtol=4e-2)
2915

2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
                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]
                )
2947

2948
                assert torch.allclose(logits_fa[1:], logits[1:], atol=4e-2, rtol=4e-2)
2949

2950
2951
                # check with inference + dropout
                model.train()
2952
                _ = model_fa(dummy_input, **other_inputs)
2953

2954
2955
2956
2957
2958
2959
2960
2961
2962
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_inference_padding_right(self):
        import torch

        for model_class in self.all_model_classes:
            if not model_class._supports_flash_attn_2:
2963
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
2964

2965
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
            model = model_class(config)

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

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

2980
2981
2982
2983
2984
                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)
2985

2986
2987
2988
2989
                if dummy_attention_mask is not None:
                    dummy_attention_mask = dummy_attention_mask[:1]
                    dummy_attention_mask[:, :-1] = 1
                    dummy_attention_mask[:, -1:] = 0
2990

2991
2992
                if model.config.is_encoder_decoder:
                    decoder_input_ids = inputs_dict.get("decoder_input_ids", dummy_input)[:1]
2993

2994
2995
2996
2997
2998
                    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)
2999

3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
                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]
                )
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
3044
3045
                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)
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055

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

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3056
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3057

3058
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3059
3060
3061
3062
3063
3064
3065
3066
            model = model_class(config)

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

3067
3068
3069
3070
3071
3072
3073
3074
                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
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087

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

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

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

3088
                self.assertTrue(torch.allclose(out, out_fa))
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098

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

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3099
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3100

3101
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3102
3103
3104
3105
3106
3107
3108
3109
            model = model_class(config)

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

3110
3111
3112
3113
3114
                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))
3115
                # make sure we do right padding
3116
3117
                dummy_attention_mask[:, :-1] = 1
                dummy_attention_mask[:, -1:] = 0
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130

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

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

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

3131
                self.assertTrue(torch.allclose(out, out_fa))
3132

3133
3134
3135
3136
3137
3138
3139
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_generate_use_cache(self):
        import torch

3140
3141
        max_new_tokens = 30

3142
3143
        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3144
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3145

3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
            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

3156
3157
3158
3159
3160
            model = model_class(config)

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

3161
                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
3162
3163

                model = model_class.from_pretrained(
3164
3165
3166
3167
                    tmpdirname,
                    torch_dtype=torch.float16,
                    use_flash_attention_2=True,
                    low_cpu_mem_usage=True,
3168
3169
3170
3171
                ).to(torch_device)

                # Just test that a large cache works as expected
                _ = model.generate(
3172
3173
3174
3175
3176
                    dummy_input,
                    attention_mask=dummy_attention_mask,
                    max_new_tokens=max_new_tokens,
                    do_sample=False,
                    use_cache=True,
3177
3178
                )

3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
    @require_flash_attn
    @require_torch_gpu
    @require_bitsandbytes
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_fp32_ln(self):
        import torch

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3189
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3190

3191
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
3192
3193
3194
3195
3196
            model = model_class(config)

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

3197
3198
3199
3200
3201
3202
                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"]
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216

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

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

3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
                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)
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
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
    @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))

3280
3281
3282
3283
3284
3285
3286
3287
3288
    @require_flash_attn
    @require_torch_gpu
    @mark.flash_attn_test
    @slow
    def test_flash_attn_2_from_config(self):
        import torch

        for model_class in self.all_generative_model_classes:
            if not model_class._supports_flash_attn_2:
3289
                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326

            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(
                config, use_flash_attention_2=True, torch_dtype=torch.bfloat16
            ).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)

                self.assertFalse(getattr(model_from_pretrained.config, "_flash_attn_2_enabled", False))

                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)

3327

3328
global_rng = random.Random()
thomwolf's avatar
thomwolf committed
3329
3330


thomwolf's avatar
thomwolf committed
3331
def ids_tensor(shape, vocab_size, rng=None, name=None):
3332
    #  Creates a random int32 tensor of the shape within the vocab size
thomwolf's avatar
thomwolf committed
3333
    if rng is None:
3334
        rng = global_rng
thomwolf's avatar
thomwolf committed
3335

thomwolf's avatar
thomwolf committed
3336
3337
3338
    total_dims = 1
    for dim in shape:
        total_dims *= dim
thomwolf's avatar
thomwolf committed
3339

thomwolf's avatar
thomwolf committed
3340
3341
3342
    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))
thomwolf's avatar
thomwolf committed
3343

3344
    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
thomwolf's avatar
thomwolf committed
3345
3346


3347
3348
3349
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
3350
3351
    # 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
3352
3353
3354
    return attn_mask


3355
def floats_tensor(shape, scale=1.0, rng=None, name=None):
Patrick von Platen's avatar
Patrick von Platen committed
3356
    """Creates a random float32 tensor"""
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
    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)

3368
    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()