fleet_base.py 47.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
#   Copyright (c) 2020 PaddlePaddle Authors. 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.

from __future__ import print_function
16
import copy
17
import warnings
18
import paddle
19
import os
20
from paddle.fluid.framework import dygraph_only
21
from paddle.fluid import compiler
22
from .role_maker import UserDefinedRoleMaker, PaddleCloudRoleMaker, RoleMakerBase
23
from .strategy_compiler import StrategyCompiler
24
from .distributed_strategy import DistributedStrategy
25 26
from .meta_optimizer_factory import MetaOptimizerFactory
from .runtime_factory import RuntimeFactory
27
from paddle.fluid.wrapped_decorator import wrap_decorator
28
from paddle.fluid.dygraph import parallel_helper
29
from . import topology as tp
30 31
from .topology import ParallelMode
from ..meta_parallel import ModelParallel
32
from ..meta_parallel import PipelineParallel
33
from ..meta_optimizers import HybridParallelOptimizer
34
from ..meta_optimizers import HybridParallelGradScaler
35

36 37
__all__ = []

38

39 40 41 42 43 44 45 46 47 48 49 50
def _inited_runtime_handler_(func):
    def __impl__(*args, **kwargs):
        cls = args[0]

        if cls._runtime_handle is None:
            raise ValueError("Fleet can not find suitable runtime handler")

        return func(*args, **kwargs)

    return __impl__


51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66
def _is_non_distributed_check_(func):
    def __impl__(*args, **kwargs):
        cls = args[0]

        if cls._role_maker is not None and cls._role_maker._is_non_distributed(
        ) is True:
            warnings.warn(
                "%s() function doesn't work when use non_distributed fleet." %
                (func.__name__))
            return

        return func(*args, **kwargs)

    return __impl__


67
inited_runtime_handler = wrap_decorator(_inited_runtime_handler_)
68
is_non_distributed_check = wrap_decorator(_is_non_distributed_check_)
69 70


71 72 73
class Fleet(object):
    """
    Unified API for distributed training of PaddlePaddle
74
    Please reference the https://github.com/PaddlePaddle/FleetX for details
75 76 77 78 79


    Returns:
        Fleet: A Fleet instance

80
    Example for collective training:
1
123malin 已提交
81

82 83
        .. code-block:: python

1
123malin 已提交
84 85
            import paddle
            paddle.enable_static()
86
            import paddle.distributed.fleet as fleet
87 88 89

            fleet.init(is_collective=True)

90 91 92
            strategy = fleet.DistributedStrategy()
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
            optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
93 94 95 96 97 98 99 100

            # do distributed training


    Example for parameter server training:

        .. code-block:: python

1
123malin 已提交
101 102
            import paddle
            paddle.enable_static()
103 104
            import paddle.distributed.fleet as fleet
            strategy = fleet.DistributedStrategy()
S
ShenLiang 已提交
105
            fleet.init(strategy=strategy)
106

107
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
108
            optimizer = fleet.distributed_optimizer(optimizer)
109

110 111
            if fleet.is_first_worker():
                print("this is first worker")
112

113 114
            print("current node index: {}".format(fleet.worker_index()))
            print("total number of worker num: {}".format(fleet.worker_num()))
115

116 117 118
            if fleet.is_worker():
                print("this is worker")
            print("worker endpoints: {}".format(fleet.worker_endpoints(to_string=True)))
119

120 121
            print("server num: {}".format(fleet.server_num()))
            print("server endpoints: {}".format(fleet.server_endpoints(to_string=True)))
122

123 124 125
            if fleet.is_server():
                print("this is server")
            fleet.stop_worker()
126 127


128 129 130
    """

    def __init__(self):
131
        self._role_maker = None
132
        self.strategy_compiler = None
133
        self._is_collective = False
134
        self._runtime_handle = None
D
Dong Daxiang 已提交
135 136
        self._util = None
        self._context = {}
137

138
    def init(self, role_maker=None, is_collective=False, strategy=None):
139 140 141
        """
        Initialize role_maker in Fleet.

142 143 144 145 146 147 148 149 150 151 152
        This function is responsible for the distributed architecture
        what you want to run your code behind.

        Args:
            role_maker (RoleMakerBase, optional): A ``RoleMakerBase`` containing the configuration
                of environment variables related to distributed training.If you did not initialize 
                the rolemaker by yourself, it will be automatically initialized to PaddleRoleMaker.
                The default value is None.
            is_collective (Boolean, optional): A ``Boolean`` variable determines whether the program 
                runs on the CPU or GPU. False means set distributed training using CPU, and True means
                GPU.The default value is False.The default value is False.
153 154 155 156
            strategy (DistributedStrategy): Extra properties for distributed training. 
                For details, please refer to paddle.distributed.fleet.DistributedStrategy. Default: None.


157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
        Returns:
            None

        Examples1:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

        Examples2:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init(is_collective=True)

        Examples3:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
1
123malin 已提交
179
                role = fleet.PaddleCloudRoleMaker()
180
                fleet.init(role)
181

182 183 184 185 186 187
        Examples4:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                strategy = fleet.DistributedStrategy()
S
ShenLiang 已提交
188
                fleet.init(strategy=strategy)
189

190
        """
S
ShenLiang 已提交
191 192 193
        if strategy is None:
            strategy = DistributedStrategy()
        self._user_defined_strategy = copy.deepcopy(strategy)
194 195

        if role_maker is None:
196 197 198 199 200 201
            if isinstance(is_collective, bool):
                self._is_collective = is_collective
                self._role_maker = PaddleCloudRoleMaker(
                    is_collective=self._is_collective)
            else:
                raise ValueError(
202 203
                    "`is_collective` should be instance of `bool`, but got {}".
                    format(type(is_collective)))
204
        else:
205 206
            if isinstance(role_maker, RoleMakerBase):
                self._role_maker = role_maker
207
                self._is_collective = role_maker._is_collective
208 209 210 211
            else:
                raise ValueError(
                    "`role_maker` should be subclass of `RoleMakerBase`, but got {}".
                    format(type(role_maker)))
212
        self._role_maker._generate_role()
213

214 215 216
        import paddle.distributed.fleet as fleet
        fleet.util._set_role_maker(self._role_maker)

217
        self.strategy_compiler = StrategyCompiler()
218 219 220 221 222 223 224 225 226

        if self._role_maker._is_non_distributed() and self._is_collective:
            if paddle.fluid.core.is_compiled_with_cuda():
                gpus_num = paddle.fluid.core.get_cuda_device_count()
                if gpus_num != 1:
                    raise ValueError(
                        "CUDA_VISIBLE_DEVICES shoule be set only 1 card if you use `python` to launch fleet program."
                    )

227
        if paddle.fluid.framework.in_dygraph_mode():
228
            if self.worker_num() == 1:
229 230 231
                # if worker_num is 1, should construct default topology & hcg
                self._topology = tp.CommunicateTopology()
                self._hcg = tp.HybridCommunicateGroup(self._topology)
232
                return
233 234 235 236
            if parallel_helper._is_parallel_ctx_initialized():
                warnings.warn(
                    "The dygraph parallel environment has been initialized.")
            else:
237 238 239 240 241 242 243 244 245
                # FLAGS_nccl_nrings is used for dynamic graph multi-stream communication
                if "FLAGS_nccl_nrings" in os.environ:
                    warnings.warn(
                        "You have set the environment variable FLAGS_nccl_nrings "
                        "outside the program, so the nccl_comm_num in "
                        "DistributedStrategy will not take effect here.")
                else:
                    os.environ["FLAGS_nccl_nrings"] = str(
                        self._user_defined_strategy.nccl_comm_num)
246
                paddle.distributed.init_parallel_env()
247

248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
            # init hybrid parallel environment in dygraph
            if tp._HYBRID_PARALLEL_GROUP is None:
                self._init_hybrid_parallel_env()
            else:
                warnings.warn(
                    "The dygraph hybrid parallel environment has been initialized."
                )

    def _init_hybrid_parallel_env(self):
        """initialize the hybrid environment
        """
        self.hybrid_configs = self._user_defined_strategy.hybrid_configs
        self.dp_degree = self.hybrid_configs["dp_degree"]
        self.mp_degree = self.hybrid_configs["mp_degree"]
        self.pp_degree = self.hybrid_configs["pp_degree"]

        assert self.mp_degree >= 0, "mp_degree should be greater or equal to 0"
        assert self.pp_degree >= 0, "pp_degree should be greater or equal to 0"

        self.mp_degree = max(self.mp_degree, 1)
        self.pp_degree = max(self.pp_degree, 1)

        if self.dp_degree < 0:
            nranks = paddle.distributed.get_world_size()
            self.dp_degree = nranks // (self.mp_degree * self.pp_degree)

        self.dp_degree = max(self.dp_degree, 1)

        self._topology = tp.CommunicateTopology(
            hybrid_group_names=["data", "pipe", "model"],
            dims=[self.dp_degree, self.pp_degree, self.mp_degree])

        self._hcg = tp.HybridCommunicateGroup(self._topology)

    def get_hybrid_communicate_group(self):
        assert self._hcg is not None
        return self._hcg

    def get_hybrid_parallel_topology(self):
        assert self._topology is not None
        return self._topology

290 291 292 293 294 295 296
    def is_first_worker(self):
        """
        Check whether the node is the first instance of worker.

        Returns:
            bool: True if this is the first node of worker,
                  False if not.
297

298 299 300 301 302 303 304 305
        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.is_first_worker()

306
        """
307
        return self._role_maker._is_first_worker()
308 309 310 311 312 313 314

    def worker_index(self):
        """
        Get current worker index.

        Returns:
            int: node id
315 316 317 318

        Examples:

            .. code-block:: python
1
123malin 已提交
319

320 321 322 323
                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.worker_index()

324
        """
325
        return self._role_maker._worker_index()
326 327 328 329 330 331 332

    def worker_num(self):
        """
        Get current total worker number.

        Returns:
            int: worker numbers
1
123malin 已提交
333

334
        Examples:
1
123malin 已提交
335

336 337 338 339 340 341
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.worker_num()

342
        """
343
        return self._role_maker._worker_num()
344

345 346 347 348 349 350 351 352 353 354 355 356
    def node_num(self):
        return self._role_maker._get_node_num()

    def local_rank(self):
        return self._role_maker._get_local_rank()

    def local_device_ids(self):
        return self._role_maker._get_local_device_ids()

    def world_device_ids(self):
        return self._role_maker._get_world_device_ids()

357 358 359 360 361 362 363
    def is_worker(self):
        """
        Check whether the node is an instance of worker.

        Returns:
            bool: True if this is a node of worker,
                  False if not.
364 365

        Examples:
1
123malin 已提交
366

367 368 369 370 371 372
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.is_worker()

373
        """
374
        return self._role_maker._is_worker()
375 376 377

    def worker_endpoints(self, to_string=False):
        """
378
        Get current worker endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"].
379 380 381

        Returns:
            list/string: server endpoints
382 383

        Examples:
1
123malin 已提交
384

385 386 387 388 389 390
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.worker_endpoints()

391 392
        """
        if to_string:
393
            return ",".join(self._role_maker._get_trainer_endpoints())
394
        else:
395
            return self._role_maker._get_trainer_endpoints()
396 397 398 399 400 401 402

    def server_num(self):
        """
        Get current total worker number.

        Returns:
            int: server number
403 404

        Examples:
1
123malin 已提交
405

406
            .. code-block:: python
1
123malin 已提交
407 408 409 410

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.server_num()
411
        """
412
        return len(self._role_maker._get_pserver_endpoints())
413 414 415 416 417 418 419

    def server_index(self):
        """
        Get current server index.

        Returns:
            int: node id
420 421

        Examples:
1
123malin 已提交
422

423 424 425 426 427 428
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.server_index()

429
        """
430
        return self._role_maker._server_index()
431 432 433 434 435 436 437

    def server_endpoints(self, to_string=False):
        """
        Get current server endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"].

        Returns:
            list/string: server endpoints
438 439

        Examples:
1
123malin 已提交
440

441 442 443 444 445 446
            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.server_endpoints()

447
        """
448

449
        if to_string:
450
            return ",".join(self._role_maker._get_pserver_endpoints())
451
        else:
452
            return self._role_maker._get_pserver_endpoints()
453 454 455 456 457 458 459 460

    def is_server(self):
        """
        Check whether the node is an instance of server.

        Returns:
            bool: True if this is a node of server,
                  False if not.
461 462 463 464

        Examples:

            .. code-block:: python
1
123malin 已提交
465

466 467 468 469
                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.is_server()

470
        """
471
        return self._role_maker._is_server(
472
        ) or self._role_maker._is_heter_worker()
473 474 475

    def barrier_worker(self):
        """
476 477 478 479
        barrier all workers

        Returns:
            None
480
        """
481
        self._role_maker._barrier("worker")
482

483
    @is_non_distributed_check
484
    @inited_runtime_handler
485 486
    def init_worker(self):
        """
487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504
        initialize `Communicator` for parameter server training.


        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_worker()

505 506 507
        """
        self._runtime_handle._init_worker()

508
    @is_non_distributed_check
509
    @inited_runtime_handler
510
    def init_server(self, *args, **kwargs):
511
        """
512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530
        init_server executor to initialize startup program,
        if the `args` is not empty, it will run load_persistables for increment training.


        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_server()

531
        """
532
        self._runtime_handle._init_server(*args, **kwargs)
533

534
    @is_non_distributed_check
535
    @inited_runtime_handler
536 537
    def run_server(self):
        """
538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555
        run server will run pserver main program with executor.

        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                if fleet.is_server():
                    fleet.init_server()

556 557 558
        """
        self._runtime_handle._run_server()

559
    @is_non_distributed_check
560
    @inited_runtime_handler
561 562
    def stop_worker(self):
        """
563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579
        stop `Communicator` and give training complete notice to parameter server.

        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_server()

580 581 582
        """
        self._runtime_handle._stop_worker()

T
tangwei12 已提交
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
    def save(self, dirname, feed=[], fetch=[], **configs):
        inference = True

        if not feed and not fetch:
            inference = False

        place = paddle.CPUPlace()
        executor = paddle.static.Executor(place)

        if inference:
            feeded_var_names = []
            fetch_var_names = []

            for var in feed:
                if isinstance(var, str):
                    feeded_var_names.append(var)
                elif isinstance(var, paddle.static.Variable):
                    feeded_var_names.append(var.name)
                else:
                    raise ValueError("feed must be [str|Variable]")

            for var in fetch:
                if isinstance(var, str):
                    fetch_var_names.append(var)
                elif isinstance(var, paddle.static.Variable):
                    fetch_var_names.append(var.name)
                else:
                    raise ValueError("feed must be [str|Variable]")

            fetch_vars = [
                paddle.static.default_main_program().global_block().var(name)
                for name in fetch_var_names
            ]

            self._runtime_handle._save_inference_model(
                executor, dirname, feeded_var_names, fetch_vars, None, True, 0)
        else:
            increment_mode = 0
            if "mode" in configs:
                increment_mode = int(configs["mode"])
            self._runtime_handle._save_persistables(
                executor, dirname, main_program=None, mode=increment_mode)

626 627 628 629 630 631
    def save_inference_model(self,
                             executor,
                             dirname,
                             feeded_var_names,
                             target_vars,
                             main_program=None,
632 633
                             export_for_deployment=True,
                             mode=0):
634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652
        """
        save inference model for inference.

        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

                fleet.init_server()

        """
T
tangwei12 已提交
653 654 655
        # warnings.warn(
        #     "'save_inference_model' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead."
        # )
656

657 658
        self._runtime_handle._save_inference_model(
            executor, dirname, feeded_var_names, target_vars, main_program,
659
            export_for_deployment, mode)
660

661
    def save_persistables(self, executor, dirname, main_program=None, mode=0):
662 663
        """

1
123malin 已提交
664
        saves all persistable tensors from :code:`main_program` to
665 666
        the folder :code:`dirname`. You can refer to

1
123malin 已提交
667 668
        The :code:`dirname` is used to specify the folder where persistable tensors
        are going to be saved. If you would like to save tensors in separate
669 670 671
        files, set :code:`filename` None.

        Args:
1
123malin 已提交
672
            executor(Executor): The executor to run for saving persistable tensors.
673 674 675 676 677
                                You can refer to :ref:`api_guide_executor_en` for
                                more details.

            dirname(str, optional): The saving directory path.
                                When you need to save the parameter to the memory, set it to None.
1
123malin 已提交
678
            main_program(Program, optional): The program whose persistbale tensors will
679 680 681 682 683 684 685 686 687 688
                                             be saved. Default: None.


        Returns:
            None

        Examples:

            .. code-block:: text

1
123malin 已提交
689 690
                import paddle
                paddle.enable_static()
691 692 693 694 695 696 697
                import paddle.distributed.fleet as fleet

                fleet.init()

                # build net
                # fleet.distributed_optimizer(...)

1
123malin 已提交
698 699
                exe = paddle.static.Executor(paddle.CPUPlace())
                fleet.save_persistables(exe, "dirname", paddle.static.default_main_program())
700 701

        """
T
tangwei12 已提交
702 703 704
        # warnings.warn(
        #     "'save_persistables' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead."
        # )
705

706 707
        self._runtime_handle._save_persistables(executor, dirname, main_program,
                                                mode)
708

709 710 711
    def shrink(self, threshold):
        self._runtime_handle._shrink(threshold)

712
    def distributed_optimizer(self, optimizer, strategy=None):
713
        """
714 715 716 717 718 719 720
        Optimizer for distributed training.

        For the distributed training, this method would rebuild a new instance of DistributedOptimizer.
        Which has basic Optimizer function and special features for distributed training.

        Args:
            optimizer(Optimizer): The executor to run for init server.
721 722 723 724 725
            strategy(DistributedStrategy): Extra properties for distributed optimizer. 
                It is recommended to use DistributedStrategy in fleet.init(). The strategy
                here is for compatibility. If the strategy in fleet.distributed_optimizer() 
                is not None, then it will overwrite the DistributedStrategy in fleet.init(), 
                which will take effect in distributed training.
726

727
        Returns:
728
            Fleet: instance of fleet.
729 730

        Examples:
731

732
            .. code-block:: python
733

1
123malin 已提交
734
                import paddle
735
                import paddle.distributed.fleet as fleet
1
123malin 已提交
736
                fleet.init(is_collective=True)
737 738 739 740
                strategy = fleet.DistributedStrategy()
                optimizer = paddle.optimizer.SGD(learning_rate=0.001)
                optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)

741 742
        """
        self.user_defined_optimizer = optimizer
743

744
        if strategy is not None:
T
tangwei12 已提交
745 746 747 748 749 750 751
            if self._is_collective:
                warnings.warn(
                    "It is recommended to use DistributedStrategy "
                    "in fleet.init(). The strategy here is only for compatibility. "
                    "If the strategy in fleet.distributed_optimizer() is "
                    "not None, then it will overwrite the DistributedStrategy in fleet.init(), "
                    "which will take effect in distributed training.")
752
            self._user_defined_strategy = copy.deepcopy(strategy)
D
Dong Daxiang 已提交
753 754

        self._context = {}
S
ShenLiang 已提交
755 756

        if paddle.fluid.framework.in_dygraph_mode():
757 758 759 760 761
            if self.worker_num() > 1:
                return HybridParallelOptimizer(optimizer, self._hcg,
                                               self._user_defined_strategy)
            else:
                return optimizer
762 763
        return self

764
    @dygraph_only
765
    def distributed_model(self, model):
766
        """
767 768 769 770 771 772 773
        Return distributed data parallel model (Only work in dygraph mode)

        Args:
            model (Layer): the user-defind model which inherits Layer.

        Returns:
            distributed data parallel model which inherits Layer.
774 775

        Examples:
776

777 778
            .. code-block:: python

779 780 781 782 783 784 785 786 787
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet

                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
788

789 790
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
791

1
123malin 已提交
792
                # 1. initialize fleet environment
793 794
                fleet.init(is_collective=True)

1
123malin 已提交
795
                # 2. create layer & optimizer
796 797 798 799 800
                layer = LinearNet()
                loss_fn = nn.MSELoss()
                adam = paddle.optimizer.Adam(
                    learning_rate=0.001, parameters=layer.parameters())

1
123malin 已提交
801
                # 3. get data_parallel model using fleet
802 803 804
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

1
123malin 已提交
805
                # 4. run layer
806 807 808 809 810 811 812 813 814 815 816 817
                inputs = paddle.randn([10, 10], 'float32')
                outputs = dp_layer(inputs)
                labels = paddle.randn([10, 1], 'float32')
                loss = loss_fn(outputs, labels)

                print("loss:", loss.numpy())

                loss.backward()

                adam.step()
                adam.clear_grad()

818

819
        """
820 821 822 823 824 825 826 827 828 829 830 831 832 833 834
        assert model is not None, "model should not be None"
        if self.worker_num() <= 1:
            return model
        if self._hcg.get_parallel_mode() == ParallelMode.DATA_PARALLEL:
            distributed_model = paddle.DataParallel(
                model,
                comm_buffer_size=self._user_defined_strategy.
                fuse_grad_size_in_MB,
                last_comm_buffer_size=self._user_defined_strategy.
                last_comm_group_size_MB,
                find_unused_parameters=self._user_defined_strategy.
                find_unused_parameters)
        elif self._hcg.get_parallel_mode() == ParallelMode.MODEL_PARALLEL:
            distributed_model = ModelParallel(
                model, self._hcg, strategy=self._user_defined_strategy)
835 836 837
        elif self._hcg.get_parallel_mode() == ParallelMode.PIPELINE_PARALLEL:
            distributed_model = PipelineParallel(
                model, self._hcg, strategy=self._user_defined_strategy)
838
        return distributed_model
839 840 841 842 843

    @dygraph_only
    def state_dict(self):
        """
        Get state dict information from optimizer.
844
        (Only work in dygraph mode)
845 846 847 848 849 850 851

        Returns: 
            state_dict(dict) : dict contains all the Tensor used by optimizer

        Examples:
            .. code-block:: python

852 853 854 855 856
                import numpy as np
                import paddle
                from paddle.distributed import fleet

                fleet.init(is_collective=True)
857

858
                value = np.arange(26).reshape(2, 13).astype("float32")
1
123malin 已提交
859
                a = paddle.to_tensor(value)
860

861 862
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
863

864 865 866
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
                state_dict = adam.state_dict()
867 868 869 870 871 872 873 874
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.state_dict()

    @dygraph_only
    def set_state_dict(self, state_dict):
        """
        Load optimizer state dict.
875
        (Only work in dygraph mode)
876 877 878 879

        Args: 
            state_dict(dict) : Dict contains all the Tensor needed by optimizer

880 881
        Returns:
            None
882 883 884 885

        Examples:
            .. code-block:: python

886 887 888
                import numpy as np
                import paddle
                from paddle.distributed import fleet
889

890 891 892
                fleet.init(is_collective=True)

                value = np.arange(26).reshape(2, 13).astype("float32")
1
123malin 已提交
893
                a = paddle.to_tensor(value)
894

895 896
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
897

898 899 900
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
                state_dict = adam.state_dict()
1
123malin 已提交
901 902 903
                paddle.save(state_dict, "paddle_dy")
                para_state_dict = paddle.load("paddle_dy")
                adam.set_state_dict(para_state_dict)
904 905 906 907 908 909 910 911
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.set_state_dict(state_dict)

    @dygraph_only
    def set_lr(self, value):
        """
        Set the value of the learning rate manually in the optimizer. 
912
        (Only work in dygraph mode)
913

914 915 916
        Args:
            value (float|Tensor): the value of learning rate

917 918
        Returns: 
            None 
919 920 921 922

        Examples:
            .. code-block:: python

923 924 925
                import numpy as np
                import paddle
                from paddle.distributed import fleet
926

927
                fleet.init(is_collective=True)
928

929
                value = np.arange(26).reshape(2, 13).astype("float32")
1
123malin 已提交
930
                a = paddle.to_tensor(value)
931

932 933
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
934

935 936 937 938 939 940 941 942 943 944 945 946 947 948
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

                lr_list = [0.2, 0.3, 0.4, 0.5, 0.6]
                for i in range(5):
                    adam.set_lr(lr_list[i])
                    lr = adam.get_lr()
                    print("current lr is {}".format(lr))
                # Print:
                #    current lr is 0.2
                #    current lr is 0.3
                #    current lr is 0.4
                #    current lr is 0.5
                #    current lr is 0.6
949 950 951 952 953 954 955 956
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.set_lr(value)

    @dygraph_only
    def get_lr(self):
        """
        Get current step learning rate.
957
        (Only work in dygraph mode)
958 959 960 961 962

        Returns:
            float: The learning rate of the current step.

        Examples:
1
123malin 已提交
963

964 965
            .. code-block:: python

966 967 968 969 970
                import numpy as np
                import paddle
                from paddle.distributed import fleet

                fleet.init(is_collective=True)
971

972
                value = np.arange(26).reshape(2, 13).astype("float32")
1
123malin 已提交
973
                a = paddle.to_tensor(value)
974

975 976
                layer = paddle.nn.Linear(13, 5)
                adam = paddle.optimizer.Adam(learning_rate=0.01, parameters=layer.parameters())
977

978 979
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)
980

981 982
                lr = adam.get_lr()
                print(lr) # 0.01
983 984 985 986 987 988 989 990
        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.get_lr()

    @dygraph_only
    def step(self):
        """
        Execute the optimizer once.
991
        (Only work in dygraph mode)
992

993 994
        Returns:
            None
995 996

        Examples:
1
123malin 已提交
997

998 999
            .. code-block:: python

1000 1001 1002
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet
1003

1004 1005 1006 1007 1008
                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
1009

1010 1011
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
1012

1
123malin 已提交
1013
                # 1. initialize fleet environment
1014 1015
                fleet.init(is_collective=True)

1
123malin 已提交
1016
                # 2. create layer & optimizer
1017 1018 1019 1020 1021
                layer = LinearNet()
                loss_fn = nn.MSELoss()
                adam = paddle.optimizer.Adam(
                    learning_rate=0.001, parameters=layer.parameters())

1
123malin 已提交
1022
                # 3. get data_parallel model using fleet
1023 1024 1025
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

1
123malin 已提交
1026
                # 4. run layer
1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046
                inputs = paddle.randn([10, 10], 'float32')
                outputs = dp_layer(inputs)
                labels = paddle.randn([10, 1], 'float32')
                loss = loss_fn(outputs, labels)

                print("loss:", loss.numpy())

                loss.backward()

                adam.step()
                adam.clear_grad()


        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.step()

    @dygraph_only
    def clear_grad(self):
        """
1047 1048
        Clear the gradients of all optimized parameters for model.
        (Only work in dygraph mode)
1049

1050 1051
        Returns: 
            None
1052 1053

        Examples:
1
123malin 已提交
1054

1055 1056
            .. code-block:: python

1057 1058 1059
                import paddle
                import paddle.nn as nn
                from paddle.distributed import fleet
1060

1061 1062 1063 1064 1065
                class LinearNet(nn.Layer):
                    def __init__(self):
                        super(LinearNet, self).__init__()
                        self._linear1 = nn.Linear(10, 10)
                        self._linear2 = nn.Linear(10, 1)
1066

1067 1068
                    def forward(self, x):
                        return self._linear2(self._linear1(x))
1069

1
123malin 已提交
1070
                # 1. initialize fleet environment
1071 1072
                fleet.init(is_collective=True)

1
123malin 已提交
1073
                # 2. create layer & optimizer
1074 1075 1076 1077 1078
                layer = LinearNet()
                loss_fn = nn.MSELoss()
                adam = paddle.optimizer.Adam(
                    learning_rate=0.001, parameters=layer.parameters())

1
123malin 已提交
1079
                # 3. get data_parallel model using fleet
1080 1081 1082
                adam = fleet.distributed_optimizer(adam)
                dp_layer = fleet.distributed_model(layer)

1
123malin 已提交
1083
                # 4. run layer
1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099
                inputs = paddle.randn([10, 10], 'float32')
                outputs = dp_layer(inputs)
                labels = paddle.randn([10, 1], 'float32')
                loss = loss_fn(outputs, labels)

                print("loss:", loss.numpy())

                loss.backward()

                adam.step()
                adam.clear_grad()

        """
        # imitate target optimizer retrieval
        return self.user_defined_optimizer.clear_grad()

1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116
    def _get_amp_optimizer(self):
        # imitate target optimizer retrieval
        amp_optimizer = None
        for optimizer in self.strategy_compiler._get_applied_meta_optimizer():
            if hasattr(optimizer, 'amp_init'):
                amp_optimizer = optimizer
                break

        if amp_optimizer is None:
            if hasattr(self.user_defined_optimizer, 'amp_init'):
                amp_optimizer = self.user_defined_optimizer

        assert amp_optimizer is not None, \
            "amp_init can only be used when the amp(auto mixed precision) strategy is turned on."
        return amp_optimizer

    def get_loss_scaling(self):
1117 1118
        """Return the real-time loss scaling factor.
        """
1119 1120 1121
        amp_optimizer = self._get_amp_optimizer()
        return amp_optimizer.get_loss_scaling()

H
huangxu96 已提交
1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181
    def amp_init(self,
                 place,
                 scope=None,
                 test_program=None,
                 use_fp16_test=False):
        """
        Init the amp training, such as cast fp32 parameters to fp16 type.
  
        Args:
            place(CUDAPlace): place is used to initialize 
                fp16 parameters with fp32 values.
            scope(Scope): The scope is used to find fp32 parameters.
            test_program(Program): The program is used for testing.
            use_fp16_test(bool): Whether to use fp16 testing.
            
        Examples:
            .. code-block:: python

                import numpy as np
                import paddle
                import paddle.nn.functional as F
                paddle.enable_static()

                def run_example_code():
                    place = paddle.CUDAPlace(0)
                    exe = paddle.static.Executor(place)
                    data = paddle.static.data(name='X', shape=[None, 1, 28, 28], dtype='float32')
                    conv2d = paddle.static.nn.conv2d(input=data, num_filters=6, filter_size=3)
                    # 1) Use fp16_guard to control the range of fp16 kernels used.
                    with paddle.static.amp.fp16_guard():
                        bn = paddle.static.nn.batch_norm(input=conv2d, act="relu")
                        pool = F.max_pool2d(bn, kernel_size=2, stride=2)
                        hidden = paddle.static.nn.fc(pool, size=10)
                        loss = paddle.mean(hidden)
                    # 2) Create the optimizer and set `multi_precision` to True.
                    # Setting `multi_precision` to True can avoid the poor accuracy
                    # or the slow convergence in a way. 
                    optimizer = paddle.optimizer.Momentum(learning_rate=0.01, multi_precision=True)
                    # 3) These ops in `custom_black_list` will keep in the float32 computation type.
                    amp_list = paddle.static.amp.CustomOpLists(
                        custom_black_list=['pool2d'])
                    # 4) The entry of Paddle AMP.
                    # Enable pure fp16 training by setting `use_pure_fp16` to True.
                    optimizer = paddle.static.amp.decorate(
                        optimizer,
                        amp_list,
                        init_loss_scaling=128.0,
                        use_dynamic_loss_scaling=True,
                        use_pure_fp16=True)
                    # If you don't use the default_startup_program(), you sholud pass
                    # your defined `startup_program` into `minimize`.
                    optimizer.minimize(loss)
                    exe.run(paddle.static.default_startup_program())
                    # 5) Use `amp_init` after FP32 parameters initialization(such as `exe.run(startup_program)`).
                    # If you want to perform the testing process, you should pass `test_program` into `amp_init`.
                    optimizer.amp_init(place, scope=paddle.static.global_scope())
                    
                if paddle.is_compiled_with_cuda() and len(paddle.static.cuda_places()) > 0:
                    run_example_code()       
        """
1182
        amp_optimizer = self._get_amp_optimizer()
1183
        return amp_optimizer.amp_init(place, scope, test_program, use_fp16_test)
H
huangxu96 已提交
1184

D
Dong Daxiang 已提交
1185 1186 1187 1188 1189 1190 1191 1192 1193
    def _final_strategy(self):
        if "valid_strategy" not in self._context:
            print(
                "WARNING: You may need to call minimize function before this function is called"
            )
            return {}
        else:
            return self._context["valid_strategy"]

1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211
    def _get_applied_meta_list(self):
        if "applied_meta_list" not in self._context:
            print(
                "WARNING: You may need to call minimize function before _get_applied_meta_list called"
            )
            return []
        else:
            return self._context["applied_meta_list"]

    def _get_applied_graph_list(self):
        if "applied_graph_list" not in self._context:
            print(
                "WARNING: You may need to call minimize function before _get_applied_graph_list called"
            )
            return []
        else:
            return self._context["applied_graph_list"]

1212 1213 1214 1215 1216 1217 1218 1219 1220
    def minimize(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None):
        """
        Add distributed operations to minimize ``loss`` by updating ``parameter_list``.

        Args:
1
123malin 已提交
1221
            loss (Tensor): A ``Tensor`` containing the value to minimize.
1222 1223 1224
            startup_program (Program, optional): :ref:`api_fluid_Program` for
                initializing parameters in ``parameter_list``. The default value
                is None, at this time :ref:`api_fluid_default_startup_program` will be used.
1
123malin 已提交
1225
            parameter_list (Iterable, optional): Iterable of ``Tensor`` or ``Tensor.name`` to update
1226 1227
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
1
123malin 已提交
1228
            no_grad_set (set, optional): Set of ``Tensor``  or ``Tensor.name`` that don't need
1229 1230 1231 1232
                to be updated. The default value is None.

        Returns:
            tuple: tuple (optimize_ops, params_grads), A list of operators appended
1
123malin 已提交
1233
            by minimize and a list of (param, grad) tensor pairs, param is
1234
            ``Parameter``, grad is the gradient value corresponding to the parameter.
1235 1236
            The returned tuple can be passed to ``fetch_list`` in ``Executor.run()`` to
            indicate program pruning. If so, the program will be pruned by ``feed`` and
1237 1238 1239
            ``fetch_list`` before run, see details in ``Executor``.

        Examples:
1
123malin 已提交
1240

1241
            .. code-block:: python
1242

1243
                import paddle
1
123malin 已提交
1244
                paddle.enable_static()
1245
                import paddle.distributed.fleet as fleet
1
123malin 已提交
1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256
                import paddle.nn.functional as F

                hid_dim = 10
                label_dim = 2
                input_x = paddle.static.data(name='x', shape=[None, 13], dtype='float32')
                input_y = paddle.static.data(name='y', shape=[None, 1], dtype='int64')
                fc_1 = paddle.static.nn.fc(x=input_x, size=hid_dim, activation='tanh')
                fc_2 = paddle.static.nn.fc(x=fc_1, size=hid_dim, activation='tanh')
                prediction = paddle.static.nn.fc(x=[fc_2], size=label_dim, activation='softmax')
                cost = F.cross_entropy(input=prediction, label=input_y)
                avg_cost = paddle.mean(x=cost)
1257

1
123malin 已提交
1258
                fleet.init(is_collective=True)
1259 1260 1261 1262
                strategy = fleet.DistributedStrategy()
                optimizer = paddle.optimizer.SGD(learning_rate=0.001)
                optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
                optimizer.minimize(avg_cost)
1263

1264
                # for more examples, please reference https://github.com/PaddlePaddle/FleetX
1265 1266

        """
D
Dong Daxiang 已提交
1267 1268 1269
        context = {}
        context["user_defined_strategy"] = copy.deepcopy(
            self._user_defined_strategy)
1270 1271 1272
        if paddle.fluid.framework.in_dygraph_mode():
            # imitate target optimizer retrieval
            target_opt = self.user_defined_optimizer
D
Dong Daxiang 已提交
1273
            self._context = context
1274 1275
            return target_opt.minimize(loss)

1276 1277
        # cache original feed forward program
        self.origin_main_program = loss.block.program
1278 1279
        context["origin_main_program"] = self.origin_main_program
        context["loss"] = loss
1280 1281
        if startup_program == None:
            self.origin_startup_program = \
1282 1283
                paddle.static.default_startup_program().clone(for_test=False)
            startup_program = paddle.static.default_startup_program()
1284 1285 1286
        else:
            self.origin_startup_program = \
                startup_program.clone(for_test=False)
1287

1288 1289
        context["origin_startup_program"] = startup_program
        context["role_maker"] = self._role_maker
1290 1291 1292 1293 1294

        # compile time
        distributed_optimizer_list = \
            MetaOptimizerFactory()._get_valid_meta_optimizers(
                self.user_defined_optimizer)
D
Dong Daxiang 已提交
1295

D
Dong Daxiang 已提交
1296 1297 1298
        context["user_defined_strategy"] = copy.deepcopy(
            self._user_defined_strategy)
        copy_user_defined_strategy = copy.deepcopy(self._user_defined_strategy)
1299 1300 1301 1302 1303 1304

        # trigger the auto-parallel in very strict condition
        # strategy = DistributedStrategy()
        # strategy.auto = True
        # optimizer = paddle.optimizer.SGD(learning_rate=0.1)
        # optimizer = fleet.distributed_optimizer(optimizer, strategy)
D
Dong Daxiang 已提交
1305
        if copy_user_defined_strategy._is_strict_auto():
1306 1307
            # turn on all the strategy for each optimizer
            for opt in distributed_optimizer_list:
D
Dong Daxiang 已提交
1308
                opt._enable_strategy(copy_user_defined_strategy, context)
1309

1310 1311
        valid_optimizer_list = []
        valid_graph_optimizer_list = []
D
Dong Daxiang 已提交
1312
        can_not_apply_optimizer_list = []
1313 1314 1315 1316
        # recall meta optimizers for ranking
        for opt in distributed_optimizer_list:
            opt._set_basic_info(loss, self._role_maker,
                                self.user_defined_optimizer,
D
Dong Daxiang 已提交
1317
                                copy_user_defined_strategy)
1318 1319
            if opt._can_apply() and not opt._is_graph_out():
                valid_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1320
            elif opt._can_apply() and opt._is_graph_out():
1321
                valid_graph_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1322 1323
            else:
                can_not_apply_optimizer_list.append(opt)
1324
        # combine recalled meta optimizers to be a valid meta optimizer
D
Dong Daxiang 已提交
1325
        meta_optimizer, graph_optimizer = \
1326 1327
            self.strategy_compiler.generate_optimizer(
                loss, self._role_maker, self.user_defined_optimizer,
D
Dong Daxiang 已提交
1328
                copy_user_defined_strategy, valid_optimizer_list,
1329
                valid_graph_optimizer_list)
D
Dong Daxiang 已提交
1330

D
Dong Daxiang 已提交
1331
        valid_strategy = self.strategy_compiler._get_valid_strategy(
D
Dong Daxiang 已提交
1332 1333 1334
            copy_user_defined_strategy, can_not_apply_optimizer_list)

        context["valid_strategy"] = copy.deepcopy(valid_strategy)
1335

1336 1337 1338 1339 1340 1341
        applied_meta_list = self.strategy_compiler._get_applied_meta_list()
        applied_graph_list = self.strategy_compiler._get_applied_graph_list()

        context['applied_meta_list'] = applied_meta_list
        context['applied_graph_list'] = applied_graph_list

D
Dong Daxiang 已提交
1342
        self._context = context
1343

D
Dong Daxiang 已提交
1344
        self.valid_strategy = valid_strategy
1345
        self.valid_strategy._enable_env()
D
Dong Daxiang 已提交
1346

1347 1348
        optimize_ops = []
        params_grads = []
1349

1350 1351 1352 1353 1354 1355 1356 1357 1358
        if self._role_maker._is_non_distributed() and not self._is_collective:
            if self._runtime_handle is None:
                self._runtime_handle = RuntimeFactory()._create_runtime(context)

            compiled_program = compiler.CompiledProgram(
                self.origin_main_program).with_data_parallel(
                    loss_name=loss.name, share_vars_from=None)
            loss.block.program._graph = compiled_program
            return self.user_defined_optimizer.minimize(
M
MRXLT 已提交
1359
                loss, startup_program, parameter_list, no_grad_set=no_grad_set)
1360

1361 1362
        if meta_optimizer:
            optimize_ops, params_grads = meta_optimizer.minimize(
M
MRXLT 已提交
1363
                loss, startup_program, parameter_list, no_grad_set=no_grad_set)
1364

1365
            default_program = paddle.static.default_main_program()
1366 1367 1368 1369

            if id(default_program) != id(loss.block.program):
                paddle.fluid.framework.switch_main_program(loss.block.program)

1370 1371
        else:
            optimize_ops, params_grads = self.user_defined_optimizer.minimize(
M
MRXLT 已提交
1372
                loss, startup_program, parameter_list, no_grad_set=no_grad_set)
1373

1374 1375
        context["program_optimize_ops"] = optimize_ops
        context["program_params_grads"] = params_grads
1376

1377
        if graph_optimizer:
D
Dong Daxiang 已提交
1378
            optimize_ops, params_grads = graph_optimizer.minimize(
M
MRXLT 已提交
1379
                loss, startup_program, parameter_list, no_grad_set=no_grad_set)
1380 1381 1382 1383
            # since we do not encourage users to use graph operations
            # if a graph optimizer takes effect, mostly
            # optimizers_ops and params_grads are None
            # i.e. users can not modify current computation graph anymore
1384 1385 1386
            context["graph_optimize_ops"] = optimize_ops
            context["graph_optimize_grads"] = params_grads

1387
        if self._runtime_handle is None:
1388
            self._runtime_handle = RuntimeFactory()._create_runtime(context)
1389

1390 1391
        import paddle.distributed.fleet as fleet
        fleet.util._set_strategy(context["valid_strategy"])
1392 1393

        return optimize_ops, params_grads
1394 1395 1396 1397

    @dygraph_only
    def distributed_scaler(self, scaler):
        return HybridParallelGradScaler(scaler, self._hcg)