test_modeling_common.py 95.7 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Aymeric Augustin's avatar
Aymeric Augustin committed
15

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

30
31
32
import numpy as np

import transformers
33
from huggingface_hub import Repository, delete_repo, login
Sylvain Gugger's avatar
Sylvain Gugger committed
34
from requests.exceptions import HTTPError
35
36
37
38
39
40
41
42
from transformers import (
    AutoConfig,
    AutoModel,
    AutoModelForSequenceClassification,
    PretrainedConfig,
    is_torch_available,
    logging,
)
43
from transformers.file_utils import WEIGHTS_NAME, is_flax_available, is_torch_fx_available
44
from transformers.models.auto import get_values
Sylvain Gugger's avatar
Sylvain Gugger committed
45
46
47
48
from transformers.testing_utils import (
    PASS,
    USER,
    CaptureLogger,
49
    TestCasePlus,
50
51
    is_pt_flax_cross_test,
    is_pt_tf_cross_test,
Sylvain Gugger's avatar
Sylvain Gugger committed
52
53
54
55
56
57
    is_staging_test,
    require_torch,
    require_torch_multi_gpu,
    slow,
    torch_device,
)
58

Aymeric Augustin's avatar
Aymeric Augustin committed
59

60
61
sys.path.append(str(Path(__file__).parent.parent / "utils"))

62
from test_module.custom_configuration import CustomConfig, NoSuperInitConfig  # noqa E402
63
64


65
if is_torch_available():
66
    import torch
67
    from torch import nn
thomwolf's avatar
thomwolf committed
68

69
    from test_module.custom_modeling import CustomModel, NoSuperInitModel
70
    from transformers import (
71
        BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
NielsRogge's avatar
NielsRogge committed
72
        MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
73
        MODEL_FOR_CAUSAL_LM_MAPPING,
74
        MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
NielsRogge's avatar
NielsRogge committed
75
        MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING,
76
        MODEL_FOR_MASKED_LM_MAPPING,
77
        MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
78
        MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
79
        MODEL_FOR_QUESTION_ANSWERING_MAPPING,
80
81
82
        MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
        MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
        MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
83
        MODEL_MAPPING,
84
85
86
87
        AdaptiveEmbedding,
        BertConfig,
        BertModel,
        PreTrainedModel,
88
        T5Config,
89
        T5ForConditionalGeneration,
90
    )
thomwolf's avatar
thomwolf committed
91

92
93
94
95
96
97
98
if is_flax_available():
    import jax.numpy as jnp
    from transformers.modeling_flax_pytorch_utils import (
        convert_pytorch_state_dict_to_flax,
        load_flax_weights_in_pytorch_model,
    )

99
if is_torch_fx_available():
100
    from transformers.utils.fx import symbolic_trace
101

102

103
104
105
def _config_zero_init(config):
    configs_no_init = copy.deepcopy(config)
    for key in configs_no_init.__dict__.keys():
106
        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
107
            setattr(configs_no_init, key, 1e-10)
108
109
    return configs_no_init

thomwolf's avatar
thomwolf committed
110

111
112
113
TINY_T5 = "patrickvonplaten/t5-tiny-random"


114
115
116
117
118
@require_torch
class ModelTesterMixin:

    model_tester = None
    all_model_classes = ()
119
    all_generative_model_classes = ()
120
    fx_compatible = False
Patrick von Platen's avatar
Patrick von Platen committed
121
122
123
    test_torchscript = True
    test_pruning = True
    test_resize_embeddings = True
124
    test_resize_position_embeddings = False
Patrick von Platen's avatar
Patrick von Platen committed
125
    test_head_masking = True
126
    test_mismatched_shapes = True
127
    test_missing_keys = True
128
    test_model_parallel = False
129
130
    is_encoder_decoder = False

131
132
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = copy.deepcopy(inputs_dict)
133
        if model_class in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
134
            inputs_dict = {
135
                k: v.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
136
                if isinstance(v, torch.Tensor) and v.ndim > 1
Sylvain Gugger's avatar
Sylvain Gugger committed
137
                else v
138
139
                for k, v in inputs_dict.items()
            }
140
141

        if return_labels:
142
            if model_class in get_values(MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
143
                inputs_dict["labels"] = torch.ones(self.model_tester.batch_size, dtype=torch.long, device=torch_device)
144
            elif model_class in get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING):
145
146
147
148
149
150
                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
                )
151
            elif model_class in [
152
153
154
                *get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING),
                *get_values(MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING),
                *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
155
            ]:
156
157
158
159
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )
            elif model_class in [
160
161
                *get_values(MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING),
                *get_values(MODEL_FOR_CAUSAL_LM_MAPPING),
NielsRogge's avatar
NielsRogge committed
162
                *get_values(MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING),
163
164
                *get_values(MODEL_FOR_MASKED_LM_MAPPING),
                *get_values(MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING),
165
166
167
168
            ]:
                inputs_dict["labels"] = torch.zeros(
                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
                )
NielsRogge's avatar
NielsRogge committed
169
170
171
172
173
            elif model_class in get_values(MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING):
                num_patches = self.model_tester.image_size // self.model_tester.patch_size
                inputs_dict["bool_masked_pos"] = torch.zeros(
                    (self.model_tester.batch_size, num_patches**2), dtype=torch.long, device=torch_device
                )
174
175
        return inputs_dict

Patrick von Platen's avatar
Patrick von Platen committed
176
    def test_save_load(self):
177
178
179
180
181
182
183
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
184
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
185

186
            out_2 = outputs[0].cpu().numpy()
187
            out_2[np.isnan(out_2)] = 0
188

189
            with tempfile.TemporaryDirectory() as tmpdirname:
190
191
                model.save_pretrained(tmpdirname)
                model = model_class.from_pretrained(tmpdirname)
192
                model.to(torch_device)
193
                with torch.no_grad():
194
                    after_outputs = model(**self._prepare_for_class(inputs_dict, model_class))
thomwolf's avatar
thomwolf committed
195

196
197
198
                # Make sure we don't have nans
                out_1 = after_outputs[0].cpu().numpy()
                out_1[np.isnan(out_1)] = 0
thomwolf's avatar
thomwolf committed
199
200
                max_diff = np.amax(np.abs(out_1 - out_2))
                self.assertLessEqual(max_diff, 1e-5)
201

202
    def test_save_load_keys_to_ignore_on_save(self):
203
204
205
206
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
207
208
            _keys_to_ignore_on_save = getattr(model, "_keys_to_ignore_on_save", None)
            if _keys_to_ignore_on_save is None:
209
210
211
                continue

            # check the keys are in the original state_dict
212
            for k in _keys_to_ignore_on_save:
213
                self.assertIn(k, model.state_dict().keys(), "\n".join(model.state_dict().keys()))
214
215
216
217
218
219

            # check that certain keys didn't get saved with the model
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                output_model_file = os.path.join(tmpdirname, WEIGHTS_NAME)
                state_dict_saved = torch.load(output_model_file)
220
                for k in _keys_to_ignore_on_save:
221
                    self.assertNotIn(k, state_dict_saved.keys(), "\n".join(state_dict_saved.keys()))
222

Sylvain Gugger's avatar
Sylvain Gugger committed
223
224
225
                # Test we can load the state dict in the model, necessary for the checkpointing API in Trainer.
                load_result = model.load_state_dict(state_dict_saved, strict=False)
                self.assertTrue(
226
227
                    len(load_result.missing_keys) == 0
                    or set(load_result.missing_keys) == set(model._keys_to_ignore_on_save)
Sylvain Gugger's avatar
Sylvain Gugger committed
228
229
230
                )
                self.assertTrue(len(load_result.unexpected_keys) == 0)

231
232
233
234
235
236
237
238
239
240
241
    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)

242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
    def test_gradient_checkpointing_enable_disable(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

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

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

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

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

261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
    def _mock_init_weights(self, module):
        if hasattr(module, "weight") and module.weight is not None:
            module.weight.data.fill_(3)
        if hasattr(module, "bias") and module.bias is not None:
            module.bias.data.fill_(3)

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

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

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

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

            model_class_copy = CopyClass

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

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

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

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

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

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

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

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

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

        for model_class in self.all_model_classes:

            if model_class == base_class:
                continue

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

            base_class_copy = CopyClass

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

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

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

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

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

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

                for key in model_fast_init.state_dict().keys():
                    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")

Patrick von Platen's avatar
Patrick von Platen committed
358
    def test_initialization(self):
359
360
361
362
363
364
365
366
        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
367
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
368
                        [0.0, 1.0],
369
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
370
                    )
thomwolf's avatar
thomwolf committed
371

Patrick von Platen's avatar
Patrick von Platen committed
372
    def test_determinism(self):
373
374
375
376
377
378
379
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
380
381
                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
382

383
384
385
386
387
388
389
            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)

390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
    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",
                ]
406
                expected_arg_names.extend(
407
408
                    ["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
409
410
411
                    else ["encoder_outputs"]
                )
                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
412
413
414
415
            else:
                expected_arg_names = ["input_ids"]
                self.assertListEqual(arg_names[:1], expected_arg_names)

416
417
418
419
420
    def test_training(self):
        if not self.model_tester.is_training:
            return

        for model_class in self.all_model_classes:
421
422
423
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.return_dict = True

424
            if model_class in get_values(MODEL_MAPPING):
425
                continue
426

427
428
429
430
431
432
433
434
            model = model_class(config)
            model.to(torch_device)
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

    def test_training_gradient_checkpointing(self):
435
        if not self.model_tester.is_training:
436
437
438
            return

        for model_class in self.all_model_classes:
439
440
441
442
            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
            config.use_cache = False
            config.return_dict = True

443
            if model_class in get_values(MODEL_MAPPING) or not model_class.supports_gradient_checkpointing:
444
445
446
                continue
            model = model_class(config)
            model.to(torch_device)
447
            model.gradient_checkpointing_enable()
448
449
450
451
452
            model.train()
            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            loss = model(**inputs).loss
            loss.backward()

Patrick von Platen's avatar
Patrick von Platen committed
453
    def test_attention_outputs(self):
454
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
Weizhen's avatar
Weizhen committed
455
456
        config.return_dict = True

sshleifer's avatar
sshleifer committed
457
        seq_len = getattr(self.model_tester, "seq_length", None)
sshleifer's avatar
sshleifer committed
458
459
        decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
Weizhen's avatar
Weizhen committed
460
        decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
461
        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
Patrick von Platen's avatar
Patrick von Platen committed
462
463
464
        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
465
466

        for model_class in self.all_model_classes:
467
            inputs_dict["output_attentions"] = True
Joseph Liu's avatar
Joseph Liu committed
468
            inputs_dict["output_hidden_states"] = False
469
            config.return_dict = True
470
471
472
473
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
474
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
475
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
476
477
478
479
480
481
482
483
484
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            # 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():
485
486
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
487
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
Patrick von Platen's avatar
Patrick von Platen committed
488
489
490
491
492
493
494
495
496
497
498

            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],
                )
499
            out_len = len(outputs)
thomwolf's avatar
thomwolf committed
500

501
            if self.is_encoder_decoder:
502
                correct_outlen = 5
503

504
505
506
507
                # 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
508
                if model_class in get_values(MODEL_FOR_QUESTION_ANSWERING_MAPPING):
509
                    correct_outlen += 1  # start_logits and end_logits instead of only 1 output
510
511
                if "past_key_values" in outputs:
                    correct_outlen += 1  # past_key_values have been returned
Weizhen's avatar
Weizhen committed
512

Sam Shleifer's avatar
Sam Shleifer committed
513
514
                self.assertEqual(out_len, correct_outlen)

515
                # decoder attentions
516
                decoder_attentions = outputs.decoder_attentions
Sam Shleifer's avatar
Sam Shleifer committed
517
                self.assertIsInstance(decoder_attentions, (list, tuple))
518
                self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
thomwolf's avatar
thomwolf committed
519
                self.assertListEqual(
520
521
                    list(decoder_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
522
                )
thomwolf's avatar
thomwolf committed
523

524
525
526
527
528
529
530
531
532
533
534
535
536
                # 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,
                    ],
                )

537
            # Check attention is always last and order is fine
538
            inputs_dict["output_attentions"] = True
Joseph Liu's avatar
Joseph Liu committed
539
            inputs_dict["output_hidden_states"] = True
540
541
542
543
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
544
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
545

Weizhen's avatar
Weizhen committed
546
547
548
549
550
551
552
553
            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))

554
555
            self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions

556
            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
Patrick von Platen's avatar
Patrick von Platen committed
557
558
559
560
561
562
563
564
565
566
            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
567

568
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
569
    def test_torchscript(self):
570
571
        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
572

573
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
574
    def test_torchscript_output_attentions(self):
575
576
577
        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
578

579
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
580
    def test_torchscript_output_hidden_state(self):
581
582
583
        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
584

585
    def _create_and_check_torchscript(self, config, inputs_dict):
Patrick von Platen's avatar
Patrick von Platen committed
586
        if not self.test_torchscript:
587
            return
588

589
590
591
592
593
594
        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()
595
            inputs = self._prepare_for_class(inputs_dict, model_class)
thomwolf's avatar
thomwolf committed
596

597
            try:
598
                if model.config.is_encoder_decoder:
599
                    model.config.use_cache = False  # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
600
601
602
603
604
605
606
607
608
609
                    input_ids = inputs["input_ids"]
                    attention_mask = inputs["attention_mask"]
                    decoder_input_ids = inputs["decoder_input_ids"]
                    decoder_attention_mask = inputs["decoder_attention_mask"]
                    traced_model = torch.jit.trace(
                        model, (input_ids, attention_mask, decoder_input_ids, decoder_attention_mask)
                    )
                else:
                    input_ids = inputs["input_ids"]
                    traced_model = torch.jit.trace(model, input_ids)
610
611
            except RuntimeError:
                self.fail("Couldn't trace module.")
thomwolf's avatar
thomwolf committed
612

613
            with tempfile.TemporaryDirectory() as tmp_dir_name:
614
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
thomwolf's avatar
thomwolf committed
615

616
                try:
617
                    torch.jit.save(traced_model, pt_file_name)
618
619
                except Exception:
                    self.fail("Couldn't save module.")
thomwolf's avatar
thomwolf committed
620

621
622
623
624
                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")
LysandreJik's avatar
LysandreJik committed
625

626
627
            model.to(torch_device)
            model.eval()
thomwolf's avatar
thomwolf committed
628

629
630
            loaded_model.to(torch_device)
            loaded_model.eval()
thomwolf's avatar
thomwolf committed
631

632
633
634
            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

635
636
637
638
639
640
641
642
643
            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
            }

644
            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
thomwolf's avatar
thomwolf committed
645

646
647
648
649
650
651
652
653
654
655
656
            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)

657
            models_equal = True
658
            for layer_name, p1 in model_state_dict.items():
659
660
661
662
                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
663

664
            self.assertTrue(models_equal)
thomwolf's avatar
thomwolf committed
665

666
667
668
669
670
671
672
673
    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)

674
675
    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:
676
677
678
679
680
            return

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

681
        for model_class in self.all_model_classes:
682
683
684
685
686
687
688
689
690
691
692
693
            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)
                    input_names = ["input_ids", "attention_mask", "decoder_input_ids", "decoder_attention_mask"]
                    if labels is not None:
                        input_names.append("labels")
694
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
695

696
                    model_output = model(**filtered_inputs)
697

698
                    traced_model = symbolic_trace(model, input_names)
699
                    traced_output = traced_model(**filtered_inputs)
700
                else:
701
                    input_names = ["input_ids", "attention_mask", "token_type_ids"]
702
                    input_ids = inputs["input_ids"]
703

704
                    labels = inputs.get("labels", None)
705
706
                    start_positions = inputs.get("start_positions", None)
                    end_positions = inputs.get("end_positions", None)
707
708
                    if labels is not None:
                        input_names.append("labels")
709
710
711
712
                    if start_positions is not None:
                        input_names.append("start_positions")
                    if end_positions is not None:
                        input_names.append("end_positions")
713

714
715
                    filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
                    input_names = filtered_inputs.keys()
716

717
                    model_output = model(**filtered_inputs)
718

719
                    rank = len(input_ids.shape)
720
                    if rank not in [2, 3]:
721
722
723
                        raise NotImplementedError(
                            f"symbolic_trace automatic parameters inference not implemented for input of rank {rank}."
                        )
724

725
                    traced_model = symbolic_trace(model, input_names)
726
                    traced_output = traced_model(**filtered_inputs)
727
728
729
730

            except RuntimeError:
                self.fail("Couldn't trace module.")

731
732
733
734
735
736
737
738
739
740
741
742
743
            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)
744
            num_outputs = len(model_output)
745
746
747
748
749
750

            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}",
                )
751

Patrick von Platen's avatar
Patrick von Platen committed
752
753
    def test_headmasking(self):
        if not self.test_head_masking:
754
            return
755

756
757
758
        global_rng.seed(42)
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        global_rng.seed()
LysandreJik's avatar
LysandreJik committed
759

760
        inputs_dict["output_attentions"] = True
761
762
763
764
765
766
        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
767

768
769
770
            # 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
771
772
773
                self.model_tester.num_hidden_layers,
                self.model_tester.num_attention_heads,
                device=torch_device,
774
775
776
777
            )
            head_mask[0, 0] = 0
            head_mask[-1, :-1] = 0
            head_mask.requires_grad_(requires_grad=True)
778
            inputs = self._prepare_for_class(inputs_dict, model_class).copy()
779
            inputs["head_mask"] = head_mask
780
781
782
783
784
            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
785
786
                if "cross_attn_head_mask" in arg_names:
                    inputs["cross_attn_head_mask"] = head_mask
787
            outputs = model(**inputs, return_dict=True)
788
789
790
791
792
793
794
795
796

            # 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)
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817

            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)
818
                check_attentions_validity(outputs.cross_attentions)
819
820
            else:
                check_attentions_validity(outputs.attentions)
821

Patrick von Platen's avatar
Patrick von Platen committed
822
823
    def test_head_pruning(self):
        if not self.test_pruning:
824
825
826
            return

        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
827
828
829
830
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
831

832
833
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
834

835
            inputs_dict["output_attentions"] = True
836
837
838
839
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
840
841
842
843
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
844
845
            model.prune_heads(heads_to_prune)
            with torch.no_grad():
846
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
847

848
            attentions = outputs[-1]
849

850
851
852
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
LysandreJik's avatar
LysandreJik committed
853

Patrick von Platen's avatar
Patrick von Platen committed
854
855
    def test_head_pruning_save_load_from_pretrained(self):
        if not self.test_pruning:
856
            return
LysandreJik's avatar
LysandreJik committed
857

858
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
859
860
861
862
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
863
864
865

            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
866

867
            inputs_dict["output_attentions"] = True
868
869
870
871
            config.output_hidden_states = False
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
872
873
874
875
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
876
            model.prune_heads(heads_to_prune)
877

878
            with tempfile.TemporaryDirectory() as temp_dir_name:
879
880
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
881
                model.to(torch_device)
882

883
            with torch.no_grad():
884
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
885
886
887
888
            attentions = outputs[-1]
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
889

Patrick von Platen's avatar
Patrick von Platen committed
890
891
    def test_head_pruning_save_load_from_config_init(self):
        if not self.test_pruning:
892
            return
893

894
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
895
896
897
898
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
899

900
901
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
902

903
            inputs_dict["output_attentions"] = True
904
            config.output_hidden_states = False
905

906
907
908
909
            heads_to_prune = {
                0: list(range(1, self.model_tester.num_attention_heads)),
                -1: [0],
            }
910
            config.pruned_heads = heads_to_prune
911

912
913
914
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
915

916
            with torch.no_grad():
917
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
918
            attentions = outputs[-1]
919

920
921
922
            self.assertEqual(attentions[0].shape[-3], 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
923

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

928
        for model_class in self.all_model_classes:
Lysandre's avatar
Lysandre committed
929
930
931
932
            (
                config,
                inputs_dict,
            ) = self.model_tester.prepare_config_and_inputs_for_common()
933

934
935
            if "head_mask" in inputs_dict:
                del inputs_dict["head_mask"]
936

937
            inputs_dict["output_attentions"] = True
938
            config.output_hidden_states = False
939

940
941
            heads_to_prune = {0: [0], 1: [1, 2]}
            config.pruned_heads = heads_to_prune
942

943
944
945
            model = model_class(config=config)
            model.to(torch_device)
            model.eval()
946

947
            with torch.no_grad():
948
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
949
            attentions = outputs[-1]
950

951
952
953
954
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
thomwolf's avatar
thomwolf committed
955

956
            with tempfile.TemporaryDirectory() as temp_dir_name:
957
958
                model.save_pretrained(temp_dir_name)
                model = model_class.from_pretrained(temp_dir_name)
959
                model.to(torch_device)
thomwolf's avatar
thomwolf committed
960

961
            with torch.no_grad():
962
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
963
            attentions = outputs[-1]
LysandreJik's avatar
LysandreJik committed
964

965
966
967
968
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
thomwolf's avatar
thomwolf committed
969

970
971
            heads_to_prune = {0: [0], 2: [1, 2]}
            model.prune_heads(heads_to_prune)
972

973
            with torch.no_grad():
974
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
975
            attentions = outputs[-1]
976

977
978
979
980
            self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
            self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads - 2)
            self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
981

982
            self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
thomwolf's avatar
thomwolf committed
983

Patrick von Platen's avatar
Patrick von Platen committed
984
    def test_hidden_states_output(self):
Joseph Liu's avatar
Joseph Liu committed
985
        def check_hidden_states_output(inputs_dict, config, model_class):
986
            model = model_class(config)
987
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
988
            model.eval()
Joseph Liu's avatar
Joseph Liu committed
989

thomwolf's avatar
thomwolf committed
990
            with torch.no_grad():
991
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
992
993

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
994

Sylvain Gugger's avatar
Sylvain Gugger committed
995
996
997
998
            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)
999

Patrick von Platen's avatar
Patrick von Platen committed
1000
1001
1002
1003
1004
1005
1006
            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

1007
            self.assertListEqual(
Lysandre's avatar
Lysandre committed
1008
1009
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
1010
            )
thomwolf's avatar
thomwolf committed
1011

1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
            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
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
        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)

1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
    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
        config.output_attentions = True

        # 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)
1050

1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
        output = outputs[0]

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

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

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

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

            self.assertIsNotNone(encoder_hidden_states.grad)
            self.assertIsNotNone(encoder_attentions.grad)
            self.assertIsNotNone(decoder_hidden_states.grad)
            self.assertIsNotNone(decoder_attentions.grad)
            self.assertIsNotNone(cross_attentions.grad)
        else:
            # Encoder-/Decoder-only models
            hidden_states = outputs.hidden_states[0]
            attentions = outputs.attentions[0]

            hidden_states.retain_grad()
            attentions.retain_grad()

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

            self.assertIsNotNone(hidden_states.grad)
            self.assertIsNotNone(attentions.grad)

Pradhy729's avatar
Pradhy729 committed
1088
    def test_feed_forward_chunking(self):
Lysandre's avatar
Lysandre committed
1089
1090
1091
1092
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Pradhy729's avatar
Pradhy729 committed
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
        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))

1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
    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
1190
    def test_resize_tokens_embeddings(self):
Lysandre's avatar
Lysandre committed
1191
1192
1193
1194
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
Patrick von Platen's avatar
Patrick von Platen committed
1195
        if not self.test_resize_embeddings:
1196
1197
1198
1199
1200
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
1201
            model.to(torch_device)
1202

Patrick von Platen's avatar
Patrick von Platen committed
1203
1204
1205
            if self.model_tester.is_training is False:
                model.eval()

1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
            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)
1216
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
1217
            model(**self._prepare_for_class(inputs_dict, model_class))
1218
1219
1220
1221
1222
1223
1224

            # 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)

1225
1226
1227
            # 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)
1228
1229
1230
1231

            # 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)
1232
            model(**self._prepare_for_class(inputs_dict, model_class))
1233

1234
1235
1236
1237
1238
1239
1240
1241
            # 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)

1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
    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
1293
    def test_model_common_attributes(self):
1294
1295
1296
1297
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1298
1299
            self.assertIsInstance(model.get_input_embeddings(), (nn.Embedding, AdaptiveEmbedding))
            model.set_input_embeddings(nn.Embedding(10, 10))
1300
            x = model.get_output_embeddings()
1301
            self.assertTrue(x is None or isinstance(x, nn.Linear))
1302

1303
1304
1305
1306
1307
1308
1309
    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)

1310
    def test_correct_missing_keys(self):
1311
1312
        if not self.test_missing_keys:
            return
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
        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):
                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)
1323
                    with self.subTest(msg=f"Missing keys for {model.__class__.__name__}"):
1324
1325
                        self.assertGreater(len(loading_info["missing_keys"]), 0)

1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
    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))

1374
1375
1376
    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
1377
1378
1379
1380
        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

1381
1382
1383
1384
1385
1386
1387
1388
1389
        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
1390
1391
1392
1393
1394
                    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)
1395
1396
1397
1398
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
Sam Shleifer's avatar
Sam Shleifer committed
1399
1400
1401
                            torch.allclose(
                                set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
                            ),
1402
                            msg=f"Tuple and dict output are not equal. Difference: {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`: {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}.",
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
                        )

                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)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            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})

            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}
            )

1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
    @is_pt_tf_cross_test
    def test_pt_tf_model_equivalence(self):
        import numpy as np
        import tensorflow as tf

        import transformers

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            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 TF version yet
                return

            tf_model_class = getattr(transformers, tf_model_class_name)

            config.output_hidden_states = True

            tf_model = tf_model_class(config)
            pt_model = model_class(config)

            # 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")

            pt_inputs = self._prepare_for_class(inputs_dict, model_class)
            pt_inputs = {k: v for k, v in pt_inputs.items() if k in tf_input_keys}

            # Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
            pt_model.eval()
            tf_inputs_dict = {}
            for key, tensor in pt_inputs.items():
                # skip key that does not exist in tf
                if type(tensor) == bool:
                    tf_inputs_dict[key] = tensor
                elif key == "input_values":
1484
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
Yih-Dar's avatar
Yih-Dar committed
1485
                elif key == "pixel_values":
1486
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
Joao Gante's avatar
Joao Gante committed
1487
                elif key == "input_features":
1488
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
1489
                else:
1490
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.int32)
1491
1492
1493

            # Check we can load pt model in tf and vice-versa with model => model functions
            tf_model = transformers.load_pytorch_model_in_tf2_model(tf_model, pt_model, tf_inputs=tf_inputs_dict)
1494
            pt_model = transformers.load_tf2_model_in_pytorch_model(pt_model, tf_model).to(torch_device)
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504

            # need to rename encoder-decoder "inputs" for PyTorch
            #            if "inputs" in pt_inputs_dict and self.is_encoder_decoder:
            #                pt_inputs_dict["input_ids"] = pt_inputs_dict.pop("inputs")

            with torch.no_grad():
                pto = pt_model(**pt_inputs)
            tfo = tf_model(tf_inputs_dict, training=False)

            tf_hidden_states = tfo[0].numpy()
1505
            pt_hidden_states = pto[0].cpu().numpy()
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526

            tf_nans = np.copy(np.isnan(tf_hidden_states))
            pt_nans = np.copy(np.isnan(pt_hidden_states))

            pt_hidden_states[tf_nans] = 0
            tf_hidden_states[tf_nans] = 0
            pt_hidden_states[pt_nans] = 0
            tf_hidden_states[pt_nans] = 0

            max_diff = np.amax(np.abs(tf_hidden_states - pt_hidden_states))
            self.assertLessEqual(max_diff, 4e-2)

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

                tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
                tf_model.save_weights(tf_checkpoint_path)
                pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(pt_model, tf_checkpoint_path)
1527
                pt_model = pt_model.to(torch_device)
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537

            # Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
            pt_model.eval()
            tf_inputs_dict = {}
            for key, tensor in pt_inputs.items():
                # skip key that does not exist in tf
                if type(tensor) == bool:
                    tensor = np.array(tensor, dtype=bool)
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor, dtype=tf.int32)
                elif key == "input_values":
1538
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
Yih-Dar's avatar
Yih-Dar committed
1539
                elif key == "pixel_values":
1540
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
Joao Gante's avatar
Joao Gante committed
1541
                elif key == "input_features":
1542
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.float32)
1543
                else:
1544
                    tf_inputs_dict[key] = tf.convert_to_tensor(tensor.cpu().numpy(), dtype=tf.int32)
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554

            # need to rename encoder-decoder "inputs" for PyTorch
            #            if "inputs" in pt_inputs_dict and self.is_encoder_decoder:
            #                pt_inputs_dict["input_ids"] = pt_inputs_dict.pop("inputs")

            with torch.no_grad():
                pto = pt_model(**pt_inputs)

            tfo = tf_model(tf_inputs_dict)
            tfo = tfo[0].numpy()
1555
            pto = pto[0].cpu().numpy()
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
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
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
            tf_nans = np.copy(np.isnan(tfo))
            pt_nans = np.copy(np.isnan(pto))

            pto[tf_nans] = 0
            tfo[tf_nans] = 0
            pto[pt_nans] = 0
            tfo[pt_nans] = 0

            max_diff = np.amax(np.abs(tfo - pto))
            self.assertLessEqual(max_diff, 4e-2)

    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}).")

    @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__):

                # 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

                fx_model_class_name = "Flax" + model_class.__name__

                if not hasattr(transformers, fx_model_class_name):
                    return

                fx_model_class = getattr(transformers, fx_model_class_name)

                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
                # 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}

                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
                fx_model.params = fx_state

                with torch.no_grad():
                    pt_outputs = pt_model(**pt_inputs).to_tuple()

                # convert inputs to Flax
                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}
                fx_outputs = fx_model(**fx_inputs).to_tuple()
                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
                for fx_output, pt_output in zip(fx_outputs, pt_outputs):
                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)

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

                fx_outputs_loaded = fx_model_loaded(**fx_inputs).to_tuple()
                self.assertEqual(
                    len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch"
                )
                for fx_output_loaded, pt_output in zip(fx_outputs_loaded, pt_outputs):
                    self.assert_almost_equals(fx_output_loaded, pt_output.numpy(), 4e-2)

    @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__):
                # load corresponding PyTorch class
                pt_model = model_class(config).eval()

                # So we disable `use_cache` here for PyTorch model.
                pt_model.config.use_cache = False

                fx_model_class_name = "Flax" + model_class.__name__

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

                fx_model_class = getattr(transformers, fx_model_class_name)

                # load Flax class
                fx_model = fx_model_class(config, dtype=jnp.float32)
                # make sure only flax inputs are forward that actually exist in function args
                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()

                pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)

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

                # 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}

                with torch.no_grad():
                    pt_outputs = pt_model(**pt_inputs).to_tuple()

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

                fx_outputs = fx_model(**fx_inputs).to_tuple()
                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")

                for fx_output, pt_output in zip(fx_outputs, pt_outputs):
                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)

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

                with torch.no_grad():
                    pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()

                self.assertEqual(
                    len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch"
                )
                for fx_output, pt_output in zip(fx_outputs, pt_outputs_loaded):
                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)

Patrick von Platen's avatar
Patrick von Platen committed
1686
    def test_inputs_embeds(self):
1687
1688
1689
1690
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
1691
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
1692
            model.eval()
1693

1694
            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
Weizhen's avatar
Weizhen committed
1695

1696
1697
1698
1699
1700
1701
1702
1703
1704
            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)

1705
1706
            wte = model.get_input_embeddings()
            if not self.is_encoder_decoder:
1707
                inputs["inputs_embeds"] = wte(input_ids)
1708
            else:
1709
1710
                inputs["inputs_embeds"] = wte(encoder_input_ids)
                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
1711

thomwolf's avatar
thomwolf committed
1712
            with torch.no_grad():
Weizhen's avatar
Weizhen committed
1713
                model(**inputs)[0]
1714

1715
1716
    @require_torch_multi_gpu
    def test_multi_gpu_data_parallel_forward(self):
1717
1718
1719
1720
        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.
1721
        blacklist_non_batched_params = ["head_mask", "decoder_head_mask", "cross_attn_head_mask"]
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
        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
1736
            model = nn.DataParallel(model)
1737
            with torch.no_grad():
1738
                _ = model(**self._prepare_for_class(inputs_dict, model_class))
1739

1740
1741
1742
    @require_torch_multi_gpu
    def test_model_parallelization(self):
        if not self.test_model_parallel:
1743
            return
1744

1745
        # a candidate for testing_utils
1746
        def get_current_gpu_memory_use():
Patrick von Platen's avatar
Patrick von Platen committed
1747
            """returns a list of cuda memory allocations per GPU in MBs"""
1748
1749
1750
1751
1752

            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)
1753
1754
1755
1756
1757
1758
1759
1760
1761

            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()

1762
1763
1764
            # 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()
1765

1766
1767
            # Put model on device 0 and take a memory snapshot
            model = model_class(config)
1768
1769
1770
            model.to("cuda:0")
            memory_after_model_load = get_current_gpu_memory_use()

1771
1772
1773
            # The memory use on device 0 should be higher than it was initially.
            self.assertGreater(memory_after_model_load[0], memory_at_start[0])

1774
            del model
1775
            gc.collect()
1776
1777
            torch.cuda.empty_cache()

1778
1779
1780
            # 2. MP test
            # it's essential to re-calibrate the usage before the next stage
            memory_at_start = get_current_gpu_memory_use()
1781
1782

            # Spread model layers over multiple devices
1783
            model = model_class(config)
1784
1785
1786
1787
1788
            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
            for n in range(torch.cuda.device_count()):
1789
                self.assertGreater(memory_after_parallelization[n], memory_at_start[n])
1790

1791
            # Assert that the memory use of device 0 is lower than it was when the entire model was loaded on it
1792
1793
            self.assertLess(memory_after_parallelization[0], memory_after_model_load[0])

1794
1795
            # 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
1796
1797
1798
            self.assertGreater(memory_after_parallelization[1], memory_after_model_load[1])

            del model
1799
            gc.collect()
1800
1801
1802
1803
1804
            torch.cuda.empty_cache()

    @require_torch_multi_gpu
    def test_model_parallel_equal_results(self):
        if not self.test_model_parallel:
1805
            return
1806
1807
1808
1809
1810
1811

        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)

1812
            def cast_to_device(dictionary, device):
1813
1814
1815
                output = {}
                for k, v in dictionary.items():
                    if isinstance(v, torch.Tensor):
1816
                        output[k] = v.to(device)
1817
1818
1819
1820
1821
                    else:
                        output[k] = v

                return output

1822
1823
1824
1825
1826
1827
            model = model_class(config)
            output = model(**cast_to_device(inputs_dict, "cpu"))

            model.parallelize()

            parallel_output = model(**cast_to_device(inputs_dict, "cuda:0"))
1828
1829
1830
1831
1832
1833
1834
1835

            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))

1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
    @require_torch_multi_gpu
    def test_model_parallel_beam_search(self):
        if not self.test_model_parallel:
            return

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

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

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

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

                return output

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

1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
    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:
            if model_class not in get_values(MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING):
                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"])

1894
1895
1896
1897
1898
1899
                    # 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
1900
1901
1902
1903
1904
                    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}"
                            )
1905

1906
1907
                    loss.backward()

1908
    def test_load_with_mismatched_shapes(self):
1909
1910
        if not self.test_mismatched_shapes:
            return
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

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

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

                    # Fails when we don't set ignore_mismatched_sizes=True
1923
                    with self.assertRaises(RuntimeError):
1924
                        new_model = AutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
1925
1926
                    with self.assertRaises(RuntimeError):
                        new_model_without_prefix = AutoModel.from_pretrained(tmp_dir, vocab_size=10)
1927
1928

                    logger = logging.get_logger("transformers.modeling_utils")
1929

1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
                    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)

1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
                    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)

1952

1953
global_rng = random.Random()
thomwolf's avatar
thomwolf committed
1954
1955


thomwolf's avatar
thomwolf committed
1956
def ids_tensor(shape, vocab_size, rng=None, name=None):
1957
    #  Creates a random int32 tensor of the shape within the vocab size
thomwolf's avatar
thomwolf committed
1958
    if rng is None:
1959
        rng = global_rng
thomwolf's avatar
thomwolf committed
1960

thomwolf's avatar
thomwolf committed
1961
1962
1963
    total_dims = 1
    for dim in shape:
        total_dims *= dim
thomwolf's avatar
thomwolf committed
1964

thomwolf's avatar
thomwolf committed
1965
1966
1967
    values = []
    for _ in range(total_dims):
        values.append(rng.randint(0, vocab_size - 1))
thomwolf's avatar
thomwolf committed
1968

1969
    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
thomwolf's avatar
thomwolf committed
1970
1971


1972
1973
1974
1975
1976
1977
1978
def random_attention_mask(shape, rng=None, name=None):
    attn_mask = ids_tensor(shape, vocab_size=2, rng=None, name=None)
    # make sure that at least one token is attended to for each batch
    attn_mask[:, -1] = 1
    return attn_mask


1979
def floats_tensor(shape, scale=1.0, rng=None, name=None):
Patrick von Platen's avatar
Patrick von Platen committed
1980
    """Creates a random float32 tensor"""
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
    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)

1992
    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()
1993
1994


1995
@require_torch
1996
class ModelUtilsTest(TestCasePlus):
1997
    @slow
Patrick von Platen's avatar
Patrick von Platen committed
1998
    def test_model_from_pretrained(self):
1999
        for model_name in BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
thomwolf's avatar
thomwolf committed
2000
2001
2002
2003
2004
2005
2006
2007
            config = BertConfig.from_pretrained(model_name)
            self.assertIsNotNone(config)
            self.assertIsInstance(config, PretrainedConfig)

            model = BertModel.from_pretrained(model_name)
            model, loading_info = BertModel.from_pretrained(model_name, output_loading_info=True)
            self.assertIsNotNone(model)
            self.assertIsInstance(model, PreTrainedModel)
Lysandre Debut's avatar
Lysandre Debut committed
2008
2009
2010
2011
2012

            self.assertEqual(len(loading_info["missing_keys"]), 0)
            self.assertEqual(len(loading_info["unexpected_keys"]), 8)
            self.assertEqual(len(loading_info["mismatched_keys"]), 0)
            self.assertEqual(len(loading_info["error_msgs"]), 0)
thomwolf's avatar
thomwolf committed
2013
2014

            config = BertConfig.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
Lysandre Debut's avatar
Lysandre Debut committed
2015
2016
2017
2018

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

thomwolf's avatar
thomwolf committed
2019
2020
2021
            model = BertModel.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
            self.assertEqual(model.config.output_hidden_states, True)
            self.assertEqual(model.config, config)
2022
2023
2024
2025
2026

    def test_model_from_pretrained_with_different_pretrained_model_name(self):
        model = T5ForConditionalGeneration.from_pretrained(TINY_T5)
        self.assertIsNotNone(model)

2027
2028
        logger = logging.get_logger("transformers.configuration_utils")
        with CaptureLogger(logger) as cl:
2029
            BertModel.from_pretrained(TINY_T5)
2030
        self.assertTrue("You are using a model of type t5 to instantiate a model of type bert" in cl.out)
Sylvain Gugger's avatar
Sylvain Gugger committed
2031

2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
    @require_torch
    def test_model_from_config_torch_dtype(self):
        # test that the model can be instantiated with dtype of user's choice - as long as it's a
        # float dtype. To make it happen config.torch_dtype needs to be set before instantiating the
        # model from the config object.

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

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

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

    @require_torch
    def test_model_from_pretrained_torch_dtype(self):
        # test that the model can be instantiated with dtype of either
2054
2055
        # 1. explicit from_pretrained's torch_dtype argument
        # 2. via autodiscovery by looking at model weights (torch_dtype="auto")
2056
        # so if a model.half() was saved, we want it to be instantiated as such.
2057
2058
        #
        # test an explicit model class, but also AutoModel separately as the latter goes through a different code path
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
        model_path = self.get_auto_remove_tmp_dir()

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

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

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

        # test fp16 save_pretrained, loaded with auto-detection
        model = model.half()
        model.save_pretrained(model_path)
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype="auto")
2081
        self.assertEqual(model.config.torch_dtype, torch.float16)
2082
2083
        self.assertEqual(model.dtype, torch.float16)

2084
2085
2086
2087
2088
        # tests `config.torch_dtype` saving
        with open(f"{model_path}/config.json") as f:
            config_dict = json.load(f)
        self.assertEqual(config_dict["torch_dtype"], "float16")

2089
2090
2091
2092
        # test fp16 save_pretrained, loaded with the explicit fp16
        model = T5ForConditionalGeneration.from_pretrained(model_path, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

2093
2094
2095
2096
2097
2098
2099
2100
        # test AutoModel separately as it goes through a different path
        # test auto-detection
        model = AutoModel.from_pretrained(TINY_T5, torch_dtype="auto")
        self.assertEqual(model.dtype, torch.float32)
        # test forcing an explicit dtype
        model = AutoModel.from_pretrained(TINY_T5, torch_dtype=torch.float16)
        self.assertEqual(model.dtype, torch.float16)

2101
2102
2103
2104
2105
2106
2107
2108
2109
    def test_no_super_init_config_and_model(self):
        config = NoSuperInitConfig(attribute=32)
        model = NoSuperInitModel(config)

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

            model = NoSuperInitModel.from_pretrained(tmp_dir)

Sylvain Gugger's avatar
Sylvain Gugger committed
2110
2111
2112
2113
2114
2115

@require_torch
@is_staging_test
class ModelPushToHubTester(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
2116
        cls._token = login(username=USER, password=PASS)
Sylvain Gugger's avatar
Sylvain Gugger committed
2117
2118
2119
2120

    @classmethod
    def tearDownClass(cls):
        try:
2121
            delete_repo(token=cls._token, name="test-model")
Sylvain Gugger's avatar
Sylvain Gugger committed
2122
2123
2124
2125
        except HTTPError:
            pass

        try:
2126
            delete_repo(token=cls._token, name="test-model-org", organization="valid_org")
Sylvain Gugger's avatar
Sylvain Gugger committed
2127
2128
2129
        except HTTPError:
            pass

2130
        try:
2131
            delete_repo(token=cls._token, name="test-dynamic-model")
2132
2133
2134
        except HTTPError:
            pass

2135
2136
2137
2138
2139
        try:
            delete_repo(token=cls._token, name="test-dynamic-model-config")
        except HTTPError:
            pass

Sylvain Gugger's avatar
Sylvain Gugger committed
2140
2141
2142
2143
2144
2145
    def test_push_to_hub(self):
        config = BertConfig(
            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
        )
        model = BertModel(config)
        with tempfile.TemporaryDirectory() as tmp_dir:
2146
            model.save_pretrained(os.path.join(tmp_dir, "test-model"), push_to_hub=True, use_auth_token=self._token)
Sylvain Gugger's avatar
Sylvain Gugger committed
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158

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

    def test_push_to_hub_in_organization(self):
        config = BertConfig(
            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
        )
        model = BertModel(config)
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(
2159
                os.path.join(tmp_dir, "test-model-org"),
Sylvain Gugger's avatar
Sylvain Gugger committed
2160
2161
2162
2163
2164
2165
2166
2167
                push_to_hub=True,
                use_auth_token=self._token,
                organization="valid_org",
            )

            new_model = BertModel.from_pretrained("valid_org/test-model-org")
            for p1, p2 in zip(model.parameters(), new_model.parameters()):
                self.assertTrue(torch.equal(p1, p2))
2168
2169

    def test_push_to_hub_dynamic_model(self):
2170
2171
2172
2173
2174
        CustomConfig.register_for_auto_class()
        CustomModel.register_for_auto_class()

        config = CustomConfig(hidden_size=32)
        model = CustomModel(config)
2175
2176
2177
2178

        with tempfile.TemporaryDirectory() as tmp_dir:
            repo = Repository(tmp_dir, clone_from=f"{USER}/test-dynamic-model", use_auth_token=self._token)
            model.save_pretrained(tmp_dir)
2179
2180
2181
2182
2183
            # checks
            self.assertDictEqual(
                config.auto_map,
                {"AutoConfig": "custom_configuration.CustomConfig", "AutoModel": "custom_modeling.CustomModel"},
            )
2184
2185
2186
2187

            repo.push_to_hub()

        new_model = AutoModel.from_pretrained(f"{USER}/test-dynamic-model", trust_remote_code=True)
2188
2189
        # Can't make an isinstance check because the new_model is from the CustomModel class of a dynamic module
        self.assertEqual(new_model.__class__.__name__, "CustomModel")
2190
2191
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))
2192

2193
        config = AutoConfig.from_pretrained(f"{USER}/test-dynamic-model", trust_remote_code=True)
2194
        new_model = AutoModel.from_config(config, trust_remote_code=True)
2195
        self.assertEqual(new_model.__class__.__name__, "CustomModel")