test_trainer.py 118 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# coding=utf-8
# Copyright 2018 the HuggingFace Inc. team.
#
# 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.

16
import dataclasses
17
import gc
18
import json
19
import math
20
import os
21
import random
Sylvain Gugger's avatar
Sylvain Gugger committed
22
import re
23
import subprocess
24
import sys
25
import tempfile
26
import time
Julien Chaumond's avatar
Julien Chaumond committed
27
import unittest
28
from itertools import product
29
from pathlib import Path
30
from unittest.mock import Mock, patch
Julien Chaumond's avatar
Julien Chaumond committed
31

Sylvain Gugger's avatar
Sylvain Gugger committed
32
import numpy as np
33
from huggingface_hub import HfFolder, Repository, delete_repo
34
from parameterized import parameterized
Sylvain Gugger's avatar
Sylvain Gugger committed
35
from requests.exceptions import HTTPError
36

37
38
39
40
41
42
43
44
from transformers import (
    AutoTokenizer,
    IntervalStrategy,
    PretrainedConfig,
    TrainingArguments,
    is_torch_available,
    logging,
)
45
from transformers.testing_utils import (
Sylvain Gugger's avatar
Sylvain Gugger committed
46
    ENDPOINT_STAGING,
47
    TOKEN,
Sylvain Gugger's avatar
Sylvain Gugger committed
48
    USER,
49
    CaptureLogger,
50
    TestCasePlus,
51
    get_gpu_count,
52
    get_tests_dir,
Sylvain Gugger's avatar
Sylvain Gugger committed
53
    is_staging_test,
Yih-Dar's avatar
Yih-Dar committed
54
    require_accelerate,
55
    require_intel_extension_for_pytorch,
56
    require_optuna,
57
    require_ray,
58
    require_safetensors,
59
    require_sentencepiece,
60
    require_sigopt,
61
62
    require_tokenizers,
    require_torch,
63
64
    require_torch_bf16_cpu,
    require_torch_bf16_gpu,
65
    require_torch_gpu,
66
    require_torch_multi_gpu,
67
    require_torch_non_multi_gpu,
68
    require_torch_tensorrt_fx,
69
    require_torch_tf32,
70
    require_torch_up_to_2_gpus,
71
    require_torchdynamo,
72
    require_wandb,
73
74
    slow,
)
75
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
76
from transformers.training_args import OptimizerNames
77
from transformers.utils import (
78
79
    SAFE_WEIGHTS_INDEX_NAME,
    SAFE_WEIGHTS_NAME,
80
81
82
83
    WEIGHTS_INDEX_NAME,
    WEIGHTS_NAME,
    is_apex_available,
    is_bitsandbytes_available,
84
    is_safetensors_available,
85
86
    is_torchdistx_available,
)
87
from transformers.utils.hp_naming import TrialShortNamer
Julien Chaumond's avatar
Julien Chaumond committed
88
89
90
91


if is_torch_available():
    import torch
92
    from torch import nn
93
94
    from torch.utils.data import IterableDataset

95
    import transformers.optimization
Julien Chaumond's avatar
Julien Chaumond committed
96
97
    from transformers import (
        AutoModelForSequenceClassification,
98
        EarlyStoppingCallback,
Julien Chaumond's avatar
Julien Chaumond committed
99
100
        GlueDataset,
        GlueDataTrainingArguments,
101
102
        GPT2Config,
        GPT2LMHeadModel,
103
        LineByLineTextDataset,
104
        PreTrainedModel,
105
        Trainer,
106
        TrainerState,
Julien Chaumond's avatar
Julien Chaumond committed
107
    )
108
    from transformers.modeling_utils import unwrap_model
Julien Chaumond's avatar
Julien Chaumond committed
109

110
111
112
    if is_safetensors_available():
        import safetensors.torch

Julien Chaumond's avatar
Julien Chaumond committed
113

114
PATH_SAMPLE_TEXT = f"{get_tests_dir()}/fixtures/sample_text.txt"
Julien Chaumond's avatar
Julien Chaumond committed
115
116


Sylvain Gugger's avatar
Sylvain Gugger committed
117
class RegressionDataset:
Sylvain Gugger's avatar
Sylvain Gugger committed
118
    def __init__(self, a=2, b=3, length=64, seed=42, label_names=None):
Sylvain Gugger's avatar
Sylvain Gugger committed
119
        np.random.seed(seed)
Sylvain Gugger's avatar
Sylvain Gugger committed
120
        self.label_names = ["labels"] if label_names is None else label_names
Sylvain Gugger's avatar
Sylvain Gugger committed
121
122
        self.length = length
        self.x = np.random.normal(size=(length,)).astype(np.float32)
Sylvain Gugger's avatar
Sylvain Gugger committed
123
124
        self.ys = [a * self.x + b + np.random.normal(scale=0.1, size=(length,)) for _ in self.label_names]
        self.ys = [y.astype(np.float32) for y in self.ys]
Julien Chaumond's avatar
Julien Chaumond committed
125

Sylvain Gugger's avatar
Sylvain Gugger committed
126
127
128
129
    def __len__(self):
        return self.length

    def __getitem__(self, i):
Sylvain Gugger's avatar
Sylvain Gugger committed
130
131
132
        result = {name: y[i] for name, y in zip(self.label_names, self.ys)}
        result["input_x"] = self.x[i]
        return result
Sylvain Gugger's avatar
Sylvain Gugger committed
133
134


135
136
137
138
139
@dataclasses.dataclass
class RegressionTrainingArguments(TrainingArguments):
    a: float = 0.0
    b: float = 0.0

140
141
142
143
144
    def __post_init__(self):
        super().__post_init__()
        # save resources not dealing with reporting (also avoids the warning when it's not set)
        self.report_to = []

145

146
147
148
149
150
151
152
153
154
155
156
157
class RepeatDataset:
    def __init__(self, x, length=64):
        self.x = x
        self.length = length

    def __len__(self):
        return self.length

    def __getitem__(self, i):
        return {"input_ids": self.x, "labels": self.x}


158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
class DynamicShapesDataset:
    def __init__(self, length=64, seed=42, batch_size=8):
        self.length = length
        np.random.seed(seed)
        sizes = np.random.randint(1, 20, (length // batch_size,))
        # For easy batching, we make every batch_size consecutive samples the same size.
        self.xs = [np.random.normal(size=(s,)) for s in sizes.repeat(batch_size)]
        self.ys = [np.random.normal(size=(s,)) for s in sizes.repeat(batch_size)]

    def __len__(self):
        return self.length

    def __getitem__(self, i):
        return {"input_x": self.xs[i], "labels": self.ys[i]}


Sylvain Gugger's avatar
Sylvain Gugger committed
174
175
176
177
178
179
180
181
class AlmostAccuracy:
    def __init__(self, thresh=0.25):
        self.thresh = thresh

    def __call__(self, eval_pred):
        predictions, labels = eval_pred
        true = np.abs(predictions - labels) <= self.thresh
        return {"accuracy": true.astype(np.float32).mean().item()}
182

Julien Chaumond's avatar
Julien Chaumond committed
183

184
class RegressionModelConfig(PretrainedConfig):
185
    def __init__(self, a=0, b=0, double_output=False, random_torch=True, **kwargs):
186
187
188
189
        super().__init__(**kwargs)
        self.a = a
        self.b = b
        self.double_output = double_output
190
        self.random_torch = random_torch
191
        self.hidden_size = 1
192
193


194
195
196
if is_torch_available():

    class SampleIterableDataset(IterableDataset):
197
198
        def __init__(self, a=2, b=3, length=64, seed=42, label_names=None):
            self.dataset = RegressionDataset(a=a, b=b, length=length, seed=seed, label_names=label_names)
199
200

        def __iter__(self):
201
202
            for i in range(len(self.dataset)):
                yield self.dataset[i]
203

204
205
206
207
208
209
210
211
212
213
    class FiniteIterableDataset(SampleIterableDataset):
        def __init__(self, a=2, b=3, length=64, seed=42, label_names=None):
            super().__init__(a, b, length, seed, label_names)
            self.current_sample = 0

        def __iter__(self):
            while self.current_sample < len(self.dataset):
                yield self.dataset[self.current_sample]
                self.current_sample += 1

214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
    class MultiLoader:
        def __init__(self, loaders):
            self.loaders = loaders

        def __len__(self):
            return sum(len(loader) for loader in self.loaders)

        def __iter__(self):
            for loader in self.loaders:
                yield from loader

    class CustomDataloaderTrainer(Trainer):
        def get_train_dataloader(self):
            dataloaders = [super().get_train_dataloader(), super().get_train_dataloader()]
            return MultiLoader(dataloaders)

        def get_eval_dataloader(self, eval_dataset):
            dataloaders = [super().get_eval_dataloader(eval_dataset), super().get_eval_dataloader(eval_dataset)]
            return MultiLoader(dataloaders)

234
    class RegressionModel(nn.Module):
235
        def __init__(self, a=0, b=0, double_output=False):
Sylvain Gugger's avatar
Sylvain Gugger committed
236
            super().__init__()
237
238
            self.a = nn.Parameter(torch.tensor(a).float())
            self.b = nn.Parameter(torch.tensor(b).float())
239
240
            self.double_output = double_output
            self.config = None
Sylvain Gugger's avatar
Sylvain Gugger committed
241

Stas Bekman's avatar
Stas Bekman committed
242
        def forward(self, input_x, labels=None, **kwargs):
Sylvain Gugger's avatar
Sylvain Gugger committed
243
244
            y = input_x * self.a + self.b
            if labels is None:
245
                return (y, y) if self.double_output else (y,)
246
            loss = nn.functional.mse_loss(y, labels)
247
            return (loss, y, y) if self.double_output else (loss, y)
Sylvain Gugger's avatar
Sylvain Gugger committed
248

249
    class RegressionDictModel(nn.Module):
250
251
        def __init__(self, a=0, b=0):
            super().__init__()
252
253
            self.a = nn.Parameter(torch.tensor(a).float())
            self.b = nn.Parameter(torch.tensor(b).float())
254
255
            self.config = None

Stas Bekman's avatar
Stas Bekman committed
256
        def forward(self, input_x, labels=None, **kwargs):
257
258
259
            y = input_x * self.a + self.b
            result = {"output": y}
            if labels is not None:
260
                result["loss"] = nn.functional.mse_loss(y, labels)
261
262
            return result

263
264
265
266
267
268
    class RegressionPreTrainedModel(PreTrainedModel):
        config_class = RegressionModelConfig
        base_model_prefix = "regression"

        def __init__(self, config):
            super().__init__(config)
269
270
            self.a = nn.Parameter(torch.tensor(config.a).float())
            self.b = nn.Parameter(torch.tensor(config.b).float())
271
272
            self.double_output = config.double_output

Stas Bekman's avatar
Stas Bekman committed
273
        def forward(self, input_x, labels=None, **kwargs):
274
275
276
            y = input_x * self.a + self.b
            if labels is None:
                return (y, y) if self.double_output else (y,)
277
            loss = nn.functional.mse_loss(y, labels)
278
279
            return (loss, y, y) if self.double_output else (loss, y)

280
281
282
283
284
285
    class RegressionRandomPreTrainedModel(PreTrainedModel):
        config_class = RegressionModelConfig
        base_model_prefix = "regression"

        def __init__(self, config):
            super().__init__(config)
286
287
            self.a = nn.Parameter(torch.tensor(config.a).float())
            self.b = nn.Parameter(torch.tensor(config.b).float())
288
            self.random_torch = config.random_torch
289
290
291

        def forward(self, input_x, labels=None, **kwargs):
            y = input_x * self.a + self.b
292
293
            if self.random_torch:
                torch_rand = torch.randn(1).squeeze()
294
295
296
            np_rand = np.random.rand()
            rand_rand = random.random()

297
298
299
            if self.random_torch:
                y += 0.05 * torch_rand
            y += 0.05 * torch.tensor(np_rand + rand_rand)
300
301
302

            if labels is None:
                return (y,)
303
            loss = nn.functional.mse_loss(y, labels)
304
305
            return (loss, y)

306
    class TstLayer(nn.Module):
307
308
        def __init__(self, hidden_size):
            super().__init__()
309
310
311
312
313
            self.linear1 = nn.Linear(hidden_size, hidden_size)
            self.ln1 = nn.LayerNorm(hidden_size)
            self.linear2 = nn.Linear(hidden_size, hidden_size)
            self.ln2 = nn.LayerNorm(hidden_size)
            self.bias = nn.Parameter(torch.zeros(hidden_size))
314
315

        def forward(self, x):
316
317
            h = self.ln1(nn.functional.relu(self.linear1(x)))
            h = nn.functional.relu(self.linear2(x))
318
319
            return self.ln2(x + h + self.bias)

320
    def get_regression_trainer(a=0, b=0, double_output=False, train_len=64, eval_len=64, pretrained=True, **kwargs):
Sylvain Gugger's avatar
Sylvain Gugger committed
321
322
323
        label_names = kwargs.get("label_names", None)
        train_dataset = RegressionDataset(length=train_len, label_names=label_names)
        eval_dataset = RegressionDataset(length=eval_len, label_names=label_names)
324
325
326
327

        model_init = kwargs.pop("model_init", None)
        if model_init is not None:
            model = None
328
        else:
329
330
331
332
333
334
            if pretrained:
                config = RegressionModelConfig(a=a, b=b, double_output=double_output)
                model = RegressionPreTrainedModel(config)
            else:
                model = RegressionModel(a=a, b=b, double_output=double_output)

Sylvain Gugger's avatar
Sylvain Gugger committed
335
336
337
        compute_metrics = kwargs.pop("compute_metrics", None)
        data_collator = kwargs.pop("data_collator", None)
        optimizers = kwargs.pop("optimizers", (None, None))
338
        output_dir = kwargs.pop("output_dir", "./regression")
339
        preprocess_logits_for_metrics = kwargs.pop("preprocess_logits_for_metrics", None)
340
341

        args = RegressionTrainingArguments(output_dir, a=a, b=b, **kwargs)
Sylvain Gugger's avatar
Sylvain Gugger committed
342
343
344
345
346
347
348
349
        return Trainer(
            model,
            args,
            data_collator=data_collator,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            compute_metrics=compute_metrics,
            optimizers=optimizers,
350
            model_init=model_init,
351
            preprocess_logits_for_metrics=preprocess_logits_for_metrics,
Sylvain Gugger's avatar
Sylvain Gugger committed
352
353
        )

354

355
class TrainerIntegrationCommon:
356
357
358
    def check_saved_checkpoints(self, output_dir, freq, total, is_pretrained=True, safe_weights=False):
        weights_file = WEIGHTS_NAME if not safe_weights else SAFE_WEIGHTS_NAME
        file_list = [weights_file, "training_args.bin", "optimizer.pt", "scheduler.pt", "trainer_state.json"]
359
360
361
362
363
364
365
366
367
        if is_pretrained:
            file_list.append("config.json")
        for step in range(freq, total, freq):
            checkpoint = os.path.join(output_dir, f"checkpoint-{step}")
            self.assertTrue(os.path.isdir(checkpoint))
            for filename in file_list:
                self.assertTrue(os.path.isfile(os.path.join(checkpoint, filename)))

    def check_best_model_has_been_loaded(
368
        self, output_dir, freq, total, trainer, metric, greater_is_better=False, is_pretrained=True, safe_weights=False
369
370
    ):
        checkpoint = os.path.join(output_dir, f"checkpoint-{(total // freq) * freq}")
371
        log_history = TrainerState.load_from_json(os.path.join(checkpoint, "trainer_state.json")).log_history
372
373
374
375
376
377
378
379
380
381

        values = [d[metric] for d in log_history]
        best_value = max(values) if greater_is_better else min(values)
        best_checkpoint = (values.index(best_value) + 1) * freq
        checkpoint = os.path.join(output_dir, f"checkpoint-{best_checkpoint}")
        if is_pretrained:
            best_model = RegressionPreTrainedModel.from_pretrained(checkpoint)
            best_model.to(trainer.args.device)
        else:
            best_model = RegressionModel()
382
383
384
385
            if not safe_weights:
                state_dict = torch.load(os.path.join(checkpoint, WEIGHTS_NAME))
            else:
                state_dict = safetensors.torch.load_file(os.path.join(checkpoint, SAFE_WEIGHTS_NAME))
386
            best_model.load_state_dict(state_dict)
387
            best_model.to(trainer.args.device)
388
389
390
391
392
393
        self.assertTrue(torch.allclose(best_model.a, trainer.model.a))
        self.assertTrue(torch.allclose(best_model.b, trainer.model.b))

        metrics = trainer.evaluate()
        self.assertEqual(metrics[metric], best_value)

394
395
396
397
398
399
400
401
    def check_trainer_state_are_the_same(self, trainer_state, trainer_state1):
        # We'll pop things so operate on copies.
        state = trainer_state.copy()
        state1 = trainer_state1.copy()
        # Log history main contain different logs for the time metrics (after resuming a training).
        log_history = state.pop("log_history", None)
        log_history1 = state1.pop("log_history", None)
        self.assertEqual(state, state1)
402
        skip_log_keys = ["train_runtime", "train_samples_per_second", "train_steps_per_second", "train_loss"]
403
        for log, log1 in zip(log_history, log_history1):
404
405
406
            for key in skip_log_keys:
                _ = log.pop(key, None)
                _ = log1.pop(key, None)
407
408
            self.assertEqual(log, log1)

409
    def convert_to_sharded_checkpoint(self, folder, save_safe=False, load_safe=False):
410
        # Converts a checkpoint of a regression model to a sharded checkpoint.
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
        if load_safe:
            loader = safetensors.torch.load_file
            weights_file = os.path.join(folder, SAFE_WEIGHTS_NAME)
        else:
            loader = torch.load
            weights_file = os.path.join(folder, WEIGHTS_NAME)

        if save_safe:
            extension = "safetensors"
            saver = safetensors.torch.save_file
            index_file = os.path.join(folder, SAFE_WEIGHTS_INDEX_NAME)
            shard_name = SAFE_WEIGHTS_NAME
        else:
            extension = "bin"
            saver = torch.save
            index_file = os.path.join(folder, WEIGHTS_INDEX_NAME)
            shard_name = WEIGHTS_NAME

        state_dict = loader(weights_file)

        os.remove(weights_file)
432
433
434
        keys = list(state_dict.keys())

        shard_files = [
435
436
            shard_name.replace(f".{extension}", f"-{idx+1:05d}-of-{len(keys):05d}.{extension}")
            for idx in range(len(keys))
437
438
439
        ]
        index = {"metadata": {}, "weight_map": {key: shard_files[i] for i, key in enumerate(keys)}}

440
        with open(index_file, "w", encoding="utf-8") as f:
441
442
443
444
            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
            f.write(content)

        for param_name, shard_file in zip(keys, shard_files):
445
            saver({param_name: state_dict[param_name]}, os.path.join(folder, shard_file))
446

447
448
449
450

@require_torch
@require_sentencepiece
@require_tokenizers
451
452
453
454
455
456
457
458
class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
    """
    Only tests that want to tap into the auto-pre-run 2 trainings:
    - self.default_trained_model
    - self.alternate_trained_model
    directly, or via check_trained_model
    """

459
460
    def setUp(self):
        super().setUp()
461
        args = TrainingArguments("..")
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
        self.n_epochs = args.num_train_epochs
        self.batch_size = args.train_batch_size
        trainer = get_regression_trainer(learning_rate=0.1)
        trainer.train()
        self.default_trained_model = (trainer.model.a, trainer.model.b)

        trainer = get_regression_trainer(learning_rate=0.1, seed=314)
        trainer.train()
        self.alternate_trained_model = (trainer.model.a, trainer.model.b)

    def check_trained_model(self, model, alternate_seed=False):
        # Checks a training seeded with learning_rate = 0.1
        (a, b) = self.alternate_trained_model if alternate_seed else self.default_trained_model
        self.assertTrue(torch.allclose(model.a, a))
        self.assertTrue(torch.allclose(model.b, b))

478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
    def test_reproducible_training(self):
        # Checks that training worked, model trained and seed made a reproducible training.
        trainer = get_regression_trainer(learning_rate=0.1)
        trainer.train()
        self.check_trained_model(trainer.model)

        # Checks that a different seed gets different (reproducible) results.
        trainer = get_regression_trainer(learning_rate=0.1, seed=314)
        trainer.train()
        self.check_trained_model(trainer.model, alternate_seed=True)

    def test_trainer_with_datasets(self):
        import datasets

        np.random.seed(42)
        x = np.random.normal(size=(64,)).astype(np.float32)
        y = 2.0 * x + 3.0 + np.random.normal(scale=0.1, size=(64,))
        train_dataset = datasets.Dataset.from_dict({"input_x": x, "label": y})

        # Base training. Should have the same results as test_reproducible_training
        model = RegressionModel()
        args = TrainingArguments("./regression", learning_rate=0.1)
        trainer = Trainer(model, args, train_dataset=train_dataset)
        trainer.train()
        self.check_trained_model(trainer.model)

        # Can return tensors.
        train_dataset.set_format(type="torch", dtype=torch.float32)
        model = RegressionModel()
        trainer = Trainer(model, args, train_dataset=train_dataset)
        trainer.train()
        self.check_trained_model(trainer.model)

        # Adding one column not used by the model should have no impact
        z = np.random.normal(size=(64,)).astype(np.float32)
        train_dataset = datasets.Dataset.from_dict({"input_x": x, "label": y, "extra": z})
        model = RegressionModel()
        trainer = Trainer(model, args, train_dataset=train_dataset)
        trainer.train()
        self.check_trained_model(trainer.model)

    def test_model_init(self):
        train_dataset = RegressionDataset()
        args = TrainingArguments("./regression", learning_rate=0.1)
        trainer = Trainer(args=args, train_dataset=train_dataset, model_init=lambda: RegressionModel())
        trainer.train()
        self.check_trained_model(trainer.model)

        # Re-training should restart from scratch, thus lead the same results.
        trainer.train()
        self.check_trained_model(trainer.model)

        # Re-training should restart from scratch, thus lead the same results and new seed should be used.
        trainer.args.seed = 314
        trainer.train()
        self.check_trained_model(trainer.model, alternate_seed=True)

    def test_gradient_accumulation(self):
        # Training with half the batch size but accumulation steps as 2 should give the same results.
        trainer = get_regression_trainer(
            gradient_accumulation_steps=2, per_device_train_batch_size=4, learning_rate=0.1
        )
        trainer.train()
        self.check_trained_model(trainer.model)

543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
    def test_training_loss(self):
        n_gpus = max(1, get_gpu_count())

        # With even logs
        trainer = get_regression_trainer(logging_steps=64 / (8 * n_gpus))
        trainer.train()
        log_history = trainer.state.log_history

        losses = [log["loss"] for log in log_history if "loss" in log]
        train_loss = log_history[-1]["train_loss"]
        self.assertAlmostEqual(sum(losses) / len(losses), train_loss, places=4)

        # With uneven logs
        trainer = get_regression_trainer(logging_steps=5)
        trainer.train()
        log_history = trainer.state.log_history

        # Training loss should be the same as before
        new_train_loss = log_history[-1]["train_loss"]
        self.assertAlmostEqual(train_loss, new_train_loss, places=4)

564
565
566
567
568
569
570
571
572
573
574
575
576
577
    def test_custom_optimizer(self):
        train_dataset = RegressionDataset()
        args = TrainingArguments("./regression")
        model = RegressionModel()
        optimizer = torch.optim.SGD(model.parameters(), lr=1.0)
        lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda x: 1.0)
        trainer = Trainer(model, args, train_dataset=train_dataset, optimizers=(optimizer, lr_scheduler))
        trainer.train()

        (a, b) = self.default_trained_model
        self.assertFalse(torch.allclose(trainer.model.a, a))
        self.assertFalse(torch.allclose(trainer.model.b, b))
        self.assertEqual(trainer.optimizer.state_dict()["param_groups"][0]["lr"], 1.0)

578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
    def test_reduce_lr_on_plateau_args(self):
        # test passed arguments for a custom ReduceLROnPlateau scheduler
        train_dataset = RegressionDataset(length=64)
        eval_dataset = RegressionDataset(length=64)
        args = TrainingArguments(
            "./regression",
            evaluation_strategy="epoch",
            metric_for_best_model="eval_loss",
        )
        model = RegressionModel()
        optimizer = torch.optim.SGD(model.parameters(), lr=1.0)
        lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=0.2, patience=5, cooldown=2)
        trainer = Trainer(
            model, args, train_dataset=train_dataset, eval_dataset=eval_dataset, optimizers=(optimizer, lr_scheduler)
        )
        trainer.train()

        self.assertIsInstance(trainer.lr_scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau)
        self.assertEqual(trainer.lr_scheduler.factor, 0.2)
        self.assertEqual(trainer.lr_scheduler.patience, 5)
        self.assertEqual(trainer.lr_scheduler.cooldown, 2)

    def test_reduce_lr_on_plateau(self):
        # test the ReduceLROnPlateau scheduler

        class TrainerWithLRLogs(Trainer):
            def log(self, logs):
                # the LR is computed after metrics and does not exist for the first epoch
                if hasattr(self.lr_scheduler, "_last_lr"):
                    logs["learning_rate"] = self.lr_scheduler._last_lr
                super().log(logs)

        train_dataset = RegressionDataset(length=64)
        eval_dataset = RegressionDataset(length=64)

        args = TrainingArguments(
            "./regression",
            lr_scheduler_type="reduce_lr_on_plateau",
            evaluation_strategy="epoch",
            metric_for_best_model="eval_loss",
            num_train_epochs=10,
            learning_rate=0.2,
        )
        model = RegressionModel()
        trainer = TrainerWithLRLogs(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)
        trainer.train()

        self.assertIsInstance(trainer.lr_scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau)
        patience = trainer.lr_scheduler.patience

        logs = trainer.state.log_history[1:]
        best_loss = logs[0]["eval_loss"]
        bad_epochs = 0
        for i, log in enumerate(logs[:-1]):  # Compare learning rate to next epoch's
            loss = log["eval_loss"]
            just_decreased = False
            if loss > best_loss:
                bad_epochs += 1
                if bad_epochs > patience:
                    self.assertLess(logs[i + 1]["learning_rate"][0], log["learning_rate"][0])
                    just_decreased = True
                    bad_epochs = 0
            else:
                best_loss = loss
                bad_epochs = 0
            if not just_decreased:
                self.assertEqual(logs[i + 1]["learning_rate"][0], log["learning_rate"][0])

646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
    def test_adafactor_lr_none(self):
        # test the special case where lr=None, since Trainer can't not have lr_scheduler

        from transformers.optimization import Adafactor, AdafactorSchedule

        train_dataset = RegressionDataset()
        args = TrainingArguments("./regression")
        model = RegressionModel()
        optimizer = Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None)
        lr_scheduler = AdafactorSchedule(optimizer)
        trainer = Trainer(model, args, train_dataset=train_dataset, optimizers=(optimizer, lr_scheduler))
        trainer.train()

        (a, b) = self.default_trained_model
        self.assertFalse(torch.allclose(trainer.model.a, a))
        self.assertFalse(torch.allclose(trainer.model.b, b))
        self.assertGreater(trainer.optimizer.state_dict()["param_groups"][0]["lr"], 0)

664
    @require_torch_gpu
665
    @require_torch_bf16_gpu
666
667
668
669
670
671
672
673
674
675
676
677
    def test_mixed_bf16(self):
        # very basic test
        trainer = get_regression_trainer(learning_rate=0.1, bf16=True)
        trainer.train()
        self.check_trained_model(trainer.model)

        # --bf16 --half_precision_backend apex can't be used together
        with self.assertRaises(ValueError):
            trainer = get_regression_trainer(learning_rate=0.1, bf16=True, half_precision_backend="apex")

        # will add more specific tests once there are some bugs to fix

678
679
680
681
682
683
684
685
    @require_torch_gpu
    @require_torch_tf32
    def test_tf32(self):
        # very basic test
        trainer = get_regression_trainer(learning_rate=0.1, tf32=True)
        trainer.train()
        self.check_trained_model(trainer.model)

686
687
688
689
690
691
692

@require_torch
@require_sentencepiece
@require_tokenizers
class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
    def setUp(self):
        super().setUp()
693
        args = TrainingArguments("..")
694
695
696
        self.n_epochs = args.num_train_epochs
        self.batch_size = args.train_batch_size

697
698
699
700
701
702
703
704
705
706
707
708
709
    def test_trainer_works_with_dict(self):
        # Edge case because Apex with mode O2 will change our models to return dicts. This test checks it doesn't break
        # anything.
        train_dataset = RegressionDataset()
        eval_dataset = RegressionDataset()
        model = RegressionDictModel()
        args = TrainingArguments("./regression")
        trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)
        trainer.train()
        _ = trainer.evaluate()
        _ = trainer.predict(eval_dataset)

    def test_evaluation_with_keys_to_drop(self):
710
        config = GPT2Config(vocab_size=100, n_positions=128, n_embd=32, n_layer=3, n_head=4)
711
712
713
714
715
716
717
718
719
720
721
722
723
        tiny_gpt2 = GPT2LMHeadModel(config)
        x = torch.randint(0, 100, (128,))
        eval_dataset = RepeatDataset(x)
        args = TrainingArguments("./test")
        trainer = Trainer(tiny_gpt2, args, eval_dataset=eval_dataset)
        # By default the past_key_values are removed
        result = trainer.predict(eval_dataset)
        self.assertTrue(isinstance(result.predictions, np.ndarray))
        # We can still get them by setting ignore_keys to []
        result = trainer.predict(eval_dataset, ignore_keys=[])
        self.assertTrue(isinstance(result.predictions, tuple))
        self.assertEqual(len(result.predictions), 2)

724
725
726
    def test_training_arguments_are_left_untouched(self):
        trainer = get_regression_trainer()
        trainer.train()
727
        args = TrainingArguments("./regression", report_to=[])
728
729
        dict1, dict2 = args.to_dict(), trainer.args.to_dict()
        for key in dict1.keys():
730
            # Logging dir can be slightly different as they default to something with the time.
Sylvain Gugger's avatar
Sylvain Gugger committed
731
            if key != "logging_dir":
732
                self.assertEqual(dict1[key], dict2[key])
733

Sylvain Gugger's avatar
Sylvain Gugger committed
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
    def test_number_of_steps_in_training(self):
        # Regular training has n_epochs * len(train_dl) steps
        trainer = get_regression_trainer(learning_rate=0.1)
        train_output = trainer.train()
        self.assertEqual(train_output.global_step, self.n_epochs * 64 / self.batch_size)

        # Check passing num_train_epochs works (and a float version too):
        trainer = get_regression_trainer(learning_rate=0.1, num_train_epochs=1.5)
        train_output = trainer.train()
        self.assertEqual(train_output.global_step, int(1.5 * 64 / self.batch_size))

        # If we pass a max_steps, num_train_epochs is ignored
        trainer = get_regression_trainer(learning_rate=0.1, max_steps=10)
        train_output = trainer.train()
        self.assertEqual(train_output.global_step, 10)

750
    @require_torch_bf16_cpu
751
752
753
754
755
756
    @require_intel_extension_for_pytorch
    def test_number_of_steps_in_training_with_ipex(self):
        for mix_bf16 in [True, False]:
            # Regular training has n_epochs * len(train_dl) steps
            trainer = get_regression_trainer(learning_rate=0.1, use_ipex=True, bf16=mix_bf16, no_cuda=True)
            train_output = trainer.train()
757
            self.assertEqual(train_output.global_step, self.n_epochs * 64 / trainer.args.train_batch_size)
758
759
760
761
762
763

            # Check passing num_train_epochs works (and a float version too):
            trainer = get_regression_trainer(
                learning_rate=0.1, num_train_epochs=1.5, use_ipex=True, bf16=mix_bf16, no_cuda=True
            )
            train_output = trainer.train()
764
            self.assertEqual(train_output.global_step, int(1.5 * 64 / trainer.args.train_batch_size))
765
766
767
768
769
770
771
772

            # If we pass a max_steps, num_train_epochs is ignored
            trainer = get_regression_trainer(
                learning_rate=0.1, max_steps=10, use_ipex=True, bf16=mix_bf16, no_cuda=True
            )
            train_output = trainer.train()
            self.assertEqual(train_output.global_step, 10)

773
    def test_logging_inf_nan_filter(self):
774
        config = GPT2Config(vocab_size=100, n_positions=128, n_embd=32, n_layer=3, n_head=4)
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
        tiny_gpt2 = GPT2LMHeadModel(config)
        x = torch.randint(0, 100, (128,))
        train_dataset = RepeatDataset(x)

        # Trainer without inf/nan filter
        args = TrainingArguments("./test", learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=False)
        trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset)
        trainer.train()
        log_history_no_filter = trainer.state.log_history

        # Trainer with inf/nan filter
        args = TrainingArguments("./test", learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=True)
        trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset)
        trainer.train()
        log_history_filter = trainer.state.log_history

        def is_any_loss_nan_or_inf(log_history):
            losses = [l["loss"] for l in log_history[:-1]]
            return any(math.isnan(x) for x in losses) or any(math.isinf(x) for x in losses)

        self.assertTrue(is_any_loss_nan_or_inf(log_history_no_filter))
        self.assertFalse(is_any_loss_nan_or_inf(log_history_filter))

Sylvain Gugger's avatar
Sylvain Gugger committed
798
    def test_train_and_eval_dataloaders(self):
799
        n_gpu = max(1, torch.cuda.device_count())
Sylvain Gugger's avatar
Sylvain Gugger committed
800
        trainer = get_regression_trainer(learning_rate=0.1, per_device_train_batch_size=16)
801
        self.assertEqual(trainer.get_train_dataloader().batch_size, 16 * n_gpu)
Sylvain Gugger's avatar
Sylvain Gugger committed
802
        trainer = get_regression_trainer(learning_rate=0.1, per_device_eval_batch_size=16)
803
        self.assertEqual(trainer.get_eval_dataloader().batch_size, 16 * n_gpu)
Sylvain Gugger's avatar
Sylvain Gugger committed
804
805
806
807
808

        # Check drop_last works
        trainer = get_regression_trainer(
            train_len=66, eval_len=74, learning_rate=0.1, per_device_train_batch_size=16, per_device_eval_batch_size=32
        )
809
810
        self.assertEqual(len(trainer.get_train_dataloader()), 66 // (16 * n_gpu) + 1)
        self.assertEqual(len(trainer.get_eval_dataloader()), 74 // (32 * n_gpu) + 1)
Sylvain Gugger's avatar
Sylvain Gugger committed
811
812
813
814
815
816
817
818
819

        trainer = get_regression_trainer(
            train_len=66,
            eval_len=74,
            learning_rate=0.1,
            per_device_train_batch_size=16,
            per_device_eval_batch_size=32,
            dataloader_drop_last=True,
        )
820
821
        self.assertEqual(len(trainer.get_train_dataloader()), 66 // (16 * n_gpu))
        self.assertEqual(len(trainer.get_eval_dataloader()), 74 // (32 * n_gpu))
Sylvain Gugger's avatar
Sylvain Gugger committed
822

823
        # Check passing a new dataset for evaluation works
Sylvain Gugger's avatar
Sylvain Gugger committed
824
        new_eval_dataset = RegressionDataset(length=128)
825
        self.assertEqual(len(trainer.get_eval_dataloader(new_eval_dataset)), 128 // (32 * n_gpu))
Sylvain Gugger's avatar
Sylvain Gugger committed
826

827
828
829
830
831
832
833
834
835
    # tests that we do not require dataloader to have a .dataset attribute
    def test_dataloader_without_dataset(self):
        train_dataset = RegressionDataset(length=128)
        trainer = CustomDataloaderTrainer(
            model=RegressionModel(), train_dataset=train_dataset, eval_dataset=train_dataset
        )
        trainer.train()
        trainer.evaluate()

836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
    def test_sampler_seed(self):
        # nb: we don't want to inherit from IterableDataset to hit the right code path
        class DummyDataset(torch.utils.data.Dataset):
            def __init__(self, length: int = 101):
                self.length = length

            def __len__(self):
                return self.length

            def __getitem__(self, i):
                if (i < 0) or (i >= self.length):
                    raise IndexError
                return {"input_ids": [i]}

        class DummyModel(PreTrainedModel):
            def __init__(self, num_params: int):
                super().__init__(PretrainedConfig())
                # Add some (unused) params. the point here is that randomness in model_init shouldn't influence
                # data loader order.
                self.params = nn.Parameter(torch.randn(num_params))

            def forward(self, input_ids, labels=None):
                if labels is not None:
                    return torch.tensor(0.0, device=input_ids.device), input_ids
                else:
                    return input_ids

        def _get_first_data_sample(num_params, seed, data_seed, **kwargs):
            with tempfile.TemporaryDirectory() as tmpdir:
                trainer = Trainer(
                    model_init=lambda: DummyModel(num_params),
                    args=TrainingArguments(
                        output_dir=tmpdir,
                        **kwargs,
                        seed=seed,
                        data_seed=data_seed,
                        local_rank=-1,
                    ),
                    train_dataset=DummyDataset(),
                )

                return next(iter(trainer.get_train_dataloader()))

        # test that the seed is passed to the sampler
        # the codepath we want to hit is world_size <= 1, and both group_by_length
        for group_by_length in [True, False]:
            sample42_1 = _get_first_data_sample(num_params=10, seed=42, data_seed=42, group_by_length=group_by_length)
            sample42_2 = _get_first_data_sample(num_params=11, seed=42, data_seed=42, group_by_length=group_by_length)
            self.assertTrue(torch.equal(sample42_1["input_ids"], sample42_2["input_ids"]))

            # should get same samples with different seed, so long as data_seed is the same
            sample42_3 = _get_first_data_sample(num_params=11, seed=11, data_seed=42, group_by_length=group_by_length)
            self.assertTrue(torch.equal(sample42_1["input_ids"], sample42_3["input_ids"]))

            # make sure we have some randomness in the samples if data_seed is different
            others = [
                _get_first_data_sample(num_params=i, seed=42, data_seed=i, group_by_length=group_by_length)
                for i in range(10)
            ]
            self.assertTrue(any(not torch.equal(sample42_1["input_ids"], sample["input_ids"]) for sample in others))

897
898
899
900
901
902
    @require_torch_multi_gpu
    def test_data_is_not_parallelized_when_model_is_parallel(self):
        model = RegressionModel()
        # Make the Trainer believe it's a parallelized model
        model.is_parallelizable = True
        model.model_parallel = True
903
904
        args = TrainingArguments("./regression", per_device_train_batch_size=16, per_device_eval_batch_size=16)
        trainer = Trainer(model, args, train_dataset=RegressionDataset(), eval_dataset=RegressionDataset())
905
906
        # Check the Trainer was fooled
        self.assertTrue(trainer.is_model_parallel)
907
        self.assertEqual(trainer.args.n_gpu, 1)
908
909
910
911
912
913
914

        # The batch size of the training and evaluation dataloaders should be 16, not 16 * n_gpu
        self.assertEqual(trainer.get_train_dataloader().batch_size, 16)
        self.assertEqual(len(trainer.get_train_dataloader()), 64 // 16)
        self.assertEqual(trainer.get_eval_dataloader().batch_size, 16)
        self.assertEqual(len(trainer.get_eval_dataloader()), 64 // 16)

Sylvain Gugger's avatar
Sylvain Gugger committed
915
916
917
918
    def test_evaluate(self):
        trainer = get_regression_trainer(a=1.5, b=2.5, compute_metrics=AlmostAccuracy())
        results = trainer.evaluate()

Sylvain Gugger's avatar
Sylvain Gugger committed
919
        x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
Sylvain Gugger's avatar
Sylvain Gugger committed
920
921
922
923
924
925
926
927
928
929
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

        # With a number of elements not a round multiple of the batch size
        trainer = get_regression_trainer(a=1.5, b=2.5, eval_len=66, compute_metrics=AlmostAccuracy())
        results = trainer.evaluate()

Sylvain Gugger's avatar
Sylvain Gugger committed
930
        x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
Sylvain Gugger's avatar
Sylvain Gugger committed
931
932
933
934
935
936
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
        # With logits preprocess
        trainer = get_regression_trainer(
            a=1.5,
            b=2.5,
            compute_metrics=AlmostAccuracy(),
            preprocess_logits_for_metrics=lambda logits, labels: logits + 1,
        )
        results = trainer.evaluate()

        x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred + 1, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
    def test_evaluate_with_jit(self):
        trainer = get_regression_trainer(a=1.5, b=2.5, compute_metrics=AlmostAccuracy(), jit_mode_eval=True)
        results = trainer.evaluate()

        x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

        # With a number of elements not a round multiple of the batch size
        trainer = get_regression_trainer(
            a=1.5, b=2.5, eval_len=66, compute_metrics=AlmostAccuracy(), jit_mode_eval=True
        )
        results = trainer.evaluate()

        x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

        # With logits preprocess
        trainer = get_regression_trainer(
            a=1.5,
            b=2.5,
            compute_metrics=AlmostAccuracy(),
            preprocess_logits_for_metrics=lambda logits, labels: logits + 1,
            jit_mode_eval=True,
        )
        results = trainer.evaluate()

        x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred + 1, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

994
    @require_torch_bf16_cpu
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
    @require_intel_extension_for_pytorch
    def test_evaluate_with_ipex(self):
        for mix_bf16 in [True, False]:
            trainer = get_regression_trainer(
                a=1.5, b=2.5, use_ipex=True, compute_metrics=AlmostAccuracy(), bf16=mix_bf16, no_cuda=True
            )
            results = trainer.evaluate()

            x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
            pred = 1.5 * x + 2.5
            expected_loss = ((pred - y) ** 2).mean()
            self.assertAlmostEqual(results["eval_loss"], expected_loss)
            expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
            self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

            # With a number of elements not a round multiple of the batch size
            trainer = get_regression_trainer(
                a=1.5,
                b=2.5,
                use_ipex=True,
                eval_len=66,
                compute_metrics=AlmostAccuracy(),
                bf16=mix_bf16,
                no_cuda=True,
            )
            results = trainer.evaluate()

            x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
            pred = 1.5 * x + 2.5
            expected_loss = ((pred - y) ** 2).mean()
            self.assertAlmostEqual(results["eval_loss"], expected_loss)
            expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
            self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

            # With logits preprocess
            trainer = get_regression_trainer(
                a=1.5,
                b=2.5,
                use_ipex=True,
                compute_metrics=AlmostAccuracy(),
                preprocess_logits_for_metrics=lambda logits, labels: logits + 1,
                bf16=mix_bf16,
                no_cuda=True,
            )
            results = trainer.evaluate()

            x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
            pred = 1.5 * x + 2.5
            expected_loss = ((pred - y) ** 2).mean()
            self.assertAlmostEqual(results["eval_loss"], expected_loss)
            expected_acc = AlmostAccuracy()((pred + 1, y))["accuracy"]
            self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

Sylvain Gugger's avatar
Sylvain Gugger committed
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
    def test_predict(self):
        trainer = get_regression_trainer(a=1.5, b=2.5)
        preds = trainer.predict(trainer.eval_dataset).predictions
        x = trainer.eval_dataset.x
        self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))

        # With a number of elements not a round multiple of the batch size
        trainer = get_regression_trainer(a=1.5, b=2.5, eval_len=66)
        preds = trainer.predict(trainer.eval_dataset).predictions
        x = trainer.eval_dataset.x
        self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))

1060
1061
1062
1063
        # With more than one output of the model
        trainer = get_regression_trainer(a=1.5, b=2.5, double_output=True)
        preds = trainer.predict(trainer.eval_dataset).predictions
        x = trainer.eval_dataset.x
1064
        self.assertEqual(len(preds), 2)
1065
1066
1067
        self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
        self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))

Sylvain Gugger's avatar
Sylvain Gugger committed
1068
1069
1070
1071
1072
1073
        # With more than one output/label of the model
        trainer = get_regression_trainer(a=1.5, b=2.5, double_output=True, label_names=["labels", "labels_2"])
        outputs = trainer.predict(trainer.eval_dataset)
        preds = outputs.predictions
        labels = outputs.label_ids
        x = trainer.eval_dataset.x
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
        self.assertEqual(len(preds), 2)
        self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
        self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))
        self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
        self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))

    def test_predict_with_jit(self):
        trainer = get_regression_trainer(a=1.5, b=2.5, jit_mode_eval=True)
        preds = trainer.predict(trainer.eval_dataset).predictions
        x = trainer.eval_dataset.x
        self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))

        # With a number of elements not a round multiple of the batch size
        trainer = get_regression_trainer(a=1.5, b=2.5, eval_len=66, jit_mode_eval=True)
        preds = trainer.predict(trainer.eval_dataset).predictions
        x = trainer.eval_dataset.x
        self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))

        # With more than one output of the model
        trainer = get_regression_trainer(a=1.5, b=2.5, double_output=True, jit_mode_eval=True)
        preds = trainer.predict(trainer.eval_dataset).predictions
        x = trainer.eval_dataset.x
        self.assertEqual(len(preds), 2)
        self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
        self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))

        # With more than one output/label of the model
        trainer = get_regression_trainer(
            a=1.5, b=2.5, double_output=True, label_names=["labels", "labels_2"], jit_mode_eval=True
        )
        outputs = trainer.predict(trainer.eval_dataset)
        preds = outputs.predictions
        labels = outputs.label_ids
        x = trainer.eval_dataset.x
1108
        self.assertEqual(len(preds), 2)
Sylvain Gugger's avatar
Sylvain Gugger committed
1109
1110
1111
1112
1113
        self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
        self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))
        self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
        self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))

1114
    @require_torch_bf16_cpu
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
    @require_intel_extension_for_pytorch
    def test_predict_with_ipex(self):
        for mix_bf16 in [True, False]:
            trainer = get_regression_trainer(a=1.5, b=2.5, use_ipex=True, bf16=mix_bf16, no_cuda=True)
            preds = trainer.predict(trainer.eval_dataset).predictions
            x = trainer.eval_dataset.x
            self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))

            # With a number of elements not a round multiple of the batch size
            trainer = get_regression_trainer(a=1.5, b=2.5, eval_len=66, use_ipex=True, bf16=mix_bf16, no_cuda=True)
            preds = trainer.predict(trainer.eval_dataset).predictions
            x = trainer.eval_dataset.x
            self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))

            # With more than one output of the model
            trainer = get_regression_trainer(
                a=1.5, b=2.5, double_output=True, use_ipex=True, bf16=mix_bf16, no_cuda=True
            )
            preds = trainer.predict(trainer.eval_dataset).predictions
            x = trainer.eval_dataset.x
            self.assertEqual(len(preds), 2)
            self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
            self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))

            # With more than one output/label of the model
            trainer = get_regression_trainer(
                a=1.5,
                b=2.5,
                double_output=True,
                label_names=["labels", "labels_2"],
                use_ipex=True,
                bf16=mix_bf16,
                no_cuda=True,
            )
            outputs = trainer.predict(trainer.eval_dataset)
            preds = outputs.predictions
            labels = outputs.label_ids
            x = trainer.eval_dataset.x
            self.assertEqual(len(preds), 2)
            self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
            self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))
            self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
            self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))

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
1190
1191
1192
1193
1194
    def test_dynamic_shapes(self):
        eval_dataset = DynamicShapesDataset(batch_size=self.batch_size)
        model = RegressionModel(a=2, b=1)
        args = TrainingArguments("./regression")
        trainer = Trainer(model, args, eval_dataset=eval_dataset)

        # Check evaluation can run to completion
        _ = trainer.evaluate()

        # Check predictions
        preds = trainer.predict(eval_dataset)
        for expected, seen in zip(eval_dataset.ys, preds.label_ids):
            self.assertTrue(np.array_equal(expected, seen[: expected.shape[0]]))
            self.assertTrue(np.all(seen[expected.shape[0] :] == -100))

        for expected, seen in zip(eval_dataset.xs, preds.predictions):
            self.assertTrue(np.array_equal(2 * expected + 1, seen[: expected.shape[0]]))
            self.assertTrue(np.all(seen[expected.shape[0] :] == -100))

        # Same tests with eval accumulation
        args = TrainingArguments("./regression", eval_accumulation_steps=2)
        trainer = Trainer(model, args, eval_dataset=eval_dataset)

        # Check evaluation can run to completion
        _ = trainer.evaluate()

        # Check predictions
        preds = trainer.predict(eval_dataset)
        for expected, seen in zip(eval_dataset.ys, preds.label_ids):
            self.assertTrue(np.array_equal(expected, seen[: expected.shape[0]]))
            self.assertTrue(np.all(seen[expected.shape[0] :] == -100))

        for expected, seen in zip(eval_dataset.xs, preds.predictions):
            self.assertTrue(np.array_equal(2 * expected + 1, seen[: expected.shape[0]]))
            self.assertTrue(np.all(seen[expected.shape[0] :] == -100))

1195
    def test_log_level(self):
1196
        # testing only --log_level (--log_level_replica requires multiple gpus and DDP and is tested elsewhere)
1197
1198
1199
        logger = logging.get_logger()
        log_info_string = "Running training"

1200
1201
        # test with the default log_level - should be the same as before and thus we test depending on is_info
        is_info = logging.get_verbosity() <= 20
1202
1203
1204
        with CaptureLogger(logger) as cl:
            trainer = get_regression_trainer()
            trainer.train()
1205
1206
1207
1208
        if is_info:
            self.assertIn(log_info_string, cl.out)
        else:
            self.assertNotIn(log_info_string, cl.out)
1209

1210
        # test with low log_level - lower than info
1211
1212
1213
1214
1215
        with CaptureLogger(logger) as cl:
            trainer = get_regression_trainer(log_level="debug")
            trainer.train()
        self.assertIn(log_info_string, cl.out)

1216
        # test with high log_level - should be quiet
1217
1218
1219
1220
1221
        with CaptureLogger(logger) as cl:
            trainer = get_regression_trainer(log_level="error")
            trainer.train()
        self.assertNotIn(log_info_string, cl.out)

1222
1223
1224
1225
1226
1227
1228
1229
    def test_save_checkpoints(self):
        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(output_dir=tmpdir, save_steps=5)
            trainer.train()
            self.check_saved_checkpoints(tmpdir, 5, int(self.n_epochs * 64 / self.batch_size))

        # With a regular model that is not a PreTrainedModel
        with tempfile.TemporaryDirectory() as tmpdir:
1230
            trainer = get_regression_trainer(output_dir=tmpdir, save_steps=5, pretrained=False)
1231
1232
1233
            trainer.train()
            self.check_saved_checkpoints(tmpdir, 5, int(self.n_epochs * 64 / self.batch_size), False)

1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
    @require_safetensors
    def test_safe_checkpoints(self):
        for save_safetensors in [True, False]:
            with tempfile.TemporaryDirectory() as tmpdir:
                trainer = get_regression_trainer(output_dir=tmpdir, save_steps=5, save_safetensors=save_safetensors)
                trainer.train()
                self.check_saved_checkpoints(
                    tmpdir, 5, int(self.n_epochs * 64 / self.batch_size), safe_weights=save_safetensors
                )

            # With a regular model that is not a PreTrainedModel
            with tempfile.TemporaryDirectory() as tmpdir:
                trainer = get_regression_trainer(
                    output_dir=tmpdir, save_steps=5, pretrained=False, save_safetensors=save_safetensors
                )
                trainer.train()
                self.check_saved_checkpoints(
                    tmpdir, 5, int(self.n_epochs * 64 / self.batch_size), False, safe_weights=save_safetensors
                )

1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
    @require_torch_multi_gpu
    def test_run_seq2seq_double_train_wrap_once(self):
        # test that we don't wrap the model more than once
        # since wrapping primarily happens on multi-gpu setup we want multiple gpus to test for
        # example DataParallel(DataParallel(model))

        trainer = get_regression_trainer()
        trainer.train()
        model_wrapped_before = trainer.model_wrapped
        trainer.train()
        model_wrapped_after = trainer.model_wrapped
        self.assertIs(model_wrapped_before, model_wrapped_after, "should be not wrapped twice")

1267
    @require_torch_up_to_2_gpus
1268
    def test_can_resume_training(self):
1269
1270
1271
        # This test will fail for more than 2 GPUs since the batch size will get bigger and with the number of
        # save_steps, the checkpoint will resume training at epoch 2 or more (so the data seen by the model
        # won't be the same since the training dataloader is shuffled).
1272

1273
        with tempfile.TemporaryDirectory() as tmpdir:
1274
1275
1276
1277
1278
1279
1280
            kwargs = {
                "output_dir": tmpdir,
                "train_len": 128,
                "save_steps": 5,
                "learning_rate": 0.1,
                "logging_steps": 5,
            }
1281
            trainer = get_regression_trainer(**kwargs)
1282
1283
1284
1285
1286
1287
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
            state = dataclasses.asdict(trainer.state)

            checkpoint = os.path.join(tmpdir, "checkpoint-5")

1288
            # Reinitialize trainer
1289
            trainer = get_regression_trainer(**kwargs)
1290

1291
            trainer.train(resume_from_checkpoint=checkpoint)
1292
1293
1294
1295
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
            state1 = dataclasses.asdict(trainer.state)
            self.assertEqual(a, a1)
            self.assertEqual(b, b1)
1296
            self.check_trainer_state_are_the_same(state, state1)
1297

1298
1299
1300
1301
            # Now check with a later checkpoint that it also works when we span over one epoch
            checkpoint = os.path.join(tmpdir, "checkpoint-15")

            # Reinitialize trainer and load model
1302
            trainer = get_regression_trainer(**kwargs)
1303

1304
            trainer.train(resume_from_checkpoint=checkpoint)
1305
1306
1307
1308
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
            state1 = dataclasses.asdict(trainer.state)
            self.assertEqual(a, a1)
            self.assertEqual(b, b1)
1309
            self.check_trainer_state_are_the_same(state, state1)
1310

1311
1312
        # With a regular model that is not a PreTrainedModel
        with tempfile.TemporaryDirectory() as tmpdir:
1313
1314
1315
1316
1317
1318
1319
            kwargs = {
                "output_dir": tmpdir,
                "train_len": 128,
                "save_steps": 5,
                "learning_rate": 0.1,
                "pretrained": False,
            }
1320
1321

            trainer = get_regression_trainer(**kwargs)
1322
1323
1324
1325
1326
1327
1328
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
            state = dataclasses.asdict(trainer.state)

            checkpoint = os.path.join(tmpdir, "checkpoint-5")

            # Reinitialize trainer and load model
1329
            trainer = get_regression_trainer(**kwargs)
1330

1331
            trainer.train(resume_from_checkpoint=checkpoint)
1332
1333
1334
1335
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
            state1 = dataclasses.asdict(trainer.state)
            self.assertEqual(a, a1)
            self.assertEqual(b, b1)
1336
            self.check_trainer_state_are_the_same(state, state1)
1337

1338
1339
1340
1341
            # Now check with a later checkpoint that it also works when we span over one epoch
            checkpoint = os.path.join(tmpdir, "checkpoint-15")

            # Reinitialize trainer and load model
1342
            trainer = get_regression_trainer(**kwargs)
1343

1344
            trainer.train(resume_from_checkpoint=checkpoint)
1345
1346
1347
1348
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
            state1 = dataclasses.asdict(trainer.state)
            self.assertEqual(a, a1)
            self.assertEqual(b, b1)
1349
            self.check_trainer_state_are_the_same(state, state1)
1350

1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
        # Now check failures

        # 1. fail to find a bogus checkpoint
        trainer = get_regression_trainer()
        with self.assertRaises(Exception) as context:
            trainer.train(resume_from_checkpoint=f"{checkpoint}-bogus")
        self.assertTrue("Can't find a valid checkpoint at" in str(context.exception))

        # 2. fail to find any checkpoint - due a fresh output_dir
        output_dir2 = self.get_auto_remove_tmp_dir()
        trainer = get_regression_trainer(output_dir=output_dir2)
        with self.assertRaises(Exception) as context:
            trainer.train(resume_from_checkpoint=True)
        self.assertTrue("No valid checkpoint found in output directory" in str(context.exception))

1366
    def test_resume_training_with_randomness(self):
1367
1368
1369
1370
        # For more than 1 GPUs, since the randomness is introduced in the model and with DataParallel (which is used
        # in this test for more than 2 GPUs), the calls to the torch RNG will happen in a random order (sometimes
        # GPU 0 will call first and sometimes GPU 1).
        random_torch = not torch.cuda.is_available() or torch.cuda.device_count() <= 1
1371
1372
1373
1374
1375
1376

        if torch.cuda.is_available():
            torch.backends.cudnn.deterministic = True
        train_dataset = RegressionDataset(length=128)
        eval_dataset = RegressionDataset()

1377
1378
1379
        with self.subTest("Test every step"):
            config = RegressionModelConfig(a=0, b=2, random_torch=random_torch)
            model = RegressionRandomPreTrainedModel(config)
1380

1381
1382
1383
            tmp_dir = self.get_auto_remove_tmp_dir()
            args = RegressionTrainingArguments(tmp_dir, save_steps=5, learning_rate=0.1)
            trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)
1384

1385
1386
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
1387

1388
1389
1390
1391
1392
            model = RegressionRandomPreTrainedModel(config)
            trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)
            trainer.train(resume_from_checkpoint=os.path.join(tmp_dir, "checkpoint-15"))
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()

1393
1394
            self.assertAlmostEqual(a, a1, delta=1e-5)
            self.assertAlmostEqual(b, b1, delta=1e-5)
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416

        with self.subTest("Test every epoch"):
            config = RegressionModelConfig(a=0, b=2, random_torch=random_torch)
            model = RegressionRandomPreTrainedModel(config)

            tmp_dir = self.get_auto_remove_tmp_dir()
            args = RegressionTrainingArguments(tmp_dir, save_strategy="epoch", learning_rate=0.1)
            trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)

            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()

            model = RegressionRandomPreTrainedModel(config)
            trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)

            checkpoints = [d for d in os.listdir(tmp_dir) if d.startswith("checkpoint-")]
            # There should be one checkpoint per epoch.
            self.assertEqual(len(checkpoints), 3)
            checkpoint_dir = sorted(checkpoints, key=lambda x: int(x.replace("checkpoint-", "")))[0]

            trainer.train(resume_from_checkpoint=os.path.join(tmp_dir, checkpoint_dir))
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
1417

1418
1419
            self.assertAlmostEqual(a, a1, delta=1e-5)
            self.assertAlmostEqual(b, b1, delta=1e-5)
1420

1421
    @slow
Yih-Dar's avatar
Yih-Dar committed
1422
    @require_accelerate
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
    @require_torch_non_multi_gpu
    def test_auto_batch_size_finder(self):
        if torch.cuda.is_available():
            torch.backends.cudnn.deterministic = True

        SRC_DIR = os.path.abspath(
            os.path.join(os.path.dirname(__file__), "..", "..", "examples", "pytorch", "text-classification")
        )
        sys.path.append(SRC_DIR)
        import run_glue

        with tempfile.TemporaryDirectory() as tmpdir:
            testargs = f"""
                run_glue.py
                --model_name_or_path distilbert-base-uncased
                --task_name mrpc
                --do_train
                --do_eval
                --max_seq_len 128
                --per_device_train_batch_size 4096
                --learning_rate 2e-5
                --num_train_epochs 1
                --output_dir {tmpdir}
                --auto_find_batch_size 0
                """.split()
            with self.assertRaises(RuntimeError):
                with patch.object(sys, "argv", testargs):
                    run_glue.main()

        testargs[-1] = "1"
        with patch.object(sys, "argv", testargs):
            run_glue.main()

1456
    # regression for this issue: https://github.com/huggingface/transformers/issues/12970
1457
    def test_training_with_resume_from_checkpoint_false(self):
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
        train_dataset = RegressionDataset(length=128)
        eval_dataset = RegressionDataset()

        config = RegressionModelConfig(a=0, b=2)
        model = RegressionRandomPreTrainedModel(config)

        tmp_dir = self.get_auto_remove_tmp_dir()
        args = RegressionTrainingArguments(tmp_dir, save_steps=5, learning_rate=0.1)
        trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)

        trainer.train(resume_from_checkpoint=False)

1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
    @require_torch_up_to_2_gpus
    def test_resume_training_with_shard_checkpoint(self):
        # This test will fail for more than 2 GPUs since the batch size will get bigger and with the number of
        # save_steps, the checkpoint will resume training at epoch 2 or more (so the data seen by the model
        # won't be the same since the training dataloader is shuffled).

        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(output_dir=tmpdir, train_len=128, save_steps=5, learning_rate=0.1)
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
            state = dataclasses.asdict(trainer.state)

            checkpoint = os.path.join(tmpdir, "checkpoint-5")
            self.convert_to_sharded_checkpoint(checkpoint)

            # Reinitialize trainer
            trainer = get_regression_trainer(output_dir=tmpdir, train_len=128, save_steps=5, learning_rate=0.1)

            trainer.train(resume_from_checkpoint=checkpoint)
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
            state1 = dataclasses.asdict(trainer.state)
            self.assertEqual(a, a1)
            self.assertEqual(b, b1)
            self.check_trainer_state_are_the_same(state, state1)

1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
    @require_safetensors
    @require_torch_up_to_2_gpus
    def test_resume_training_with_safe_checkpoint(self):
        # This test will fail for more than 2 GPUs since the batch size will get bigger and with the number of
        # save_steps, the checkpoint will resume training at epoch 2 or more (so the data seen by the model
        # won't be the same since the training dataloader is shuffled).

        for initial_safe in [False, True]:
            for loaded_safe in [False, True]:
                with tempfile.TemporaryDirectory() as tmpdir:
                    trainer = get_regression_trainer(
                        output_dir=tmpdir,
                        train_len=128,
                        save_steps=5,
                        learning_rate=0.1,
                        save_safetensors=initial_safe,
                    )
                    trainer.train()
                    (a, b) = trainer.model.a.item(), trainer.model.b.item()
                    state = dataclasses.asdict(trainer.state)

                    checkpoint = os.path.join(tmpdir, "checkpoint-5")
                    self.convert_to_sharded_checkpoint(checkpoint, load_safe=initial_safe, save_safe=loaded_safe)

                    # Reinitialize trainer
                    trainer = get_regression_trainer(
                        output_dir=tmpdir, train_len=128, save_steps=5, learning_rate=0.1, save_safetensors=loaded_safe
                    )

                    trainer.train(resume_from_checkpoint=checkpoint)
                    (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
                    state1 = dataclasses.asdict(trainer.state)
                    self.assertEqual(a, a1)
                    self.assertEqual(b, b1)
                    self.check_trainer_state_are_the_same(state, state1)

1531
    @require_torch_up_to_2_gpus
1532
    def test_resume_training_with_gradient_accumulation(self):
1533
1534
1535
1536
        # This test will fail for more than 2 GPUs since the batch size will get bigger and with the number of
        # save_steps, the checkpoint will resume training at epoch 2 or more (so the data seen by the model
        # won't be the same since the training dataloader is shuffled).

1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(
                output_dir=tmpdir,
                train_len=128,
                gradient_accumulation_steps=2,
                per_device_train_batch_size=4,
                save_steps=5,
                learning_rate=0.1,
            )
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
            state = dataclasses.asdict(trainer.state)

            checkpoint = os.path.join(tmpdir, "checkpoint-5")

1552
1553
1554
1555
1556
1557
1558
1559
1560
            # Reinitialize trainer
            trainer = get_regression_trainer(
                output_dir=tmpdir,
                train_len=128,
                gradient_accumulation_steps=2,
                per_device_train_batch_size=4,
                save_steps=5,
                learning_rate=0.1,
            )
1561

1562
            trainer.train(resume_from_checkpoint=checkpoint)
1563
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
1564
1565
1566
1567
1568
            state1 = dataclasses.asdict(trainer.state)
            self.assertEqual(a, a1)
            self.assertEqual(b, b1)
            self.check_trainer_state_are_the_same(state, state1)

1569
    @require_torch_up_to_2_gpus
1570
    def test_resume_training_with_frozen_params(self):
1571
1572
1573
1574
        # This test will fail for more than 2 GPUs since the batch size will get bigger and with the number of
        # save_steps, the checkpoint will resume training at epoch 2 or more (so the data seen by the model
        # won't be the same since the training dataloader is shuffled).

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
        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(
                output_dir=tmpdir,
                train_len=128,
                per_device_train_batch_size=4,
                save_steps=5,
                learning_rate=0.1,
            )
            trainer.model.a.requires_grad_(False)
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
            state = dataclasses.asdict(trainer.state)

            checkpoint = os.path.join(tmpdir, "checkpoint-5")

            # Reinitialize trainer
            trainer = get_regression_trainer(
                output_dir=tmpdir,
                train_len=128,
                per_device_train_batch_size=4,
                save_steps=5,
                learning_rate=0.1,
            )
            trainer.model.a.requires_grad_(False)

            trainer.train(resume_from_checkpoint=checkpoint)

            self.assertFalse(trainer.model.a.requires_grad)
            (a1, b1) = trainer.model.a.item(), trainer.model.b.item()
1604
1605
1606
            state1 = dataclasses.asdict(trainer.state)
            self.assertEqual(a, a1)
            self.assertEqual(b, b1)
1607
            self.check_trainer_state_are_the_same(state, state1)
1608

1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
    def test_load_best_model_at_end(self):
        total = int(self.n_epochs * 64 / self.batch_size)
        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(
                a=1.5,
                b=2.5,
                output_dir=tmpdir,
                learning_rate=0.1,
                eval_steps=5,
                evaluation_strategy="steps",
1619
                save_steps=5,
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
                load_best_model_at_end=True,
            )
            self.assertFalse(trainer.args.greater_is_better)
            trainer.train()
            self.check_saved_checkpoints(tmpdir, 5, total)
            self.check_best_model_has_been_loaded(tmpdir, 5, total, trainer, "eval_loss")

        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(
                a=1.5,
                b=2.5,
                output_dir=tmpdir,
                learning_rate=0.1,
                eval_steps=5,
                evaluation_strategy="steps",
1635
                save_steps=5,
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
                load_best_model_at_end=True,
                metric_for_best_model="accuracy",
                compute_metrics=AlmostAccuracy(),
            )
            self.assertTrue(trainer.args.greater_is_better)
            trainer.train()
            self.check_saved_checkpoints(tmpdir, 5, total)
            self.check_best_model_has_been_loaded(tmpdir, 5, total, trainer, "eval_accuracy", greater_is_better=True)

        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(
                a=1.5,
                b=2.5,
                output_dir=tmpdir,
                learning_rate=0.1,
                evaluation_strategy="epoch",
1652
                save_strategy="epoch",
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
                load_best_model_at_end=True,
                metric_for_best_model="accuracy",
                compute_metrics=AlmostAccuracy(),
            )
            self.assertTrue(trainer.args.greater_is_better)
            trainer.train()
            self.check_saved_checkpoints(tmpdir, 64 // self.batch_size, total)
            self.check_best_model_has_been_loaded(
                tmpdir, 64 // self.batch_size, total, trainer, "eval_accuracy", greater_is_better=True
            )

        # Test this works with a non PreTrainedModel
        with tempfile.TemporaryDirectory() as tmpdir:
            trainer = get_regression_trainer(
                output_dir=tmpdir,
                learning_rate=0.1,
                eval_steps=5,
                evaluation_strategy="steps",
1671
                save_steps=5,
1672
                load_best_model_at_end=True,
1673
                pretrained=False,
1674
1675
1676
1677
1678
1679
            )
            self.assertFalse(trainer.args.greater_is_better)
            trainer.train()
            self.check_saved_checkpoints(tmpdir, 5, total, is_pretrained=False)
            self.check_best_model_has_been_loaded(tmpdir, 5, total, trainer, "eval_loss", is_pretrained=False)

1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
    @require_safetensors
    def test_load_best_model_from_safetensors(self):
        total = int(self.n_epochs * 64 / self.batch_size)
        for save_safetensors, pretrained in product([False, True], [False, True]):
            with tempfile.TemporaryDirectory() as tmpdir:
                trainer = get_regression_trainer(
                    a=1.5,
                    b=2.5,
                    output_dir=tmpdir,
                    learning_rate=0.1,
                    eval_steps=5,
                    evaluation_strategy="steps",
                    save_steps=5,
                    load_best_model_at_end=True,
                    save_safetensors=save_safetensors,
                    pretrained=pretrained,
                )
                self.assertFalse(trainer.args.greater_is_better)
                trainer.train()
                self.check_saved_checkpoints(tmpdir, 5, total, is_pretrained=pretrained, safe_weights=save_safetensors)
                self.check_best_model_has_been_loaded(
                    tmpdir, 5, total, trainer, "eval_loss", is_pretrained=pretrained, safe_weights=save_safetensors
                )

1704
    @slow
Julien Chaumond's avatar
Julien Chaumond committed
1705
1706
1707
1708
1709
    def test_trainer_eval_mrpc(self):
        MODEL_ID = "bert-base-cased-finetuned-mrpc"
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
        data_args = GlueDataTrainingArguments(
1710
            task_name="mrpc", data_dir=f"{get_tests_dir()}/fixtures/tests_samples/MRPC", overwrite_cache=True
Julien Chaumond's avatar
Julien Chaumond committed
1711
        )
1712
        eval_dataset = GlueDataset(data_args, tokenizer=tokenizer, mode="dev")
Julien Chaumond's avatar
Julien Chaumond committed
1713
1714
1715
1716

        training_args = TrainingArguments(output_dir="./examples", no_cuda=True)
        trainer = Trainer(model=model, args=training_args, eval_dataset=eval_dataset)
        result = trainer.evaluate()
1717
        self.assertLess(result["eval_loss"], 0.2)
Julien Chaumond's avatar
Julien Chaumond committed
1718

1719
    @slow
Julien Chaumond's avatar
Julien Chaumond committed
1720
1721
1722
1723
    def test_trainer_eval_lm(self):
        MODEL_ID = "distilroberta-base"
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        dataset = LineByLineTextDataset(
Lysandre's avatar
Lysandre committed
1724
1725
1726
            tokenizer=tokenizer,
            file_path=PATH_SAMPLE_TEXT,
            block_size=tokenizer.max_len_single_sentence,
Julien Chaumond's avatar
Julien Chaumond committed
1727
1728
        )
        self.assertEqual(len(dataset), 31)
1729

1730
    def test_training_iterable_dataset(self):
1731
1732
        config = RegressionModelConfig()
        model = RegressionPreTrainedModel(config)
1733
1734
        # Adding one column not used by the model should have no impact
        train_dataset = SampleIterableDataset(label_names=["labels", "extra"])
1735

1736
        args = RegressionTrainingArguments(output_dir="./examples", max_steps=4)
1737
        trainer = Trainer(model=model, args=args, train_dataset=train_dataset)
1738
        trainer.train()
1739
        self.assertEqual(trainer.state.global_step, 4)
1740

1741
1742
        loader = trainer.get_train_dataloader()
        self.assertIsInstance(loader, torch.utils.data.DataLoader)
1743
1744
        self.assertIsInstance(loader.sampler, torch.utils.data.dataloader._InfiniteConstantSampler)

1745
1746
1747
1748
1749
1750
1751
    def test_training_finite_iterable_dataset(self):
        config = RegressionModelConfig()
        model = RegressionPreTrainedModel(config)

        batch_size = 1
        num_samples = 10

1752
        available_steps = num_samples // batch_size
1753
1754
1755

        data = FiniteIterableDataset(length=num_samples)
        train_args = TrainingArguments(
1756
            "..",
1757
1758
1759
1760
1761
1762
1763
1764
            max_steps=available_steps + 1,  # set a higher number than actually available
            per_device_train_batch_size=batch_size,
        )
        trainer = Trainer(model, train_dataset=data, args=train_args)
        with self.assertLogs("transformers.trainer", level="WARNING") as logs:
            trainer.train()
        self.assertIn(f"stopping training at step {available_steps}!", logs.output[0])

1765
1766
1767
    def test_evaluation_iterable_dataset(self):
        config = RegressionModelConfig(a=1.5, b=2.5)
        model = RegressionPreTrainedModel(config)
1768
1769
        # Adding one column not used by the model should have no impact
        eval_dataset = SampleIterableDataset(label_names=["labels", "extra"])
1770
1771
1772
1773

        args = RegressionTrainingArguments(output_dir="./examples")
        trainer = Trainer(model=model, args=args, eval_dataset=eval_dataset, compute_metrics=AlmostAccuracy())
        results = trainer.evaluate()
1774

1775
1776
1777
1778
1779
1780
        x, y = trainer.eval_dataset.dataset.x, trainer.eval_dataset.dataset.ys[0]
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
1781

1782
1783
1784
        # With a number of elements not a round multiple of the batch size
        eval_dataset = SampleIterableDataset(length=66)
        results = trainer.evaluate(eval_dataset)
1785

1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
        x, y = eval_dataset.dataset.x, eval_dataset.dataset.ys[0]
        pred = 1.5 * x + 2.5
        expected_loss = ((pred - y) ** 2).mean()
        self.assertAlmostEqual(results["eval_loss"], expected_loss)
        expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
        self.assertAlmostEqual(results["eval_accuracy"], expected_acc)

    def test_predict_iterable_dataset(self):
        config = RegressionModelConfig(a=1.5, b=2.5)
        model = RegressionPreTrainedModel(config)
        eval_dataset = SampleIterableDataset()

        args = RegressionTrainingArguments(output_dir="./examples")
        trainer = Trainer(model=model, args=args, eval_dataset=eval_dataset, compute_metrics=AlmostAccuracy())

        preds = trainer.predict(trainer.eval_dataset).predictions
        x = eval_dataset.dataset.x
        self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))

        # With a number of elements not a round multiple of the batch size
1806
1807
        # Adding one column not used by the model should have no impact
        test_dataset = SampleIterableDataset(length=66, label_names=["labels", "extra"])
1808
1809
1810
        preds = trainer.predict(test_dataset).predictions
        x = test_dataset.dataset.x
        self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825

    def test_num_train_epochs_in_training(self):
        # len(train_dl) < gradient_accumulation_steps shouldn't give ``ZeroDivisionError`` when ``max_steps`` is given.
        # It should give 1 update step for each epoch.
        trainer = get_regression_trainer(
            max_steps=3, train_len=64, per_device_train_batch_size=16, gradient_accumulation_steps=5
        )
        train_output = trainer.train()
        self.assertEqual(train_output.global_step, 3)

        # Even ``max_steps`` is not specified, we still expect 1 update step for each epoch if
        # len(train_dl) < gradient_accumulation_steps.
        trainer = get_regression_trainer(train_len=64, per_device_train_batch_size=16, gradient_accumulation_steps=5)
        train_output = trainer.train()
        self.assertEqual(train_output.global_step, int(self.n_epochs))
Marcin Zab艂ocki's avatar
Marcin Zab艂ocki committed
1826

1827
1828
    def test_early_stopping_callback(self):
        # early stopping stops training before num_training_epochs
1829
1830
1831
1832
1833
1834
1835
        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=tmp_dir,
                num_train_epochs=20,
                gradient_accumulation_steps=1,
                per_device_train_batch_size=16,
                load_best_model_at_end=True,
1836
                evaluation_strategy=IntervalStrategy.EPOCH,
1837
                save_strategy=IntervalStrategy.EPOCH,
1838
1839
1840
1841
1842
1843
                compute_metrics=AlmostAccuracy(),
                metric_for_best_model="accuracy",
            )
            trainer.add_callback(EarlyStoppingCallback(1, 0.0001))
            train_output = trainer.train()
            self.assertLess(train_output.global_step, 20 * 64 / 16)
1844
1845

        # Invalid inputs to trainer with early stopping callback result in assertion error
1846
1847
1848
1849
1850
1851
        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=tmp_dir,
                num_train_epochs=20,
                gradient_accumulation_steps=1,
                per_device_train_batch_size=16,
1852
                evaluation_strategy=IntervalStrategy.EPOCH,
1853
1854
1855
1856
                compute_metrics=AlmostAccuracy(),
                metric_for_best_model="accuracy",
            )
            trainer.add_callback(EarlyStoppingCallback(1))
1857
            self.assertEqual(trainer.state.global_step, 0)
1858
1859
1860
1861
            try:
                trainer.train()
            except AssertionError:
                self.assertEqual(trainer.state.global_step, 0)
1862

Marcin Zab艂ocki's avatar
Marcin Zab艂ocki committed
1863
1864
1865
1866
    def test_flos_extraction(self):
        trainer = get_regression_trainer(learning_rate=0.1)

        def assert_flos_extraction(trainer, wrapped_model_to_check):
1867
1868
            self.assertEqual(trainer.model, unwrap_model(wrapped_model_to_check))
            self.assertGreaterEqual(getattr(unwrap_model(wrapped_model_to_check).config, "total_flos", 0), 0)
Marcin Zab艂ocki's avatar
Marcin Zab艂ocki committed
1869
1870
1871
1872
1873

        # with plain model
        assert_flos_extraction(trainer, trainer.model)

        # with enforced DataParallel
1874
        assert_flos_extraction(trainer, nn.DataParallel(trainer.model))
1875

1876
1877
1878
        trainer.train()
        self.assertTrue(isinstance(trainer.state.total_flos, float))

1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
    def check_checkpoint_deletion(self, trainer, output_dir, expected):
        # Make fake checkpoints
        for n in [5, 10, 15, 20, 25]:
            os.makedirs(os.path.join(output_dir, f"{PREFIX_CHECKPOINT_DIR}-{n}"), exist_ok=True)
        trainer._rotate_checkpoints(output_dir=output_dir)
        glob_checkpoints = [str(x) for x in Path(output_dir).glob(f"{PREFIX_CHECKPOINT_DIR}-*")]
        values = [int(re.match(f".*{PREFIX_CHECKPOINT_DIR}-([0-9]+)", d).groups()[0]) for d in glob_checkpoints]
        self.assertSetEqual(set(values), set(expected))

    def test_checkpoint_rotation(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            # Without best model at end
            trainer = get_regression_trainer(output_dir=tmp_dir, save_total_limit=2)
            self.check_checkpoint_deletion(trainer, tmp_dir, [20, 25])

            # With best model at end
1895
1896
1897
            trainer = get_regression_trainer(
                output_dir=tmp_dir, evaluation_strategy="steps", load_best_model_at_end=True, save_total_limit=2
            )
1898
1899
1900
1901
1902
            trainer.state.best_model_checkpoint = os.path.join(tmp_dir, "checkpoint-5")
            self.check_checkpoint_deletion(trainer, tmp_dir, [5, 25])

            # Edge case: we don't always honor save_total_limit=1 if load_best_model_at_end=True to be able to resume
            # from checkpoint
1903
1904
1905
            trainer = get_regression_trainer(
                output_dir=tmp_dir, evaluation_strategy="steps", load_best_model_at_end=True, save_total_limit=1
            )
1906
1907
1908
1909
1910
1911
            trainer.state.best_model_checkpoint = os.path.join(tmp_dir, "checkpoint-25")
            self.check_checkpoint_deletion(trainer, tmp_dir, [25])

            trainer.state.best_model_checkpoint = os.path.join(tmp_dir, "checkpoint-5")
            self.check_checkpoint_deletion(trainer, tmp_dir, [5, 25])

1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
    def check_mem_metrics(self, trainer, check_func):
        metrics = trainer.train().metrics
        check_func("init_mem_cpu_alloc_delta", metrics)
        check_func("train_mem_cpu_alloc_delta", metrics)
        if torch.cuda.device_count() > 0:
            check_func("init_mem_gpu_alloc_delta", metrics)
            check_func("train_mem_gpu_alloc_delta", metrics)

        metrics = trainer.evaluate()
        check_func("eval_mem_cpu_alloc_delta", metrics)
        if torch.cuda.device_count() > 0:
            check_func("eval_mem_gpu_alloc_delta", metrics)

        metrics = trainer.predict(RegressionDataset()).metrics
        check_func("test_mem_cpu_alloc_delta", metrics)
        if torch.cuda.device_count() > 0:
            check_func("test_mem_gpu_alloc_delta", metrics)

    def test_mem_metrics(self):
        # with mem metrics enabled
1932
        trainer = get_regression_trainer(skip_memory_metrics=False)
1933
1934
1935
1936
1937
1938
        self.check_mem_metrics(trainer, self.assertIn)

        # with mem metrics disabled
        trainer = get_regression_trainer(skip_memory_metrics=True)
        self.check_mem_metrics(trainer, self.assertNotIn)

1939
1940
1941
1942
1943
    @require_torch_gpu
    def test_fp16_full_eval(self):
        # this is a sensitive test so let's keep debugging printouts in place for quick diagnosis.
        # it's using pretty large safety margins, but small enough to detect broken functionality.
        debug = 0
1944
        n_gpus = get_gpu_count()
1945
1946

        bs = 8
1947
        eval_len = 16 * n_gpus
1948
1949
1950
1951
1952
        # make the params somewhat big so that there will be enough RAM consumed to be able to
        # measure things. We should get about 64KB for a+b in fp32
        a = torch.ones(1000, bs) + 0.001
        b = torch.ones(1000, bs) - 0.001

1953
        # 1. with fp16_full_eval disabled
1954
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, skip_memory_metrics=False)
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
        metrics = trainer.evaluate()
        del trainer
        gc.collect()

        fp32_init = metrics["init_mem_gpu_alloc_delta"]
        fp32_eval = metrics["eval_mem_gpu_alloc_delta"]

        if debug:
            print(f"fp32_init {fp32_init}")
            print(f"fp32_eval {fp32_eval}")

        # here we expect the model to be preloaded in trainer.__init__ and consume around 64K gpu ram.
        # perfect world: fp32_init == 64<<10
        self.assertGreater(fp32_init, 59_000)
        # after eval should be no extra memory allocated - with a small margin (other than the peak
        # memory consumption for the forward calculation that gets recovered)
        # perfect world: fp32_eval == close to zero
        self.assertLess(fp32_eval, 5_000)

1974
        # 2. with fp16_full_eval enabled
1975
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, fp16_full_eval=True, skip_memory_metrics=False)
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
        metrics = trainer.evaluate()
        fp16_init = metrics["init_mem_gpu_alloc_delta"]
        fp16_eval = metrics["eval_mem_gpu_alloc_delta"]

        if debug:
            print(f"fp16_init {fp16_init}")
            print(f"fp16_eval {fp16_eval}")

        # here we expect the model to not be preloaded in trainer.__init__, so with a small margin it should be close to 0
        # perfect world: fp16_init == close to zero
        self.assertLess(fp16_init, 5_000)
        # here we put the model on device in eval and only `half()` of it, i.e. about 32K,(again we ignore the peak margin which gets returned back)
        # perfect world: fp32_init == 32<<10
        self.assertGreater(fp16_eval, 27_000)

        # 3. relative comparison fp32 vs full fp16
        # should be about half of fp16_init
        # perfect world: fp32_init/2 == fp16_eval
        self.assertAlmostEqual(fp16_eval, fp32_init / 2, delta=5_000)

1996
1997
    @require_torch_non_multi_gpu
    @require_torchdynamo
1998
    @require_torch_tensorrt_fx
1999
    def test_torchdynamo_full_eval(self):
Yih-Dar's avatar
Yih-Dar committed
2000
2001
        import torchdynamo

2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
        # torchdynamo at the moment doesn't support DP/DDP, therefore require a single gpu
        n_gpus = get_gpu_count()

        bs = 8
        eval_len = 16 * n_gpus
        # make the params are somewhat big so that there will be enough RAM consumed to be able to
        # measure things. We should get about 64KB for a+b in fp32
        a = torch.ones(1000, bs) + 0.001
        b = torch.ones(1000, bs) - 0.001

        # 1. Default - without TorchDynamo
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len)
        metrics = trainer.evaluate()
        original_eval_loss = metrics["eval_loss"]
        del trainer

        # 2. TorchDynamo eager
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="eager")
        metrics = trainer.evaluate()
        self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
        del trainer
Yih-Dar's avatar
Yih-Dar committed
2023
        torchdynamo.reset()
2024
2025
2026
2027
2028

        # 3. TorchDynamo nvfuser
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="nvfuser")
        metrics = trainer.evaluate()
        self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
Yih-Dar's avatar
Yih-Dar committed
2029
        torchdynamo.reset()
2030

2031
2032
2033
2034
        # 4. TorchDynamo fx2trt
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="fx2trt")
        metrics = trainer.evaluate()
        self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
Yih-Dar's avatar
Yih-Dar committed
2035
        torchdynamo.reset()
2036

2037
    @unittest.skip("torch 2.0.0 gives `ModuleNotFoundError: No module named 'torchdynamo'`.")
2038
2039
2040
2041
    @require_torch_non_multi_gpu
    @require_torchdynamo
    def test_torchdynamo_memory(self):
        # torchdynamo at the moment doesn't support DP/DDP, therefore require a single gpu
Yih-Dar's avatar
Yih-Dar committed
2042
2043
        import torchdynamo

2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
        class CustomTrainer(Trainer):
            def compute_loss(self, model, inputs, return_outputs=False):
                x = inputs["x"]
                output = model(x)
                if self.args.n_gpu == 1:
                    return output.mean()
                return output

        class MyModule(torch.nn.Module):
            """Simple module that does aggressive fusion"""

            def __init__(self):
                super().__init__()

            def forward(self, x):
                for _ in range(20):
Yih-Dar's avatar
Yih-Dar committed
2060
                    x = torch.cos(x)
2061
2062
2063
2064
                return x

        mod = MyModule()

2065
        # 1. without TorchDynamo (eager baseline)
2066
2067
2068
2069
2070
2071
2072
        a = torch.ones(1024, 1024, device="cuda", requires_grad=True)
        a.grad = None
        trainer = CustomTrainer(model=mod)
        # warmup
        for _ in range(10):
            orig_loss = trainer.training_step(mod, {"x": a})

2073
2074
2075
        # resets
        gc.collect()
        torch.cuda.empty_cache()
2076
        torch.cuda.reset_peak_memory_stats()
2077

2078
2079
        orig_loss = trainer.training_step(mod, {"x": a})
        orig_peak_mem = torch.cuda.max_memory_allocated()
Yih-Dar's avatar
Yih-Dar committed
2080
        torchdynamo.reset()
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
        del trainer

        # 2. TorchDynamo nvfuser
        a = torch.ones(1024, 1024, device="cuda", requires_grad=True)
        a.grad = None
        args = TrainingArguments(output_dir="None", torchdynamo="nvfuser")
        trainer = CustomTrainer(model=mod, args=args)
        # warmup
        for _ in range(10):
            loss = trainer.training_step(mod, {"x": a})

2092
2093
2094
        # resets
        gc.collect()
        torch.cuda.empty_cache()
2095
        torch.cuda.reset_peak_memory_stats()
2096

2097
2098
        loss = trainer.training_step(mod, {"x": a})
        peak_mem = torch.cuda.max_memory_allocated()
Yih-Dar's avatar
Yih-Dar committed
2099
        torchdynamo.reset()
2100
2101
2102
2103
2104
2105
2106
2107
2108
        del trainer

        # Functional check
        self.assertAlmostEqual(loss, orig_loss)

        # AOT Autograd recomputaion and nvfuser recomputation optimization
        # aggressively fuses the operations and reduce the memory footprint.
        self.assertGreater(orig_peak_mem, peak_mem * 2)

2109
    @require_torch_gpu
2110
    @require_torch_bf16_gpu
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
    def test_bf16_full_eval(self):
        # note: most of the logic is the same as test_fp16_full_eval

        # this is a sensitive test so let's keep debugging printouts in place for quick diagnosis.
        # it's using pretty large safety margins, but small enough to detect broken functionality.
        debug = 0
        n_gpus = get_gpu_count()

        bs = 8
        eval_len = 16 * n_gpus
        # make the params somewhat big so that there will be enough RAM consumed to be able to
        # measure things. We should get about 64KB for a+b in fp32
        a = torch.ones(1000, bs) + 0.001
        b = torch.ones(1000, bs) - 0.001

2126
        # 1. with bf16_full_eval disabled
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, skip_memory_metrics=False)
        metrics = trainer.evaluate()
        del trainer
        gc.collect()

        fp32_init = metrics["init_mem_gpu_alloc_delta"]
        fp32_eval = metrics["eval_mem_gpu_alloc_delta"]

        if debug:
            print(f"fp32_init {fp32_init}")
            print(f"fp32_eval {fp32_eval}")

        # here we expect the model to be preloaded in trainer.__init__ and consume around 64K gpu ram.
        # perfect world: fp32_init == 64<<10
        self.assertGreater(fp32_init, 59_000)
        # after eval should be no extra memory allocated - with a small margin (other than the peak
        # memory consumption for the forward calculation that gets recovered)
        # perfect world: fp32_eval == close to zero
        self.assertLess(fp32_eval, 5_000)

2147
        # 2. with bf16_full_eval enabled
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
        trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, bf16_full_eval=True, skip_memory_metrics=False)
        metrics = trainer.evaluate()
        bf16_init = metrics["init_mem_gpu_alloc_delta"]
        bf16_eval = metrics["eval_mem_gpu_alloc_delta"]

        if debug:
            print(f"bf16_init {bf16_init}")
            print(f"bf16_eval {bf16_eval}")

        # here we expect the model to not be preloaded in trainer.__init__, so with a small margin it should be close to 0
        # perfect world: bf16_init == close to zero
        self.assertLess(bf16_init, 5_000)
        # here we put the model on device in eval and only `half()` of it, i.e. about 32K,(again we ignore the peak margin which gets returned back)
        # perfect world: fp32_init == 32<<10
        self.assertGreater(bf16_eval, 27_000)

        # 3. relative comparison fp32 vs full bf16
        # should be about half of bf16_init
        # perfect world: fp32_init/2 == bf16_eval
        self.assertAlmostEqual(bf16_eval, fp32_init / 2, delta=5_000)

2169
    def test_no_wd_param_group(self):
2170
        model = nn.Sequential(TstLayer(128), nn.ModuleList([TstLayer(128), TstLayer(128)]))
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
        trainer = Trainer(model=model)
        trainer.create_optimizer_and_scheduler(10)
        # fmt: off
        wd_names = ['0.linear1.weight', '0.linear2.weight', '1.0.linear1.weight', '1.0.linear2.weight', '1.1.linear1.weight', '1.1.linear2.weight']
        # fmt: on
        wd_params = [p for n, p in model.named_parameters() if n in wd_names]
        no_wd_params = [p for n, p in model.named_parameters() if n not in wd_names]
        self.assertListEqual(trainer.optimizer.param_groups[0]["params"], wd_params)
        self.assertListEqual(trainer.optimizer.param_groups[1]["params"], no_wd_params)

2181

Sylvain Gugger's avatar
Sylvain Gugger committed
2182
2183
2184
2185
2186
@require_torch
@is_staging_test
class TrainerIntegrationWithHubTester(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
2187
2188
        cls._token = TOKEN
        HfFolder.save_token(TOKEN)
Sylvain Gugger's avatar
Sylvain Gugger committed
2189
2190
2191

    @classmethod
    def tearDownClass(cls):
2192
2193
        for model in ["test-trainer", "test-trainer-epoch", "test-trainer-step"]:
            try:
2194
                delete_repo(token=cls._token, repo_id=model)
2195
2196
            except HTTPError:
                pass
Sylvain Gugger's avatar
Sylvain Gugger committed
2197
2198

        try:
2199
            delete_repo(token=cls._token, repo_id="valid_org/test-trainer-org")
Sylvain Gugger's avatar
Sylvain Gugger committed
2200
2201
2202
2203
2204
        except HTTPError:
            pass

    def test_push_to_hub(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
2205
2206
2207
            trainer = get_regression_trainer(
                output_dir=os.path.join(tmp_dir, "test-trainer"),
                push_to_hub=True,
2208
                hub_token=self._token,
2209
2210
            )
            url = trainer.push_to_hub()
Sylvain Gugger's avatar
Sylvain Gugger committed
2211
2212
2213
2214
2215
2216

            # Extract repo_name from the url
            re_search = re.search(ENDPOINT_STAGING + r"/([^/]+/[^/]+)/", url)
            self.assertTrue(re_search is not None)
            repo_name = re_search.groups()[0]

2217
            self.assertEqual(repo_name, f"{USER}/test-trainer")
Sylvain Gugger's avatar
Sylvain Gugger committed
2218
2219
2220
2221
2222
2223
2224
2225
2226

            model = RegressionPreTrainedModel.from_pretrained(repo_name)
            self.assertEqual(model.a.item(), trainer.model.a.item())
            self.assertEqual(model.b.item(), trainer.model.b.item())

    def test_push_to_hub_in_organization(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(output_dir=tmp_dir)
            trainer.save_model()
2227
2228
2229
            trainer = get_regression_trainer(
                output_dir=os.path.join(tmp_dir, "test-trainer-org"),
                push_to_hub=True,
2230
2231
                hub_model_id="valid_org/test-trainer-org",
                hub_token=self._token,
2232
            )
2233
            url = trainer.push_to_hub()
Sylvain Gugger's avatar
Sylvain Gugger committed
2234
2235
2236
2237
2238

            # Extract repo_name from the url
            re_search = re.search(ENDPOINT_STAGING + r"/([^/]+/[^/]+)/", url)
            self.assertTrue(re_search is not None)
            repo_name = re_search.groups()[0]
2239
            self.assertEqual(repo_name, "valid_org/test-trainer-org")
Sylvain Gugger's avatar
Sylvain Gugger committed
2240

2241
            model = RegressionPreTrainedModel.from_pretrained("valid_org/test-trainer-org")
Sylvain Gugger's avatar
Sylvain Gugger committed
2242
2243
2244
            self.assertEqual(model.a.item(), trainer.model.a.item())
            self.assertEqual(model.b.item(), trainer.model.b.item())

2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
    def get_commit_history(self, repo):
        commit_logs = subprocess.run(
            "git log".split(),
            stderr=subprocess.PIPE,
            stdout=subprocess.PIPE,
            check=True,
            encoding="utf-8",
            cwd=repo,
        ).stdout
        commits = commit_logs.split("\n\n")[1::2]
        return [commit.strip() for commit in commits]

    def test_push_to_hub_with_saves_each_epoch(self):
        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=os.path.join(tmp_dir, "test-trainer-epoch"),
                push_to_hub=True,
                hub_token=self._token,
                save_strategy="epoch",
            )
            trainer.train()

2267
2268
2269
2270
            # Wait for the async pushes to be finished
            while trainer.push_in_progress is not None and not trainer.push_in_progress.is_done:
                time.sleep(0.5)

2271
        with tempfile.TemporaryDirectory() as tmp_dir:
2272
            _ = Repository(tmp_dir, clone_from=f"{USER}/test-trainer-epoch", token=self._token)
2273
            commits = self.get_commit_history(tmp_dir)
2274
2275
2276
2277
            self.assertIn("initial commit", commits)
            # We can't test that epoch 2 and 3 are in the commits without being flaky as those might be skipped if
            # the push for epoch 1 wasn't finished at the time.
            self.assertIn("Training in progress, epoch 1", commits)
2278
2279

    def test_push_to_hub_with_saves_each_n_steps(self):
2280
2281
2282
2283
        num_gpus = max(1, get_gpu_count())
        if num_gpus > 2:
            return

2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=os.path.join(tmp_dir, "test-trainer-step"),
                push_to_hub=True,
                hub_token=self._token,
                save_strategy="steps",
                save_steps=5,
            )
            trainer.train()

2294
2295
2296
2297
            # Wait for the async pushes to be finished
            while trainer.push_in_progress is not None and not trainer.push_in_progress.is_done:
                time.sleep(0.5)

2298
        with tempfile.TemporaryDirectory() as tmp_dir:
2299
            _ = Repository(tmp_dir, clone_from=f"{USER}/test-trainer-step", token=self._token)
2300
            commits = self.get_commit_history(tmp_dir)
2301
2302
2303
2304
            self.assertIn("initial commit", commits)
            # We can't test that epoch 2 and 3 are in the commits without being flaky as those might be skipped if
            # the push for epoch 1 wasn't finished at the time.
            self.assertIn("Training in progress, step 5", commits)
2305

Sylvain Gugger's avatar
Sylvain Gugger committed
2306

2307
2308
@require_torch
@require_optuna
2309
class TrainerHyperParameterOptunaIntegrationTest(unittest.TestCase):
2310
    def setUp(self):
2311
        args = TrainingArguments("..")
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
        self.n_epochs = args.num_train_epochs
        self.batch_size = args.train_batch_size

    def test_hyperparameter_search(self):
        class MyTrialShortNamer(TrialShortNamer):
            DEFAULTS = {"a": 0, "b": 0}

        def hp_space(trial):
            return {}

        def model_init(trial):
            if trial is not None:
                a = trial.suggest_int("a", -4, 4)
                b = trial.suggest_int("b", -4, 4)
            else:
                a = 0
                b = 0
            config = RegressionModelConfig(a=a, b=b, double_output=False)

            return RegressionPreTrainedModel(config)

        def hp_name(trial):
            return MyTrialShortNamer.shortname(trial.params)

2336
2337
2338
2339
2340
        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=tmp_dir,
                learning_rate=0.1,
                logging_steps=1,
2341
                evaluation_strategy=IntervalStrategy.EPOCH,
2342
                save_strategy=IntervalStrategy.EPOCH,
2343
2344
2345
2346
2347
2348
2349
2350
                num_train_epochs=4,
                disable_tqdm=True,
                load_best_model_at_end=True,
                logging_dir="runs",
                run_name="test",
                model_init=model_init,
            )
            trainer.hyperparameter_search(direction="minimize", hp_space=hp_space, hp_name=hp_name, n_trials=4)
2351
2352
2353
2354
2355
2356


@require_torch
@require_ray
class TrainerHyperParameterRayIntegrationTest(unittest.TestCase):
    def setUp(self):
2357
        args = TrainingArguments("..")
2358
2359
2360
        self.n_epochs = args.num_train_epochs
        self.batch_size = args.train_batch_size

2361
    def ray_hyperparameter_search(self):
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
        class MyTrialShortNamer(TrialShortNamer):
            DEFAULTS = {"a": 0, "b": 0}

        def hp_space(trial):
            from ray import tune

            return {
                "a": tune.randint(-4, 4),
                "b": tune.randint(-4, 4),
            }

        def model_init(config):
2374
2375
2376
2377
2378
2379
2380
            if config is None:
                a = 0
                b = 0
            else:
                a = config["a"]
                b = config["b"]
            model_config = RegressionModelConfig(a=a, b=b, double_output=False)
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391

            return RegressionPreTrainedModel(model_config)

        def hp_name(params):
            return MyTrialShortNamer.shortname(params)

        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=tmp_dir,
                learning_rate=0.1,
                logging_steps=1,
2392
                evaluation_strategy=IntervalStrategy.EPOCH,
2393
                save_strategy=IntervalStrategy.EPOCH,
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
                num_train_epochs=4,
                disable_tqdm=True,
                load_best_model_at_end=True,
                logging_dir="runs",
                run_name="test",
                model_init=model_init,
            )
            trainer.hyperparameter_search(
                direction="minimize", hp_space=hp_space, hp_name=hp_name, backend="ray", n_trials=4
            )
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414

    def test_hyperparameter_search(self):
        self.ray_hyperparameter_search()

    def test_hyperparameter_search_ray_client(self):
        import ray
        from ray.util.client.ray_client_helpers import ray_start_client_server

        with ray_start_client_server():
            assert ray.util.client.ray.is_connected()
            self.ray_hyperparameter_search()
2415
2416


2417
@slow
2418
2419
2420
2421
@require_torch
@require_sigopt
class TrainerHyperParameterSigOptIntegrationTest(unittest.TestCase):
    def setUp(self):
2422
        args = TrainingArguments("..")
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
        self.n_epochs = args.num_train_epochs
        self.batch_size = args.train_batch_size

    def test_hyperparameter_search(self):
        class MyTrialShortNamer(TrialShortNamer):
            DEFAULTS = {"a": 0, "b": 0}

        def hp_space(trial):
            return [
                {"bounds": {"min": -4, "max": 4}, "name": "a", "type": "int"},
                {"bounds": {"min": -4, "max": 4}, "name": "b", "type": "int"},
            ]

        def model_init(trial):
            if trial is not None:
                a = trial.assignments["a"]
                b = trial.assignments["b"]
            else:
                a = 0
                b = 0
            config = RegressionModelConfig(a=a, b=b, double_output=False)

            return RegressionPreTrainedModel(config)

        def hp_name(trial):
            return MyTrialShortNamer.shortname(trial.assignments)

        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=tmp_dir,
                learning_rate=0.1,
                logging_steps=1,
                evaluation_strategy=IntervalStrategy.EPOCH,
                save_strategy=IntervalStrategy.EPOCH,
                num_train_epochs=4,
                disable_tqdm=True,
                load_best_model_at_end=True,
                logging_dir="runs",
                run_name="test",
                model_init=model_init,
            )
            trainer.hyperparameter_search(
                direction="minimize", hp_space=hp_space, hp_name=hp_name, backend="sigopt", n_trials=4
            )
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476


optim_test_params = []
if is_torch_available():
    default_adam_kwargs = {
        "betas": (TrainingArguments.adam_beta1, TrainingArguments.adam_beta2),
        "eps": TrainingArguments.adam_epsilon,
        "lr": TrainingArguments.learning_rate,
    }

2477
2478
2479
2480
2481
    default_lion_kwargs = {
        "betas": (TrainingArguments.adam_beta1, TrainingArguments.adam_beta2),
        "lr": TrainingArguments.learning_rate,
    }

2482
2483
2484
2485
2486
2487
2488
    default_anyprecision_kwargs = {
        "use_kahan_summation": False,
        "momentum_dtype": torch.float32,
        "variance_dtype": torch.float32,
        "compensation_buffer_dtype": torch.bfloat16,
    }

2489
2490
    optim_test_params = [
        (
2491
            TrainingArguments(optim=OptimizerNames.ADAMW_HF, output_dir="None"),
2492
2493
2494
2495
            transformers.optimization.AdamW,
            default_adam_kwargs,
        ),
        (
2496
            TrainingArguments(optim=OptimizerNames.ADAMW_HF.value, output_dir="None"),
2497
2498
2499
2500
            transformers.optimization.AdamW,
            default_adam_kwargs,
        ),
        (
2501
            TrainingArguments(optim=OptimizerNames.ADAMW_TORCH, output_dir="None"),
2502
2503
2504
2505
            torch.optim.AdamW,
            default_adam_kwargs,
        ),
        (
2506
            TrainingArguments(optim=OptimizerNames.ADAFACTOR, output_dir="None"),
2507
2508
2509
2510
2511
2512
2513
2514
            transformers.optimization.Adafactor,
            {
                "scale_parameter": False,
                "relative_step": False,
                "lr": TrainingArguments.learning_rate,
            },
        ),
    ]
2515

2516
2517
2518
2519
2520
    if is_apex_available():
        import apex

        optim_test_params.append(
            (
2521
                TrainingArguments(optim=OptimizerNames.ADAMW_APEX_FUSED, output_dir="None"),
2522
2523
2524
2525
2526
                apex.optimizers.FusedAdam,
                default_adam_kwargs,
            )
        )

2527
2528
2529
2530
2531
    if is_bitsandbytes_available():
        import bitsandbytes as bnb

        optim_test_params.append(
            (
2532
                TrainingArguments(optim=OptimizerNames.ADAMW_BNB, output_dir="None"),
2533
                bnb.optim.AdamW,
2534
2535
2536
2537
                default_adam_kwargs,
            )
        )

2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
        optim_test_params.append(
            (
                TrainingArguments(optim=OptimizerNames.ADAMW_8BIT, output_dir="None"),
                bnb.optim.AdamW,
                default_adam_kwargs,
            )
        )

        optim_test_params.append(
            (
                TrainingArguments(optim=OptimizerNames.PAGED_ADAMW, output_dir="None"),
                bnb.optim.AdamW,
                default_adam_kwargs,
            )
        )

        optim_test_params.append(
            (
                TrainingArguments(optim=OptimizerNames.PAGED_ADAMW_8BIT, output_dir="None"),
                bnb.optim.AdamW,
                default_adam_kwargs,
            )
        )

        optim_test_params.append(
            (
                TrainingArguments(optim=OptimizerNames.LION, output_dir="None"),
                bnb.optim.Lion,
                default_lion_kwargs,
            )
        )

        optim_test_params.append(
            (
                TrainingArguments(optim=OptimizerNames.LION_8BIT, output_dir="None"),
                bnb.optim.Lion,
                default_lion_kwargs,
            )
        )

        optim_test_params.append(
            (
                TrainingArguments(optim=OptimizerNames.PAGED_LION_8BIT, output_dir="None"),
                bnb.optim.Lion,
                default_lion_kwargs,
            )
        )

2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
    if is_torchdistx_available():
        import torchdistx

        optim_test_params.append(
            (
                TrainingArguments(optim=OptimizerNames.ADAMW_ANYPRECISION, output_dir="None"),
                torchdistx.optimizers.AnyPrecisionAdamW,
                dict(default_adam_kwargs, **default_anyprecision_kwargs),
            )
        )

2597
2598
2599

@require_torch
class TrainerOptimizerChoiceTest(unittest.TestCase):
2600
2601
    def check_optim_and_kwargs(self, training_args: TrainingArguments, expected_cls, expected_kwargs):
        actual_cls, optim_kwargs = Trainer.get_optimizer_cls_and_kwargs(training_args)
2602
2603
2604
        self.assertEqual(expected_cls, actual_cls)
        self.assertIsNotNone(optim_kwargs)

2605
        for p, v in expected_kwargs.items():
2606
2607
2608
2609
2610
            self.assertTrue(p in optim_kwargs)
            actual_v = optim_kwargs[p]
            self.assertTrue(actual_v == v, f"Failed check for {p}. Expected {v}, but got {actual_v}.")

    @parameterized.expand(optim_test_params, skip_on_empty=True)
2611
    def test_optim_supported(self, training_args: TrainingArguments, expected_cls, expected_kwargs):
2612
        # exercises all the valid --optim options
2613
        self.check_optim_and_kwargs(training_args, expected_cls, expected_kwargs)
2614

2615
        trainer = get_regression_trainer(**training_args.to_dict())
2616
2617
2618
2619
        trainer.train()

    def test_fused_adam(self):
        # Pretend that apex is installed and mock apex.optimizers.FusedAdam exists.
2620
2621
        # Trainer.get_optimizer_cls_and_kwargs does not use FusedAdam. It only has to return the
        # class given, so mocking apex.optimizers.FusedAdam should be fine for testing and allow
2622
2623
2624
2625
2626
2627
2628
2629
2630
        # the test to run without requiring an apex installation.
        mock = Mock()
        modules = {
            "apex": mock,
            "apex.optimizers": mock.optimizers,
            "apex.optimizers.FusedAdam": mock.optimizers.FusedAdam,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
2631
                TrainingArguments(optim=OptimizerNames.ADAMW_APEX_FUSED, output_dir="None"),
2632
                mock.optimizers.FusedAdam,
2633
                default_adam_kwargs,
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
            )

    def test_fused_adam_no_apex(self):
        args = TrainingArguments(optim=OptimizerNames.ADAMW_APEX_FUSED, output_dir="None")

        # Pretend that apex does not exist, even if installed. By setting apex to None, importing
        # apex will fail even if apex is installed.
        with patch.dict("sys.modules", {"apex.optimizers": None}):
            with self.assertRaises(ValueError):
                Trainer.get_optimizer_cls_and_kwargs(args)
2644

2645
2646
2647
2648
2649
2650
2651
2652
2653
    def test_bnb_adam8bit(self):
        # Pretend that Bits and Bytes is installed and mock bnb.optim.Adam8bit exists.
        # Trainer.get_optimizer_cls_and_kwargs does not use Adam8bit. It only has to return the
        # class given, so mocking bnb.optim.Adam8bit should be fine for testing and allow
        # the test to run without requiring a bnb installation.
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
2654
            "bitsandbytes.optim.AdamW": mock.optim.AdamW,
2655
2656
2657
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
2658
                TrainingArguments(optim=OptimizerNames.ADAMW_BNB, output_dir="None"),
2659
                mock.optim.AdamW,
2660
                default_adam_kwargs,
2661
2662
            )

2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
    def test_bnb_paged_adam8bit_alias(self):
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
            "bitsandbytes.optim.AdamW": mock.optim.AdamW,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.ADAMW_8BIT, output_dir="None"),
                mock.optim.AdamW,
                default_adam_kwargs,
            )

    def test_bnb_paged_adam(self):
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
            "bitsandbytes.optim.AdamW": mock.optim.AdamW,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.PAGED_ADAMW, output_dir="None"),
                mock.optim.AdamW,
                default_adam_kwargs,
            )

    def test_bnb_paged_adam8bit(self):
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
            "bitsandbytes.optim.AdamW": mock.optim.AdamW,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.PAGED_ADAMW_8BIT, output_dir="None"),
                mock.optim.AdamW,
                default_adam_kwargs,
            )

    def test_bnb_lion(self):
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
            "bitsandbytes.optim.Lion": mock.optim.Lion,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.LION, output_dir="None"),
                mock.optim.Lion,
                default_lion_kwargs,
            )

    def test_bnb_lion8bit(self):
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
            "bitsandbytes.optim.Lion": mock.optim.Lion,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.LION_8BIT, output_dir="None"),
                mock.optim.Lion,
                default_lion_kwargs,
            )

    def test_bnb_paged_lion8bit(self):
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
            "bitsandbytes.optim.Lion": mock.optim.Lion,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.PAGED_LION_8BIT, output_dir="None"),
                mock.optim.Lion,
                default_lion_kwargs,
            )

    def test_bnb_paged_lion(self):
        mock = Mock()
        modules = {
            "bitsandbytes": mock,
            "bitsandbytes.optim": mock.optim,
            "bitsandbytes.optim.Lion": mock.optim.Lion,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.PAGED_LION, output_dir="None"),
                mock.optim.Lion,
                default_lion_kwargs,
            )

2761
2762
2763
2764
2765
    def test_bnb_adam8bit_no_bnb(self):
        args = TrainingArguments(optim=OptimizerNames.ADAMW_BNB, output_dir="None")

        # Pretend that bnb does not exist, even if installed. By setting bnb to None, importing
        # bnb will fail even if bnb is installed.
Younes Belkada's avatar
Younes Belkada committed
2766
        with patch.dict("sys.modules", {"bitsandbytes.optim": None}):
2767
2768
2769
            with self.assertRaises(ValueError):
                Trainer.get_optimizer_cls_and_kwargs(args)

2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
    def test_bnb_paged_adam_no_bnb(self):
        args = TrainingArguments(optim=OptimizerNames.PAGED_ADAMW, output_dir="None")

        # Pretend that bnb does not exist, even if installed. By setting bnb to None, importing
        # bnb will fail even if bnb is installed.
        with patch.dict("sys.modules", {"bitsandbytes.optim": None}):
            with self.assertRaises(ValueError):
                Trainer.get_optimizer_cls_and_kwargs(args)

    def test_bnb_paged_adam8bit_no_bnb(self):
        args = TrainingArguments(optim=OptimizerNames.PAGED_ADAMW_8BIT, output_dir="None")

        # Pretend that bnb does not exist, even if installed. By setting bnb to None, importing
        # bnb will fail even if bnb is installed.
        with patch.dict("sys.modules", {"bitsandbytes.optim": None}):
            with self.assertRaises(ValueError):
                Trainer.get_optimizer_cls_and_kwargs(args)

    def test_bnb_paged_lion_no_bnb(self):
        args = TrainingArguments(optim=OptimizerNames.PAGED_LION, output_dir="None")

        # Pretend that bnb does not exist, even if installed. By setting bnb to None, importing
        # bnb will fail even if bnb is installed.
        with patch.dict("sys.modules", {"bitsandbytes.optim": None}):
            with self.assertRaises(ValueError):
                Trainer.get_optimizer_cls_and_kwargs(args)

    def test_bnb_paged_lion8bit_no_bnb(self):
        args = TrainingArguments(optim=OptimizerNames.PAGED_LION_8BIT, output_dir="None")

        # Pretend that bnb does not exist, even if installed. By setting bnb to None, importing
        # bnb will fail even if bnb is installed.
        with patch.dict("sys.modules", {"bitsandbytes.optim": None}):
            with self.assertRaises(ValueError):
                Trainer.get_optimizer_cls_and_kwargs(args)

2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
    def test_anyprecision_adamw(self):
        # Pretend that torchdistx is installed and mock torchdistx.optimizers.AnyPrecisionAdamW exists.
        # Trainer.get_optimizer_cls_and_kwargs does not use AnyPrecisioinAdamW. It only has to return the
        # class given, so mocking torchdistx.optimizers.AnyPrecisionAdamW should be fine for testing and allow
        # the test to run without requiring a bnb installation.
        mock = Mock()
        modules = {
            "torchdistx": mock,
            "torchdistx.optimizers": mock.optimizers,
            "torchdistx.optimizers.AnyPrecisionAdamW.": mock.optimizers.AnyPrecisionAdamW,
        }
        with patch.dict("sys.modules", modules):
            self.check_optim_and_kwargs(
                TrainingArguments(optim=OptimizerNames.ADAMW_ANYPRECISION, output_dir="None"),
                mock.optimizers.AnyPrecisionAdamW,
                dict(default_adam_kwargs, **default_anyprecision_kwargs),
            )

    def test_no_torchdistx_anyprecision_adamw(self):
        args = TrainingArguments(optim=OptimizerNames.ADAMW_ANYPRECISION, output_dir="None")

        # Pretend that torchdistx does not exist, even if installed. By setting torchdistx to None, importing
        # torchdistx.optimizers will fail even if torchdistx is installed.
        with patch.dict("sys.modules", {"torchdistx.optimizers": None}):
            with self.assertRaises(ValueError):
                Trainer.get_optimizer_cls_and_kwargs(args)

2833
2834
2835
2836
2837

@require_torch
@require_wandb
class TrainerHyperParameterWandbIntegrationTest(unittest.TestCase):
    def setUp(self):
2838
        args = TrainingArguments("..")
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
        self.n_epochs = args.num_train_epochs
        self.batch_size = args.train_batch_size

    def test_hyperparameter_search(self):
        class MyTrialShortNamer(TrialShortNamer):
            DEFAULTS = {"a": 0, "b": 0}

        def hp_space(trial):
            return {
                "method": "random",
                "metric": {},
                "parameters": {
                    "a": {"distribution": "uniform", "min": 1e-6, "max": 1e-4},
                    "b": {"distribution": "int_uniform", "min": 1, "max": 6},
                },
            }

        def model_init(config):
            if config is None:
                a = 0
                b = 0
            else:
                a = config["a"]
                b = config["b"]
            model_config = RegressionModelConfig(a=a, b=b, double_output=False)

            return RegressionPreTrainedModel(model_config)

        def hp_name(params):
            return MyTrialShortNamer.shortname(params)

        with tempfile.TemporaryDirectory() as tmp_dir:
            trainer = get_regression_trainer(
                output_dir=tmp_dir,
                learning_rate=0.1,
                logging_steps=1,
                evaluation_strategy=IntervalStrategy.EPOCH,
                save_strategy=IntervalStrategy.EPOCH,
                num_train_epochs=4,
                disable_tqdm=True,
                load_best_model_at_end=True,
                logging_dir="runs",
                run_name="test",
                model_init=model_init,
            )
            trainer.hyperparameter_search(
                direction="minimize", hp_space=hp_space, hp_name=hp_name, backend="wandb", n_trials=4, anonymous="must"
            )