test_deepspeed.py 47.9 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# Copyright 2020 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15
import dataclasses
16
import io
17
import itertools
18
import json
19
20
import os
import unittest
21
from copy import deepcopy
22
from functools import partial
23

24
import datasets
25
from parameterized import parameterized
26

27
import tests.trainer.test_trainer
Stas Bekman's avatar
Stas Bekman committed
28
from tests.trainer.test_trainer import TrainerIntegrationCommon  # noqa
29
from transformers import AutoModel, TrainingArguments, is_torch_available, logging
30
31
32
33
34
from transformers.integrations.deepspeed import (
    HfDeepSpeedConfig,
    is_deepspeed_available,
    unset_hf_deepspeed_config,
)
35
from transformers.testing_utils import (
36
    CaptureLogger,
37
    CaptureStd,
38
    CaptureStderr,
39
    LoggingLevel,
40
41
42
    TestCasePlus,
    execute_subprocess_async,
    get_gpu_count,
43
    mockenv_context,
44
    require_deepspeed,
45
    require_optuna,
46
47
48
49
    require_torch_gpu,
    require_torch_multi_gpu,
    slow,
)
50
from transformers.trainer_utils import get_last_checkpoint, set_seed
51
from transformers.utils import SAFE_WEIGHTS_NAME, is_torch_bf16_gpu_available
52

53

54
if is_torch_available():
Stas Bekman's avatar
Stas Bekman committed
55
56
57
58
    from tests.trainer.test_trainer import (  # noqa
        RegressionModelConfig,
        RegressionPreTrainedModel,
    )
59

60
61
62
    # hack to restore original logging level pre #21700
    get_regression_trainer = partial(tests.trainer.test_trainer.get_regression_trainer, log_level="info")

63

64
set_seed(42)
65

66
67
68
# default torch.distributed port
DEFAULT_MASTER_PORT = "10999"

69
T5_SMALL = "t5-small"
70
T5_TINY = "patrickvonplaten/t5-tiny-random"
71
GPT2_TINY = "sshleifer/tiny-gpt2"
72
73


74
75
76
77
78
def load_json(path):
    with open(path) as f:
        return json.load(f)


Stas Bekman's avatar
Stas Bekman committed
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
def get_master_port(real_launcher=False):
    """
    When using a single gpu launcher emulation (i.e. not deepspeed or python -m torch.distributed)
    the issue is that once the port is tied it can't be used anywhere else outside of this process,
    since torch.dist doesn't free the port until the process exits. Therefore for the sake of being
    able to run both emulated launcher and normal launcher tests we need 2 distinct ports.

    This function will give the right port in the right context. For real launcher it'll give the
    base port, for emulated launcher it'll give the base port + 1. In both cases a string is
    returned.

    Args:
        `real_launcher`: whether a real launcher is going to be used, or the emulated one

    """

    master_port_base = os.environ.get("DS_TEST_PORT", DEFAULT_MASTER_PORT)
    if not real_launcher:
        master_port_base = str(int(master_port_base) + 1)
    return master_port_base


101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
def require_deepspeed_aio(test_case):
    """
    Decorator marking a test that requires deepspeed aio (nvme)
    """
    if not is_deepspeed_available():
        return unittest.skip("test requires deepspeed")(test_case)

    import deepspeed
    from deepspeed.ops.aio import AsyncIOBuilder

    if not deepspeed.ops.__compatible_ops__[AsyncIOBuilder.NAME]:
        return unittest.skip("test requires deepspeed async-io")(test_case)
    else:
        return test_case


117
118
if is_deepspeed_available():
    from deepspeed.utils import logger as deepspeed_logger  # noqa
119
    from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
120
    from transformers.integrations.deepspeed import deepspeed_config, is_deepspeed_zero3_enabled  # noqa
121

122
123
124
125
126
127
128

def get_launcher(distributed=False):
    # 1. explicitly set --num_nodes=1 just in case these tests end up run on a multi-node setup
    # - it won't be able to handle that
    # 2. for now testing with just 2 gpus max (since some quality tests may give different
    # results with mode gpus because we use very little data)
    num_gpus = min(2, get_gpu_count()) if distributed else 1
Stas Bekman's avatar
Stas Bekman committed
129
    master_port = get_master_port(real_launcher=True)
130
    return f"deepspeed --num_nodes 1 --num_gpus {num_gpus} --master_port {master_port}".split()
131
132


133
134
ZERO2 = "zero2"
ZERO3 = "zero3"
135
136
137
138

FP16 = "fp16"
BF16 = "bf16"

139
140
141
142
143
144
145
146
HF_OPTIM = "hf_optim"
HF_SCHEDULER = "hf_scheduler"
DS_OPTIM = "ds_optim"
DS_SCHEDULER = "ds_scheduler"

optims = [HF_OPTIM, DS_OPTIM]
schedulers = [HF_SCHEDULER, DS_SCHEDULER]

147
stages = [ZERO2, ZERO3]
148
if is_torch_bf16_gpu_available():
149
150
151
152
153
154
155
156
157
158
159
160
161
162
    dtypes = [FP16, BF16]
else:
    dtypes = [FP16]


def parameterized_custom_name_func(func, param_num, param):
    # customize the test name generator function as we want both params to appear in the sub-test
    # name, as by default it shows only the first param
    param_based_name = parameterized.to_safe_name("_".join(str(x) for x in param.args))
    return f"{func.__name__}_{param_based_name}"


# Cartesian-product of zero stages with models to test
params = list(itertools.product(stages, dtypes))
163

164
165
params_with_optims_and_schedulers = list(itertools.product(stages, dtypes, optims, schedulers))

166

167
168
169
170
171
172
173
174
175
176
@require_deepspeed
@require_torch_gpu
class CoreIntegrationDeepSpeed(TestCasePlus, TrainerIntegrationCommon):
    """
    Testing non-Trainer DeepSpeed integration
    """

    def setUp(self):
        super().setUp()

Stas Bekman's avatar
Stas Bekman committed
177
        master_port = get_master_port(real_launcher=False)
178
179
180
181
182
183
184
        self.dist_env_1_gpu = {
            "MASTER_ADDR": "localhost",
            "MASTER_PORT": master_port,
            "RANK": "0",
            "LOCAL_RANK": "0",
            "WORLD_SIZE": "1",
        }
185

186
187
188
189
190
191
    def tearDown(self):
        super().tearDown()

        # reset the ds config global so that tests state doesn't leak
        unset_hf_deepspeed_config()

192
193
    def test_init_zero3_fp16(self):
        # test that zero.Init() works correctly under zero3/fp16
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
        ds_config = {
            "train_batch_size": 1,
            "zero_optimization": {
                "stage": 3,
            },
        }

        dschf = HfDeepSpeedConfig(ds_config)

        self.assertTrue(dschf.is_zero3())
        self.assertTrue(is_deepspeed_zero3_enabled())

        with LoggingLevel(logging.INFO):
            with mockenv_context(**self.dist_env_1_gpu):
                logger = logging.get_logger("transformers.modeling_utils")
                with CaptureLogger(logger) as cl:
                    AutoModel.from_pretrained(T5_TINY)
        self.assertIn("Detected DeepSpeed ZeRO-3", cl.out)

        # now remove zero optimization
        del ds_config["zero_optimization"]
        dschf = HfDeepSpeedConfig(ds_config)

        self.assertFalse(dschf.is_zero3())
        self.assertFalse(is_deepspeed_zero3_enabled())

        with LoggingLevel(logging.INFO):
            with mockenv_context(**self.dist_env_1_gpu):
                logger = logging.get_logger("transformers.modeling_utils")
                with CaptureLogger(logger) as cl:
                    AutoModel.from_pretrained(T5_TINY)
        self.assertNotIn("Detected DeepSpeed ZeRO-3", cl.out)


228
class TrainerIntegrationDeepSpeedWithCustomConfig(TestCasePlus):
229
230
    def setUp(self):
        super().setUp()
231
232
233
234
235

        args = TrainingArguments(".")
        self.n_epochs = args.num_train_epochs
        self.batch_size = args.train_batch_size

Stas Bekman's avatar
Stas Bekman committed
236
        master_port = get_master_port(real_launcher=False)
237
238
239
240
241
242
243
        self.dist_env_1_gpu = {
            "MASTER_ADDR": "localhost",
            "MASTER_PORT": master_port,
            "RANK": "0",
            "LOCAL_RANK": "0",
            "WORLD_SIZE": "1",
        }
244

245
246
247
248
        self.ds_config_file = {
            "zero2": f"{self.test_file_dir_str}/ds_config_zero2.json",
            "zero3": f"{self.test_file_dir_str}/ds_config_zero3.json",
        }
249
250
251

        # use self.get_config_dict(stage) to use these to ensure the original is not modified
        with io.open(self.ds_config_file[ZERO2], "r", encoding="utf-8") as f:
252
            config_zero2 = json.load(f)
253
        with io.open(self.ds_config_file[ZERO3], "r", encoding="utf-8") as f:
254
            config_zero3 = json.load(f)
255
            # The following setting slows things down, so don't enable it by default unless needed by a test.
256
            # It's in the file as a demo for users since we want everything to work out of the box even if slower.
257
258
            config_zero3["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = False

259
260
261
262
        self.ds_config_dict = {
            "zero2": config_zero2,
            "zero3": config_zero3,
        }
263

264
265
266
267
268
269
    def tearDown(self):
        super().tearDown()

        # reset the ds config global so that tests state doesn't leak
        unset_hf_deepspeed_config()

270
271
272
    def get_config_dict(self, stage):
        # As some tests modify the dict, always make a copy
        return deepcopy(self.ds_config_dict[stage])
273

274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296

@require_deepspeed
@require_torch_gpu
class TrainerIntegrationDeepSpeed(TrainerIntegrationDeepSpeedWithCustomConfig, TrainerIntegrationCommon):
    """

    This class is for testing directly via get_regression_trainer

    It mixes in `TrainerIntegrationCommon` which already has a lot of helper validation methods
    which we can re-use here.

    Important: this class' setup can only work with a single gpu because it runs within the current
    pytest worker. For multi-gpu tests use TestDeepSpeedWithLauncher.

    Note: if any of the tests of this class get run there will be at least one gpu occupied by them
    until this pytest worker exits. This is because the gpu memory allocated by the cuda-kernels
    won't be released until this pytest worker exits.

    This may appear as some run-away tests if you watch `nvidia-smi` while other tests that fork new
    processes are run. So there will be one or two "stale" processes reported in `nvidia-smi`. This
    is not a bug.
    """

297
    # --- These tests are enough to run on one of zero stages --- #
298

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
    def test_hf_ds_config_mismatch(self):
        ds_config = self.get_config_dict(ZERO2)

        # Purposefully configure these values to mismatch TrainingArguments values.
        # This currently doesn't cover all keys (but it could)
        per_device_train_batch_size = 2
        ds_config["train_micro_batch_size_per_gpu"] = per_device_train_batch_size + 2

        ds_config["train_batch_size"] = 1000

        gradient_accumulation_steps = 2
        ds_config["gradient_accumulation_steps"] = gradient_accumulation_steps + 2

        max_grad_norm = 1.0
        ds_config["gradient_clipping"] = max_grad_norm + 0.1

        adam_beta1, adam_beta2 = 0.9, 0.99
        ds_config["optimizer"]["params"]["betas"] = [adam_beta1 - 0.1, adam_beta2 - 0.1]

        fp16 = True
        ds_config["fp16"]["enabled"] = not fp16

        keys = [
            "per_device_train_batch_size",
            "train_batch_size",
            "gradient_accumulation_steps",
            "max_grad_norm",
            "betas",
            "fp16",
        ]

        with mockenv_context(**self.dist_env_1_gpu):
            trainer = get_regression_trainer(
                local_rank=0,
                fp16=fp16,
                deepspeed=ds_config,
                per_device_train_batch_size=per_device_train_batch_size,
                gradient_accumulation_steps=gradient_accumulation_steps,
                max_grad_norm=max_grad_norm,
                adam_beta1=adam_beta1,
                adam_beta2=adam_beta2,
            )
            with self.assertRaises(Exception) as context:
                trainer.train()

        for key in keys:
            self.assertTrue(
                key in str(context.exception),
                f"{key} is not in the exception message:\n{context.exception}",
            )

350
351
352
353
354
355
356
357
358
    # Test various combos
    # 1. DS scheduler + DS optimizer: this is already tested by most other tests
    # 2. HF scheduler + HF optimizer:
    # 3. DS scheduler + HF optimizer:
    # 4. HF scheduler + DS optimizer:

    def test_hf_scheduler_hf_optimizer(self):
        a = 0
        with mockenv_context(**self.dist_env_1_gpu):
359
360
361
            ds_config_zero2_dict = self.get_config_dict(ZERO2)
            del ds_config_zero2_dict["optimizer"]  # force default HF Trainer optimizer
            del ds_config_zero2_dict["scheduler"]  # force default HF Trainer scheduler
362
            ds_config_zero2_dict["zero_optimization"]["offload_optimizer"]["device"] = "none"
363
            ds_config_zero2_dict["fp16"]["initial_scale_power"] = 1  # force optimizer on the first step
364
            trainer = get_regression_trainer(a=a, local_rank=0, fp16=True, deepspeed=ds_config_zero2_dict)
365
366
367
368
369
370
371
            trainer.train()
        new_a = trainer.model.a.item()
        self.assertNotEqual(new_a, a)

    def test_ds_scheduler_hf_optimizer(self):
        a = 0
        with mockenv_context(**self.dist_env_1_gpu):
372
373
            ds_config_zero2_dict = self.get_config_dict(ZERO2)
            del ds_config_zero2_dict["optimizer"]  # force default HF Trainer optimizer
374
            ds_config_zero2_dict["zero_optimization"]["offload_optimizer"]["device"] = "none"
375
            ds_config_zero2_dict["fp16"]["initial_scale_power"] = 1  # force optimizer on the first step
376
            trainer = get_regression_trainer(a=a, local_rank=0, fp16=True, deepspeed=ds_config_zero2_dict)
377
378
379
380
381
            trainer.train()
        new_a = trainer.model.a.item()
        self.assertNotEqual(new_a, a)

    def test_hf_scheduler_ds_optimizer(self):
382
        a = 0
383
        with mockenv_context(**self.dist_env_1_gpu):
384
385
            ds_config_zero2_dict = self.get_config_dict(ZERO2)
            del ds_config_zero2_dict["scheduler"]  # force default HF Trainer scheduler
386
            ds_config_zero2_dict["zero_optimization"]["offload_optimizer"]["device"] = "none"
387
            ds_config_zero2_dict["fp16"]["initial_scale_power"] = 1  # force optimizer on the first step
388
389
390
391
            trainer = get_regression_trainer(a=a, local_rank=0, fp16=True, deepspeed=ds_config_zero2_dict)
            trainer.train()
        new_a = trainer.model.a.item()
        self.assertNotEqual(new_a, a)
392

393
    @require_deepspeed_aio
394
395
396
397
398
    def test_stage3_nvme_offload(self):
        with mockenv_context(**self.dist_env_1_gpu):
            # this actually doesn't have to be on NVMe, any storage will do since this test only
            # runs a simple check that we can use some directory as if it were NVMe
            nvme_path = self.get_auto_remove_tmp_dir()
399
            nvme_config = {"device": "nvme", "nvme_path": nvme_path}
400
401
402
            ds_config_zero3_dict = self.get_config_dict(ZERO3)
            ds_config_zero3_dict["zero_optimization"]["offload_optimizer"] = nvme_config
            ds_config_zero3_dict["zero_optimization"]["offload_param"] = nvme_config
403
            trainer = get_regression_trainer(local_rank=0, fp16=True, deepspeed=ds_config_zero3_dict)
404
            with CaptureLogger(deepspeed_logger) as cl:
405
                trainer.train()
406
            self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
407

408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
    @require_optuna
    def test_hyperparameter_search(self):
        with mockenv_context(**self.dist_env_1_gpu):
            ds_config_zero3_dict = self.get_config_dict(ZERO3)

            # hyperparameter_search requires model_init() to recreate the model for each trial
            def model_init():
                config = RegressionModelConfig(a=0, b=0, double_output=False)
                model = RegressionPreTrainedModel(config)
                return model

            trainer = get_regression_trainer(
                local_rank=0,
                fp16=True,
                model_init=model_init,
                deepspeed=ds_config_zero3_dict,
            )

            n_trials = 3
            with CaptureLogger(deepspeed_logger) as cl:
                with CaptureStd() as cs:
                    trainer.hyperparameter_search(direction="maximize", n_trials=n_trials)
            self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
            self.assertIn(f"Trial {n_trials-1} finished with value", cs.err, "expected hyperparameter_search output")
            self.assertIn("Best is trial", cs.err, "expected hyperparameter_search output")

434
435
    # --- These tests need to run on both zero stages --- #

436
437
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_hf_optimizer_with_offload(self, stage, dtype):
438
        # non-DS optimizers can be used with ZERO-offload (as long as they have both CPU and GPU implementation (except LAMB))
439
440
441
        ds_config_dict = self.get_config_dict(stage)
        del ds_config_dict["optimizer"]  # force default HF Trainer optimizer
        # force cpu offload
442
        ds_config_dict["zero_optimization"]["offload_optimizer"]["device"] = "cpu"
443
        ds_config_dict["zero_force_ds_cpu_optimizer"] = False  # offload is not efficient w/o CPUAdam
444
        with mockenv_context(**self.dist_env_1_gpu):
445
            kwargs = {"local_rank": 0, "deepspeed": ds_config_dict}
446
447
            kwargs[dtype] = True
            trainer = get_regression_trainer(**kwargs)
448
            with CaptureLogger(deepspeed_logger) as cl:
449
                trainer.train()
450
            self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
451

452
453
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_fake_notebook_no_launcher(self, stage, dtype):
454
455
456
457
458
        # this setup emulates a notebook where a launcher needs to be emulated by hand

        # note that unittest resets sys.stdout each test, so `CaptureStd` will work here to capture
        # DeepSpeed log if this test happens to run first in this pytest worker. But it will fail if
        # it's run not as a first test as `sys.stdout` will no longer be the same. So we either have
459
460
        # to reset `deepspeed_logger.handlers[0].setStream(sys.stdout)` or directly capture from the deepspeed_logger.
        with mockenv_context(**self.dist_env_1_gpu):
461
            kwargs = {"local_rank": 0, "deepspeed": self.get_config_dict(stage)}
462
463
464
            kwargs[dtype] = True
            trainer = get_regression_trainer(**kwargs)

465
            with CaptureLogger(deepspeed_logger) as cl:
466
                trainer.train()
467
            self.assertIn("DeepSpeed info", cl.out, "expected DeepSpeed logger output but got none")
468

469
470
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_early_get_last_lr(self, stage, dtype):
471
472
473
474
475
476
477
478
        # with deepspeed's fp16 and dynamic loss scale enabled the optimizer/scheduler steps may
        # not run for the first few dozen steps while loss scale is too large, and thus during
        # that time `get_last_lr` will fail if called during that warm up stage,
        #
        # setting `logging_steps=1` forces an early `trainer._maybe_log_save_evaluate()` which calls
        # `self.lr_scheduler.get_last_lr()` and originally it'd fail on the very first step.
        with mockenv_context(**self.dist_env_1_gpu):
            a = b = 0.0
479
480
481
482
483
484
485
486
487
            kwargs = {
                "a": a,
                "b": b,
                "local_rank": 0,
                "train_len": 8,
                "deepspeed": self.get_config_dict(stage),
                "per_device_train_batch_size": 8,
                "logging_steps": 1,
            }
488
489
490
            kwargs[dtype] = True
            trainer = get_regression_trainer(**kwargs)

491
            trainer.train()
492
493
            post_train_a = trainer.model.a.item()

494
495
            # XXX: for some reason the following check fails with zero3/fp16 and any/bf16 - not a
            # broken but a different qualitative outcome - as if optimizer did run
496
497
498
499
            # oddly getting 1.0 for both a and b from 0.0 - there is a bug somewhere
            # print(trainer.model.a.item())
            # print(trainer.model.b.item())
            # need to investigate at some point
500
            if (stage == ZERO3 and dtype == FP16) or (dtype == BF16):
501
                return
502
503
504

            # it's enough that train didn't fail for this test, but we must check that
            # optimizer/scheduler didn't run (since if it did this test isn't testing the right thing)
505
            self.assertEqual(post_train_a, a)
506

507
508
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_gradient_accumulation(self, stage, dtype):
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
        # this test measures that we get identical weights and similar loss with:
        # 1. per_device_train_batch_size=8, gradient_accumulation_steps=1
        # 2. per_device_train_batch_size=4, gradient_accumulation_steps=2
        # since the 2nd should produce the effective batch of 1st, with the same results
        #
        # I can get an identical loss for a small train_len=32, plus the power of the initial
        # dynamic loss scale value set to:
        #   "fp16.initial_scale_power": 1
        # plus having the same WarmupLR's warmup_min_lr == warmup_max_lr in the config file
        # but for some reason going to train_len=64 the weights, weights start to mismatch with this setup.
        # the culprit seems to be `initial_scale_power` - putting it back to its default 32 keeps the weights identical

        train_len = 64
        a = b = 0.0

524
525
526
527
528
529
530
        kwargs = {
            "a": a,
            "b": b,
            "local_rank": 0,
            "train_len": train_len,
            "deepspeed": self.get_config_dict(stage),
        }
531
        kwargs[dtype] = True
532

533
534
        with mockenv_context(**self.dist_env_1_gpu):
            no_grad_accum_trainer = get_regression_trainer(
535
536
                **kwargs,
                per_device_train_batch_size=16,
537
538
539
540
541
542
543
544
545
546
547
                gradient_accumulation_steps=1,
            )
            no_grad_accum_result = no_grad_accum_trainer.train()
            no_grad_accum_loss = no_grad_accum_result.training_loss
            no_grad_accum_a = no_grad_accum_trainer.model.a.item()
            no_grad_accum_b = no_grad_accum_trainer.model.b.item()
            # make sure the optimizer kicked in - if it hasn't changed from the original value of a then make train_len bigger
            self.assertNotEqual(no_grad_accum_a, a)

        with mockenv_context(**self.dist_env_1_gpu):
            yes_grad_accum_trainer = get_regression_trainer(
548
                **kwargs,
549
                per_device_train_batch_size=4,
550
                gradient_accumulation_steps=4,
551
552
553
554
555
556
557
            )
            yes_grad_accum_result = yes_grad_accum_trainer.train()
            yes_grad_accum_loss = yes_grad_accum_result.training_loss
            yes_grad_accum_a = yes_grad_accum_trainer.model.a.item()
            yes_grad_accum_b = yes_grad_accum_trainer.model.b.item()
            self.assertNotEqual(yes_grad_accum_a, a)

558
559
560
561
        # training with half the batch size but accumulation steps as 2 should give the same
        # weights, but sometimes get a slight difference still of 1e-6
        self.assertAlmostEqual(no_grad_accum_a, yes_grad_accum_a, places=5)
        self.assertAlmostEqual(no_grad_accum_b, yes_grad_accum_b, places=5)
562
563

        # see the note above how to get identical loss on a small bs
564
        self.assertAlmostEqual(no_grad_accum_loss, yes_grad_accum_loss, places=2)
565

566
    def check_saved_checkpoints_deepspeed(self, output_dir, freq, total, stage, dtype):
567
        # adapted from TrainerIntegrationCommon.check_saved_checkpoints
568
        file_list = [SAFE_WEIGHTS_NAME, "training_args.bin", "trainer_state.json", "config.json"]
569
570
571
572
573
574
575
576

        if stage == ZERO2:
            ds_file_list = ["mp_rank_00_model_states.pt"]
        elif stage == ZERO3:
            ds_file_list = ["zero_pp_rank_0_mp_rank_00_model_states.pt"]
        else:
            raise ValueError(f"unknown stage {stage}")

577
578
        if dtype == "bf16":
            ds_file_list.append("bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt")
579
580
581

        for step in range(freq, total, freq):
            checkpoint = os.path.join(output_dir, f"checkpoint-{step}")
582
            self.assertTrue(os.path.isdir(checkpoint), f"[{stage}] {checkpoint} dir is not found")
583
584
            # common files
            for filename in file_list:
585
586
                path = os.path.join(checkpoint, filename)
                self.assertTrue(os.path.isfile(path), f"[{stage}] {path} is not found")
587
588
589
590
591
592

            # ds files
            ds_path = os.path.join(checkpoint, f"global_step{step}")
            for filename in ds_file_list:
                # filename = os.path.join(path, filename)
                # print(filename)
593
594
                path = os.path.join(ds_path, filename)
                self.assertTrue(os.path.isfile(path), f"[{stage}] {path} is not found")
595

596
597
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_save_checkpoints(self, stage, dtype):
598
599
        # adapted from  TrainerIntegrationTest.test_save_checkpoints

600
        freq = 5
601
        output_dir = self.get_auto_remove_tmp_dir()
602
        ds_config_dict = self.get_config_dict(stage)
603
604
605
        if dtype == FP16:
            ds_config_dict["fp16"]["initial_scale_power"] = 1  # force optimizer on the first step
        # XXX:
606
        if stage == ZERO3:
607
            ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
608
609
610

        # save checkpoints
        with mockenv_context(**self.dist_env_1_gpu):
611
612
613
614
615
            kwargs = {
                "output_dir": output_dir,
                "save_steps": freq,
                "deepspeed": ds_config_dict,
            }
616
617
            kwargs[dtype] = True
            trainer = get_regression_trainer(**kwargs)
618
619
620
            trainer.train()

        total = int(self.n_epochs * 64 / self.batch_size)
621
        self.check_saved_checkpoints_deepspeed(output_dir, freq, total, stage, dtype)
622

623
624
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_can_resume_training_errors(self, stage, dtype):
625
626
627
        with mockenv_context(**self.dist_env_1_gpu):
            ds_config_dict = self.get_config_dict(stage)
            output_dir = self.get_auto_remove_tmp_dir()
628
            kwargs = {"output_dir": output_dir, "deepspeed": ds_config_dict}
629
630
            kwargs[dtype] = True
            trainer = get_regression_trainer(**kwargs)
631
632
633
634
635
636
637
638

            # 1. fail to find any checkpoint - due a fresh output_dir
            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),
                f"got exception: {context.exception}",
            )
639

640
641
642
643
644
645
646
647
            # 2. fail to find a bogus checkpoint
            with self.assertRaises(Exception) as context:
                checkpoint = os.path.join(output_dir, "checkpoint-5")
                trainer.train(resume_from_checkpoint=f"{checkpoint}-bogus")
            self.assertTrue(
                "Can't find a valid checkpoint at" in str(context.exception), f"got exception: {context.exception}"
            )

648
649
    @parameterized.expand(params_with_optims_and_schedulers, name_func=parameterized_custom_name_func)
    def test_can_resume_training_normal(self, stage, dtype, optim, scheduler):
650
651
        # adapted from TrainerIntegrationTest.test_can_resume_training
        # test normal resume for each stage separately, error-handling is tested in a different test
652
653
654
655
656
657

        # ToDo: Currently, hf_optim + hf_scheduler resumes with the correct states and
        # also has same losses for few steps but then slowly diverges. Need to figure it out.
        if optim == HF_OPTIM and scheduler == HF_SCHEDULER:
            return

658
        output_dir = self.get_auto_remove_tmp_dir("./xxx", after=False)
659
        ds_config_dict = self.get_config_dict(stage)
660
661
662
        if dtype == FP16:
            ds_config_dict["fp16"]["initial_scale_power"] = 1  # force optimizer on the first step
        # XXX:
663
        if stage == ZERO3:
664
            ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
665

666
667
668
669
670
671
        if optim == HF_OPTIM:
            del ds_config_dict["optimizer"]

        if scheduler == HF_SCHEDULER:
            del ds_config_dict["scheduler"]

672
673
674
675
676
677
678
        kwargs = {
            "output_dir": output_dir,
            "train_len": 128,
            "save_steps": 5,
            "learning_rate": 0.1,
            "deepspeed": ds_config_dict,
        }
679
        kwargs[dtype] = True
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711

        with mockenv_context(**self.dist_env_1_gpu):
            trainer = get_regression_trainer(**kwargs)
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
            state = dataclasses.asdict(trainer.state)

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

            # Reinitialize trainer
            trainer = get_regression_trainer(**kwargs)

            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)

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

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

            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)

712
713
714
715
716
            # Finally, should be able to resume with the same trainer/same deepspeed engine instance
            # XXX: but currently this not possible due DS bug: https://github.com/microsoft/DeepSpeed/issues/1612
            # trainer.train(resume_from_checkpoint=checkpoint)
            # a workaround needs to be used that re-creates the deepspeed engine

717
718
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_load_state_dict_from_zero_checkpoint(self, stage, dtype):
719
720
721
722
723
724
        # test that we can load fp32 weights directly from the zero checkpoint into the current model

        output_dir = self.get_auto_remove_tmp_dir()  # "./xxx", after=False, before=False)

        ds_config_dict = self.get_config_dict(stage)

725
726
727
728
729
730
731
732
733
734
        kwargs = {
            "output_dir": output_dir,
            "train_len": 4,
            "per_device_train_batch_size": 4,
            "num_train_epochs": 1,
            "save_strategy": "steps",
            "save_steps": 1,
            "learning_rate": 0.1,
            "deepspeed": ds_config_dict,
        }
735
        kwargs[dtype] = True
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751

        with mockenv_context(**self.dist_env_1_gpu):
            trainer = get_regression_trainer(**kwargs)
            trainer.train()
            (a, b) = trainer.model.a.item(), trainer.model.b.item()
            state = dataclasses.asdict(trainer.state)

            checkpoint_dir = get_last_checkpoint(output_dir)
            model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)

            (a1, b1) = model.a.item(), 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)

752
753
754
755
    def test_config_object(self):
        # test that we can switch from zero2 to zero3 in the same process for example
        # test is_zero, etc.
        output_dir = self.get_auto_remove_tmp_dir()
756
        kwargs = {"output_dir": output_dir, "train_len": 8, "fp16": True}
757

758
759
        ds_config_zero3_dict = self.get_config_dict(ZERO3)
        ds_config_zero2_dict = self.get_config_dict(ZERO2)
760

761
        with mockenv_context(**self.dist_env_1_gpu):
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
            trainer = get_regression_trainer(deepspeed=ds_config_zero3_dict, **kwargs)
            self.assertTrue(is_deepspeed_zero3_enabled())

            # test we can repeat that and with train this time
            trainer = get_regression_trainer(deepspeed=ds_config_zero3_dict, **kwargs)
            trainer.train()
            self.assertTrue(is_deepspeed_zero3_enabled())

            # test zero3 is disabled
            trainer = get_regression_trainer(deepspeed=ds_config_zero2_dict, **kwargs)
            self.assertFalse(is_deepspeed_zero3_enabled())

            # check config obj
            config = deepspeed_config()
            self.assertTrue(bool(config), "Deepspeed config should be accessible")

778
779
            # with accelerate integration below line is additionally required for this test to pass
            trainer.accelerator.state._reset_state()
780
781
782
783
784
785
            del trainer
            # now weakref should gc the global and we shouldn't get anything here
            config = deepspeed_config()
            self.assertFalse(is_deepspeed_zero3_enabled())
            self.assertFalse(bool(config), "Deepspeed config should not be accessible")

786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_load_best_model(self, stage, dtype):
        # Test that forced deepspeed reinit doesn't break the model. the forced re-init after
        # loading the best model in Trainer is there to workaround this bug in Deepspeed
        # https://github.com/microsoft/DeepSpeed/issues/1612
        #
        # The test is derived from a repro script submitted in this Issue:
        # https://github.com/huggingface/transformers/issues/17114
        #
        # One additional feature of this test is that we use a non-AdamW optimizer to test that
        # deepspeed doesn't fallback to AdamW, which would prevent the optimizer states from loading
        # correctly

        from transformers import T5ForConditionalGeneration, T5Tokenizer, Trainer  # noqa

        output_dir = self.get_auto_remove_tmp_dir()  # "./xxx", after=False, before=False)

        ds_config_dict = self.get_config_dict(stage)
        del ds_config_dict["optimizer"]  # will use HF Trainer optimizer
        del ds_config_dict["scheduler"]  # will use HF Trainer scheduler
806
        ds_config_dict["zero_force_ds_cpu_optimizer"] = False  # offload is not efficient w/o CPUAdam
807
808
809
        # must use this setting to get the reload path exercised
        ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True

810
        with mockenv_context(**self.dist_env_1_gpu):
811
            args_dict = {
812
813
                "per_device_train_batch_size": 1,
                "per_device_eval_batch_size": 1,
814
815
816
817
818
819
820
821
822
823
824
825
826
                "gradient_accumulation_steps": 1,
                "learning_rate": 1e-4,
                "num_train_epochs": 1,
                "do_train": True,
                "do_eval": True,
                "optim": "adafactor",
                "evaluation_strategy": "steps",
                "eval_steps": 1,
                "save_strategy": "steps",
                "save_steps": 1,
                "load_best_model_at_end": True,
                "max_steps": 1,
                "deepspeed": ds_config_dict,
827
                "report_to": "none",
828
829
830
            }

            training_args = TrainingArguments(output_dir, **args_dict)
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
            tokenizer = T5Tokenizer.from_pretrained(T5_TINY)
            model = T5ForConditionalGeneration.from_pretrained(T5_TINY)

            def _add_eos_to_examples(example):
                example["input_text"] = f"question: {example['question']}  context: {example['context']}"
                example["target_text"] = example["answers"]["text"][0] if len(example["answers"]["text"]) > 0 else ""
                return example

            def _convert_to_features(example_batch):
                input_encodings = tokenizer.batch_encode_plus(
                    example_batch["input_text"], pad_to_max_length=True, max_length=512, truncation=True
                )
                target_encodings = tokenizer.batch_encode_plus(
                    example_batch["target_text"], pad_to_max_length=True, max_length=16, truncation=True
                )

                encodings = {
                    "input_ids": input_encodings["input_ids"],
                    "attention_mask": input_encodings["attention_mask"],
                    "labels": target_encodings["input_ids"],
                }

                return encodings

            def get_dataset():
                data_file = str(self.tests_dir / "fixtures/tests_samples/SQUAD/sample.json")
857
                data_files = {"train": data_file, "validation": data_file}
858
859
860
861
862
863
864
                raw_datasets = datasets.load_dataset("json", data_files=data_files, field="data")
                train_dataset = raw_datasets["train"].map(_add_eos_to_examples).map(_convert_to_features, batched=True)
                valid_dataset = deepcopy(train_dataset)
                return train_dataset, valid_dataset

            train_dataset, eval_dataset = get_dataset()

865
866
867
868
869
870
871
872
873
874
            trainer = Trainer(
                model=model,
                tokenizer=tokenizer,
                args=training_args,
                train_dataset=train_dataset,
                eval_dataset=eval_dataset,
            )
            trainer.train()  # crash 1 was here
            trainer.evaluate()  # crash 2 was here

875
876
877
878

@slow
@require_deepspeed
@require_torch_gpu
879
class TestDeepSpeedWithLauncher(TestCasePlus):
Patrick von Platen's avatar
Patrick von Platen committed
880
    """This class is for testing via an external script - can do multiple gpus"""
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896

    # Tests to devise #
    #
    # 1. predict_with_generate on multigpu - need to figure out how to give input sequences so that
    # the 2 gpus will generate prediction sequences that aren't of the same length - this is because
    # we had to code a special feature to sync the gpus when the predicted sequences aren't of the
    # same length. In general this will tested as a side-effect through a variety of other tests -
    # it'll simply hang trying to synchronize with other gpus if this problem is encountered. So as
    # long as we have a few full tests running on zero3 + predict_with_generate this should be
    # mostly covered.
    #
    # but there are 5 variations on beam search in `generate`- with identical code branched with `if
    # synced_gpus`
    #
    # 2. most tests should probably be run on both: zero2 and zero3 configs
    #
897

898
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
Yih-Dar's avatar
Yih-Dar committed
899
    @require_torch_multi_gpu
900
901
    def test_basic_distributed(self, stage, dtype):
        self.run_and_check(stage=stage, dtype=dtype, distributed=True)
902

903
904
    def test_do_eval_no_train(self):
        # testing only zero3 since zero2 makes no sense with inference
905
        self.run_and_check(
906
            stage=ZERO3,
907
            dtype=FP16,
908
909
            eval_steps=1,
            distributed=False,
910
911
            do_train=False,
            do_eval=True,
912
        )
913

914
915
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_fp32_non_distributed(self, stage, dtype):
916
917
918
919
        # real model needs too much GPU memory under stage2+fp32, so using tiny random model here -
        # therefore no quality checks, just basic completion checks are done
        self.run_and_check(
            stage=stage,
920
            dtype=dtype,
921
922
923
924
925
            model_name=T5_TINY,
            distributed=False,
            do_train=True,
            do_eval=True,
            quality_checks=False,
926
            fp32=True,
927
928
        )

929
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
Yih-Dar's avatar
Yih-Dar committed
930
    @require_torch_multi_gpu
931
    def test_fp32_distributed(self, stage, dtype):
932
933
934
935
        # real model needs too much GPU memory under stage2+fp32, so using tiny random model here -
        # therefore no quality checks, just basic completion checks are done
        self.run_and_check(
            stage=stage,
936
            dtype=dtype,
937
938
939
940
941
            model_name=T5_TINY,
            distributed=True,
            do_train=True,
            do_eval=True,
            quality_checks=False,
942
            fp32=True,
943
944
        )

945
946
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_resume_train_not_from_ds_checkpoint(self, stage, dtype):
947
948
949
950
951
        # do normal training and then resume not from the deepspeed checkpoint but explicitly from
        # the saved model dir

        do_train = True
        do_eval = False
952
953
954
955
956
957
958
959
        kwargs = {
            "stage": stage,
            "dtype": dtype,
            "eval_steps": 1,
            "distributed": True,
            "do_train": do_train,
            "do_eval": do_eval,
        }
960
961
962
963
964
965
966
967
968
969

        # 1. normal training
        output_dir = self.run_and_check(**kwargs)

        # 2. now resume explicitly from the saved weights, by passing --model_name_or_path output_dir
        # - i.e. the same path the model was saved to in step 1
        output_dir = self.run_trainer(**kwargs, model_name=output_dir)

        self.do_checks(output_dir, do_train=do_train, do_eval=do_eval)

970
    @parameterized.expand(["bf16", "fp16", "fp32"])
Yih-Dar's avatar
Yih-Dar committed
971
    @require_torch_multi_gpu
972
    def test_inference(self, dtype):
973
        if dtype == "bf16" and not is_torch_bf16_gpu_available():
974
975
            self.skipTest("test requires bfloat16 hardware support")

976
977
        # this is just inference, so no optimizer should be loaded
        # it only works for z3 (makes no sense with z1-z2)
978
        fp32 = True if dtype == "fp32" else False
979
980
        self.run_and_check(
            stage=ZERO3,
981
            dtype=FP16,
982
983
984
985
986
            model_name=T5_TINY,
            distributed=True,
            do_train=False,
            do_eval=True,
            quality_checks=False,
987
            fp32=fp32,
988
989
        )

990
    def do_checks(self, output_dir, do_train=True, do_eval=True, quality_checks=True):
991
992
993
        if do_train:
            train_metrics = load_json(os.path.join(output_dir, "train_results.json"))
            self.assertIn("train_samples_per_second", train_metrics)
994
995
            if quality_checks:
                self.assertGreater(train_metrics["train_samples_per_second"], 0.5)
996
997
998
999

        if do_eval:
            eval_metrics = load_json(os.path.join(output_dir, "eval_results.json"))
            self.assertIn("eval_bleu", eval_metrics)
1000
1001
            if quality_checks:
                self.assertGreater(eval_metrics["eval_bleu"], 1)
1002
1003

    # XXX: need to do better validation beyond just that the run was successful
1004
1005
1006
    def run_and_check(
        self,
        stage,
1007
        dtype,
1008
1009
1010
1011
1012
1013
        model_name: str = T5_SMALL,
        eval_steps: int = 10,
        distributed: bool = True,
        do_train: bool = True,
        do_eval: bool = True,
        quality_checks: bool = True,
1014
        fp32: bool = False,
1015
1016
        extra_args_str: str = None,
        remove_args_str: str = None,
1017
1018
    ):
        # we are doing quality testing so using a small real model
1019
        output_dir = self.run_trainer(
1020
            stage=stage,
1021
            dtype=dtype,
1022
            model_name=model_name,
1023
            eval_steps=eval_steps,
1024
            num_train_epochs=1,
1025
1026
            do_train=do_train,
            do_eval=do_eval,
1027
            distributed=distributed,
1028
            fp32=fp32,
1029
1030
1031
            extra_args_str=extra_args_str,
            remove_args_str=remove_args_str,
        )
1032

1033
        self.do_checks(output_dir, do_train=do_train, do_eval=do_eval, quality_checks=quality_checks)
1034
1035

        return output_dir
1036
1037
1038

    def run_trainer(
        self,
1039
        stage: str,
1040
        dtype: str,
1041
        model_name: str,
1042
1043
1044
1045
        eval_steps: int = 10,
        num_train_epochs: int = 1,
        do_train: bool = False,
        do_eval: bool = True,
1046
        distributed: bool = True,
1047
        fp32: bool = False,
1048
1049
1050
        extra_args_str: str = None,
        remove_args_str: str = None,
    ):
1051
        max_len = 32
Sylvain Gugger's avatar
Sylvain Gugger committed
1052
        data_dir = self.test_file_dir / "../fixtures/tests_samples/wmt_en_ro"
1053
1054
1055
        output_dir = self.get_auto_remove_tmp_dir()
        args = f"""
            --model_name_or_path {model_name}
1056
1057
            --train_file {data_dir}/train.json
            --validation_file {data_dir}/val.json
1058
1059
1060
1061
1062
1063
1064
            --output_dir {output_dir}
            --overwrite_output_dir
            --max_source_length {max_len}
            --max_target_length {max_len}
            --val_max_target_length {max_len}
            --warmup_steps 8
            --predict_with_generate
1065
1066
            --save_steps 0
            --eval_steps {eval_steps}
1067
1068
            --group_by_length
            --label_smoothing_factor 0.1
1069
1070
            --source_lang en
            --target_lang ro
1071
            --report_to none
1072
        """.split()
1073
1074
        args.extend(["--source_prefix", '"translate English to Romanian: "'])

1075
1076
        if not fp32:
            args.extend([f"--{dtype}"])
1077

1078
1079
1080
1081
1082
1083
1084
        actions = 0
        if do_train:
            actions += 1
            args.extend(
                f"""
            --do_train
            --num_train_epochs {str(num_train_epochs)}
1085
            --max_train_samples 16
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
            --per_device_train_batch_size 2
            --learning_rate 3e-3
            """.split()
            )

        if do_eval:
            actions += 1
            args.extend(
                """
            --do_eval
1096
            --max_eval_samples 16
1097
1098
1099
1100
1101
            --per_device_eval_batch_size 2
            """.split()
            )

        assert actions > 0, "need at least do_train or do_eval for the test to run"
1102
1103
1104
1105

        if extra_args_str is not None:
            args.extend(extra_args_str.split())

1106
        # currently only works for bool args
1107
1108
1109
1110
        if remove_args_str is not None:
            remove_args = remove_args_str.split()
            args = [x for x in args if x not in remove_args]

1111
        ds_args = f"--deepspeed {self.test_file_dir_str}/ds_config_{stage}.json".split()
Sylvain Gugger's avatar
Sylvain Gugger committed
1112
        script = [f"{self.examples_dir_str}/pytorch/translation/run_translation.py"]
1113
        launcher = get_launcher(distributed)
1114
1115

        cmd = launcher + script + args + ds_args
1116
        # keep for quick debug
1117
1118
1119
1120
1121
        # print(" ".join([f"\nPYTHONPATH={self.src_dir_str}"] +cmd)); die
        execute_subprocess_async(cmd, env=self.get_env())

        return output_dir

1122
1123
    @parameterized.expand(params, name_func=parameterized_custom_name_func)
    def test_clm(self, stage, dtype):
1124
1125
1126
1127
1128
1129
        # this test exercises model.resize_token_embeddings() which requires param gathering outside
        # of forward - it's not used by `run_translation.py`, but it is in `run_clm.py`

        data_dir = self.tests_dir / "fixtures"
        output_dir = self.get_auto_remove_tmp_dir()
        args = f"""
1130
            --model_name_or_path {GPT2_TINY}
1131
1132
1133
1134
1135
1136
            --train_file {data_dir}/sample_text.txt
            --validation_file {data_dir}/sample_text.txt
            --output_dir {output_dir}
            --overwrite_output_dir
            --do_train
            --do_eval
1137
1138
1139
1140
            --max_train_samples 16
            --max_eval_samples 16
            --per_device_train_batch_size 2
            --per_device_eval_batch_size 2
1141
1142
            --num_train_epochs 1
            --warmup_steps 8
1143
            --block_size 64
1144
            --report_to none
1145
1146
            """.split()

1147
1148
        args.extend([f"--{dtype}"])

1149
        ds_args = f"--deepspeed {self.test_file_dir_str}/ds_config_{stage}.json".split()
Sylvain Gugger's avatar
Sylvain Gugger committed
1150
        script = [f"{self.examples_dir_str}/pytorch/language-modeling/run_clm.py"]
1151
        launcher = get_launcher(distributed=True)
1152
1153
1154
1155

        cmd = launcher + script + args + ds_args
        # keep for quick debug
        # print(" ".join([f"\nPYTHONPATH={self.src_dir_str}"] +cmd)); die
1156
1157
        execute_subprocess_async(cmd, env=self.get_env())

1158
    def test_clm_from_config_zero3_fp16(self):
1159
1160
1161
1162
1163
1164
        # this test exercises AutoModel.from_config(config) - to ensure zero.Init is called

        data_dir = self.tests_dir / "fixtures"
        output_dir = self.get_auto_remove_tmp_dir()
        args = f"""
            --model_type gpt2
1165
            --tokenizer_name {GPT2_TINY}
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
            --train_file {data_dir}/sample_text.txt
            --validation_file {data_dir}/sample_text.txt
            --output_dir {output_dir}
            --overwrite_output_dir
            --do_train
            --max_train_samples 4
            --per_device_train_batch_size 2
            --num_train_epochs 1
            --warmup_steps 8
            --block_size 8
            --fp16
            --report_to none
            """.split()

        ds_args = f"--deepspeed {self.test_file_dir_str}/ds_config_zero3.json".split()
        script = [f"{self.examples_dir_str}/pytorch/language-modeling/run_clm.py"]
1182
        launcher = get_launcher(distributed=True)
1183
1184
1185
1186
1187
1188

        cmd = launcher + script + args + ds_args
        # keep for quick debug
        # print(" ".join([f"\nPYTHONPATH={self.src_dir_str}"] +cmd)); die
        with CaptureStderr() as cs:
            execute_subprocess_async(cmd, env=self.get_env())
1189
        self.assertIn("Detected DeepSpeed ZeRO-3", cs.err)