fleet.py 51.4 KB
Newer Older
W
wuhuachaocoding 已提交
1
#   Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14
#
# 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 copy
16
import warnings
17
import paddle
18
import os
19
from types import MethodType
20
import numpy as np
W
wuhuachaocoding 已提交
21
from paddle.fluid.framework import _global_flags
22
from paddle.fluid import compiler
W
wuhuachaocoding 已提交
23 24 25 26 27
from .base.role_maker import UserDefinedRoleMaker, PaddleCloudRoleMaker, RoleMakerBase
from .base.strategy_compiler import StrategyCompiler
from .base.distributed_strategy import DistributedStrategy
from .base.meta_optimizer_factory import MetaOptimizerFactory
from .base.runtime_factory import RuntimeFactory
28
from paddle.fluid.wrapped_decorator import wrap_decorator
29
from paddle.fluid.dygraph import parallel_helper
30
from paddle.fluid.ir import apply_build_strategy
W
wuhuachaocoding 已提交
31 32
from .base import topology as tp
from .meta_parallel import model_parallel_random_seed
33
from paddle import _C_ops, _legacy_C_ops
34
from paddle.fluid import core
35

36 37
__all__ = []

38

39 40 41 42 43 44 45 46 47 48 49 50 51 52
def apply_ir_passes(main_program, startup_program, config):
    build_strategy = config._user_defined_strategy.build_strategy._copy()
    if not _global_flags()['FLAGS_apply_pass_to_program']:
        return build_strategy

    pipeline_opt = getattr(main_program, "_pipeline_opt", {})
    if pipeline_opt:
        main_program = pipeline_opt["section_program"]
        startup_program = startup_program._pipeline_opt["startup_program"]

    pass_attrs = {"use_cuda": config._is_collective}
    fuse_all_reduce = config._user_defined_strategy.fuse_all_reduce_ops
    if fuse_all_reduce and build_strategy.fuse_all_optimizer_ops:
        # FIXME(zjl): currently, fuse_all_optimizer_ops
53 54 55 56
        # have conflict with fuse_all_reduce_ops because
        # RawProgramOptimizer also inserts coalesce_tensor
        # into program. These two procedures may conflict
        # in which vars are to be fused.
57 58 59 60 61 62 63 64 65
        warnings.warn(
            'Currently, the fuse_all_optimizer_ops pass has conflict with fuse_all_reduce_ops pass. Disable the fuse_all_optimizer_ops pass temporarily.'
        )
        build_strategy.fuse_all_optimizer_ops = False

    return apply_build_strategy(main_program, startup_program, build_strategy,
                                pass_attrs)


66
def _inited_runtime_handler_(func):
67

68 69 70 71 72 73 74 75 76 77 78
    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__


79
def _is_non_distributed_check_(func):
80

81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
    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__


96
inited_runtime_handler = wrap_decorator(_inited_runtime_handler_)
97
is_non_distributed_check = wrap_decorator(_is_non_distributed_check_)
98 99


100 101 102
class Fleet(object):
    """
    Unified API for distributed training of PaddlePaddle
103
    Please reference the https://github.com/PaddlePaddle/PaddleFleetX for details
104 105 106 107 108


    Returns:
        Fleet: A Fleet instance

109
    Example for collective training:
1
123malin 已提交
110

111 112
        .. code-block:: python

1
123malin 已提交
113 114
            import paddle
            paddle.enable_static()
115
            import paddle.distributed.fleet as fleet
116 117 118

            fleet.init(is_collective=True)

119 120 121
            strategy = fleet.DistributedStrategy()
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
            optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
122 123 124 125 126 127 128 129

            # do distributed training


    Example for parameter server training:

        .. code-block:: python

1
123malin 已提交
130 131
            import paddle
            paddle.enable_static()
132 133
            import paddle.distributed.fleet as fleet
            strategy = fleet.DistributedStrategy()
S
ShenLiang 已提交
134
            fleet.init(strategy=strategy)
135

136
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
137
            optimizer = fleet.distributed_optimizer(optimizer)
138

139 140
            if fleet.is_first_worker():
                print("this is first worker")
141

142 143
            print("current node index: {}".format(fleet.worker_index()))
            print("total number of worker num: {}".format(fleet.worker_num()))
144

145 146 147
            if fleet.is_worker():
                print("this is worker")
            print("worker endpoints: {}".format(fleet.worker_endpoints(to_string=True)))
148

149 150
            print("server num: {}".format(fleet.server_num()))
            print("server endpoints: {}".format(fleet.server_endpoints(to_string=True)))
151

152 153 154
            if fleet.is_server():
                print("this is server")
            fleet.stop_worker()
155 156


157 158 159
    """

    def __init__(self):
160
        self._role_maker = None
161
        self.strategy_compiler = None
162
        self._is_collective = False
163
        self._runtime_handle = None
D
Dong Daxiang 已提交
164 165
        self._util = None
        self._context = {}
W
wuhuachaocoding 已提交
166
        self.user_defined_optimizer = paddle.optimizer.Optimizer(0.0)
167

168
    def init(self, role_maker=None, is_collective=False, strategy=None):
169 170 171
        """
        Initialize role_maker in Fleet.

172 173 174 175 176
        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
177
                of environment variables related to distributed training.If you did not initialize
178 179
                the rolemaker by yourself, it will be automatically initialized to PaddleRoleMaker.
                The default value is None.
180
            is_collective (Boolean, optional): A ``Boolean`` variable determines whether the program
181
                runs on Collective mode or ParameterServer mode. True means the program runs on
182
                Collective mode, and False means running on ParameterServer mode. The default value
183
                is False.
184
            strategy (DistributedStrategy): Extra properties for distributed training.
185 186 187
                For details, please refer to paddle.distributed.fleet.DistributedStrategy. Default: None.


188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209
        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
210
                role = fleet.PaddleCloudRoleMaker()
211
                fleet.init(role)
212

213 214 215 216 217 218
        Examples4:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                strategy = fleet.DistributedStrategy()
S
ShenLiang 已提交
219
                fleet.init(strategy=strategy)
220

221
        """
S
ShenLiang 已提交
222 223 224
        if strategy is None:
            strategy = DistributedStrategy()
        self._user_defined_strategy = copy.deepcopy(strategy)
225 226

        if role_maker is None:
227 228 229 230 231 232
            if isinstance(is_collective, bool):
                self._is_collective = is_collective
                self._role_maker = PaddleCloudRoleMaker(
                    is_collective=self._is_collective)
            else:
                raise ValueError(
233 234
                    "`is_collective` should be instance of `bool`, but got {}".
                    format(type(is_collective)))
235
        else:
236 237
            if isinstance(role_maker, RoleMakerBase):
                self._role_maker = role_maker
238
                self._is_collective = role_maker._is_collective
239 240
            else:
                raise ValueError(
241 242
                    "`role_maker` should be subclass of `RoleMakerBase`, but got {}"
                    .format(type(role_maker)))
243
        self._role_maker._generate_role()
244

245 246 247
        import paddle.distributed.fleet as fleet
        fleet.util._set_role_maker(self._role_maker)

248
        self.strategy_compiler = StrategyCompiler()
249 250 251 252 253 254 255 256 257

        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."
                    )

J
Jiabin Yang 已提交
258
        if paddle.fluid.framework._non_static_mode():
259
            if self.worker_num() == 1:
260 261 262
                # if worker_num is 1, should construct default topology & hcg
                self._topology = tp.CommunicateTopology()
                self._hcg = tp.HybridCommunicateGroup(self._topology)
263
                return
264 265 266 267
            if parallel_helper._is_parallel_ctx_initialized():
                warnings.warn(
                    "The dygraph parallel environment has been initialized.")
            else:
268 269 270 271 272 273 274 275 276
                # 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)
277
                paddle.distributed.init_parallel_env()
278

K
kuizhiqing 已提交
279 280 281 282 283 284 285 286 287
            # hybrid parallel not support for npu/xpu
            if self._user_defined_strategy.heter_ccl_mode == False:
                # 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."
                    )
W
WangXi 已提交
288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
        elif self._is_collective:
            use_sharding = self._user_defined_strategy.sharding

            # global group
            global_rank = self.worker_index()
            global_world_size = self.worker_num()
            # NOTE(wangxi): see sharding_optimizer
            global_ring_id = 3 if use_sharding else 0
            global_ranks = list(range(global_world_size))

            if tp._HYBRID_PARALLEL_GROUP is None: tp._CommunicateGroup()
            cg = tp._HYBRID_PARALLEL_GROUP
            self._hcg = cg
            cg.set_comm_group('global', global_rank, global_world_size,
                              global_ring_id, global_ranks)

Y
Yuang Liu 已提交
304 305 306
            use_tensor_parallel = self._user_defined_strategy.tensor_parallel
            use_mp = use_sharding or use_tensor_parallel

W
WangXi 已提交
307
            # hybrid group
Y
Yuang Liu 已提交
308 309 310 311 312 313 314 315 316 317
            if use_mp is False: return

            mp_degree_sharding = 1
            mp_degree_tensor_parallel = 1
            if use_sharding:
                sharding_configs = self._user_defined_strategy.sharding_configs
                mp_degree_sharding = int(sharding_configs['mp_degree'])

            if use_tensor_parallel:
                tensor_parallel_configs = self._user_defined_strategy.tensor_parallel_configs
318 319
                mp_degree_tensor_parallel = int(
                    tensor_parallel_configs['tensor_parallel_degree'])
Y
Yuang Liu 已提交
320 321 322

            if use_sharding and use_tensor_parallel:
                assert mp_degree_sharding == mp_degree_tensor_parallel
W
WangXi 已提交
323

Y
Yuang Liu 已提交
324
            mp_degree = mp_degree_sharding if use_sharding else mp_degree_tensor_parallel
W
WangXi 已提交
325 326 327 328 329 330 331 332 333 334 335 336 337

            if mp_degree > 1:
                assert global_world_size % mp_degree == 0
                # NOTE(wangxi): mp_ring_id sync with sharding_optimizer.py _build_groups
                mp_ring_id = 0
                mp_rank = global_rank % mp_degree
                mp_group_id = global_rank // mp_degree
                mp_group_ranks = [
                    idx for idx in global_ranks
                    if idx // mp_degree == mp_group_id
                ]
                cg.set_comm_group('model', mp_rank, mp_degree, mp_ring_id,
                                  mp_group_ranks)
W
wuhuachaocoding 已提交
338
        return self
339 340 341 342 343 344 345 346

    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"]
J
JZ-LIANG 已提交
347
        self.sharding_degree = self.hybrid_configs["sharding_degree"]
348 349 350

        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"
J
JZ-LIANG 已提交
351
        assert self.sharding_degree >= 0, "sharding_degree should be greater or equal to 0"
352 353 354 355 356 357 358 359 360 361 362

        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(
J
JZ-LIANG 已提交
363 364 365 366 367
            hybrid_group_names=["data", "pipe", "sharding", "model"],
            dims=[
                self.dp_degree, self.pp_degree, self.sharding_degree,
                self.mp_degree
            ])
368 369 370

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

371 372 373 374 375 376 377 378
        if self.mp_degree > 1:
            tensor_parallel_configs = self._user_defined_strategy.tensor_parallel_configs
            tensor_init_seed = tensor_parallel_configs["tensor_init_seed"]
            if tensor_init_seed == -1:
                model_parallel_random_seed()
            else:
                model_parallel_random_seed(tensor_init_seed)

379 380 381 382 383 384 385 386
    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

387 388 389 390 391 392 393
    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.
394

395 396 397 398 399 400 401 402
        Examples:

            .. code-block:: python

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

403
        """
404
        return self._role_maker._is_first_worker()
405 406 407 408 409 410 411

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

        Returns:
            int: node id
412 413 414 415

        Examples:

            .. code-block:: python
1
123malin 已提交
416

417 418 419 420
                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.worker_index()

421
        """
422
        return self._role_maker._worker_index()
423 424 425 426 427 428 429

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

        Returns:
            int: worker numbers
1
123malin 已提交
430

431
        Examples:
1
123malin 已提交
432

433 434 435 436 437 438
            .. code-block:: python

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

439
        """
440
        return self._role_maker._worker_num()
441

442 443 444 445 446 447 448 449 450 451 452 453
    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()

454 455 456 457 458 459 460
    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.
461 462

        Examples:
1
123malin 已提交
463

464 465 466 467 468 469
            .. code-block:: python

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

470
        """
471
        return self._role_maker._is_worker()
472

473 474 475
    def is_coordinator(self):
        return self._role_maker._is_coordinator()

476 477
    def worker_endpoints(self, to_string=False):
        """
478
        Get current worker endpoints, such as ["127.0.0.1:1001", "127.0.0.1:1002"].
479 480 481

        Returns:
            list/string: server endpoints
482 483

        Examples:
1
123malin 已提交
484

485 486 487 488 489 490
            .. code-block:: python

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

491 492
        """
        if to_string:
493
            return ",".join(self._role_maker._get_trainer_endpoints())
494
        else:
495
            return self._role_maker._get_trainer_endpoints()
496 497 498 499 500 501 502

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

        Returns:
            int: server number
503 504

        Examples:
1
123malin 已提交
505

506
            .. code-block:: python
1
123malin 已提交
507 508 509 510

                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.server_num()
511
        """
512
        return len(self._role_maker._get_pserver_endpoints())
513 514 515 516 517 518 519

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

        Returns:
            int: node id
520 521

        Examples:
1
123malin 已提交
522

523 524 525 526 527 528
            .. code-block:: python

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

529
        """
530
        return self._role_maker._server_index()
531 532 533 534 535 536 537

    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
538 539

        Examples:
1
123malin 已提交
540

541 542 543 544 545 546
            .. code-block:: python

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

547
        """
548

549
        if to_string:
550
            return ",".join(self._role_maker._get_pserver_endpoints())
551
        else:
552
            return self._role_maker._get_pserver_endpoints()
553 554 555 556 557 558 559 560

    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.
561 562 563 564

        Examples:

            .. code-block:: python
1
123malin 已提交
565

566 567 568 569
                import paddle.distributed.fleet as fleet
                fleet.init()
                fleet.is_server()

570
        """
571 572
        return self._role_maker._is_server()

573 574
    def barrier_worker(self):
        """
575 576 577 578
        barrier all workers

        Returns:
            None
579
        """
580
        self._role_maker._barrier("worker")
581

582
    @is_non_distributed_check
583
    @inited_runtime_handler
584
    def init_worker(self, scopes=None):
585
        """
586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603
        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()

604
        """
605
        self._runtime_handle._init_worker(scopes)
606

607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625
    @is_non_distributed_check
    @inited_runtime_handler
    def init_coordinator(self, scopes=None):
        """
        initialize coordinator node
        """
        self._runtime_handle._init_coordinator(scopes)

    def make_fl_strategy(self):
        self._runtime_handle._make_fl_strategy()

    @is_non_distributed_check
    @inited_runtime_handler
    def get_fl_client(self):
        """
        get worker(training node) ptr
        """
        return self._runtime_handle._worker

626
    @is_non_distributed_check
627
    @inited_runtime_handler
628
    def init_server(self, *args, **kwargs):
629
        """
630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648
        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()

649
        """
650
        self._runtime_handle._init_server(*args, **kwargs)
651

Z
zmxdream 已提交
652 653
    @is_non_distributed_check
    @inited_runtime_handler
T
Thunderbrook 已提交
654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671
    def load_model(self, path, mode):
        """
        load fleet model from path


        Returns:
            None

        Examples:

            .. code-block:: python

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

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

672
                fleet.load_model("path", mode=0)
T
Thunderbrook 已提交
673 674

        """
675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725
        self._runtime_handle._load_persistables(path, mode)

    @is_non_distributed_check
    @inited_runtime_handler
    def load_one_table(self, table_id, path, mode):
        """
        load fleet one table from path


        Returns:
            None

        Examples:

            .. code-block:: python

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

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

                fleet.load_one_table(0, "path", mode=0)

        """
        self._runtime_handle._load_one_table(table_id, path, mode)

    @is_non_distributed_check
    @inited_runtime_handler
    def load_inference_model(self, path, mode):
        """
        load fleet inference model from path


        Returns:
            None

        Examples:

            .. code-block:: python

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

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

                fleet.load_inference_model("path", mode=1)

        """
        self._runtime_handle._load_inference_model(path, mode)
T
Thunderbrook 已提交
726

727
    @is_non_distributed_check
728
    @inited_runtime_handler
729 730
    def run_server(self):
        """
731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748
        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()

749 750 751
        """
        self._runtime_handle._run_server()

752
    @is_non_distributed_check
753
    @inited_runtime_handler
754 755
    def stop_worker(self):
        """
756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772
        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()

773 774 775
        """
        self._runtime_handle._stop_worker()

Z
zmxdream 已提交
776 777
    @is_non_distributed_check
    @inited_runtime_handler
T
tangwei12 已提交
778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811
    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
            ]

812 813 814 815
            self._runtime_handle._save_inference_model(executor, dirname,
                                                       feeded_var_names,
                                                       fetch_vars, None, True,
                                                       0)
T
tangwei12 已提交
816 817 818 819
        else:
            increment_mode = 0
            if "mode" in configs:
                increment_mode = int(configs["mode"])
820 821 822 823
            self._runtime_handle._save_persistables(executor,
                                                    dirname,
                                                    main_program=None,
                                                    mode=increment_mode)
T
tangwei12 已提交
824

Z
zmxdream 已提交
825 826
    @is_non_distributed_check
    @inited_runtime_handler
827 828 829 830 831 832
    def save_inference_model(self,
                             executor,
                             dirname,
                             feeded_var_names,
                             target_vars,
                             main_program=None,
833 834
                             export_for_deployment=True,
                             mode=0):
835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853
        """
        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
854 855 856
        # warnings.warn(
        #     "'save_inference_model' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead."
        # )
857

858 859 860 861
        self._runtime_handle._save_inference_model(executor, dirname,
                                                   feeded_var_names,
                                                   target_vars, main_program,
                                                   export_for_deployment, mode)
862

Z
zmxdream 已提交
863 864
    @is_non_distributed_check
    @inited_runtime_handler
865
    def save_persistables(self, executor, dirname, main_program=None, mode=0):
866 867
        """

1
123malin 已提交
868
        saves all persistable tensors from :code:`main_program` to
869 870
        the folder :code:`dirname`. You can refer to

1
123malin 已提交
871 872
        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
873 874 875
        files, set :code:`filename` None.

        Args:
1
123malin 已提交
876
            executor(Executor): The executor to run for saving persistable tensors.
877 878 879 880 881
                                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
882
            main_program(Program, optional): The program whose persistbale tensors will
883 884 885 886 887 888 889 890 891 892
                                             be saved. Default: None.


        Returns:
            None

        Examples:

            .. code-block:: text

1
123malin 已提交
893 894
                import paddle
                paddle.enable_static()
895 896 897 898 899 900 901
                import paddle.distributed.fleet as fleet

                fleet.init()

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

1
123malin 已提交
902 903
                exe = paddle.static.Executor(paddle.CPUPlace())
                fleet.save_persistables(exe, "dirname", paddle.static.default_main_program())
904 905

        """
T
tangwei12 已提交
906 907 908
        # warnings.warn(
        #     "'save_persistables' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead."
        # )
909

910 911
        self._runtime_handle._save_persistables(executor, dirname, main_program,
                                                mode)
912

Z
zhaocaibei123 已提交
913 914 915 916 917
    @is_non_distributed_check
    @inited_runtime_handler
    def save_cache_model(self, dirname, **configs):
        return self._runtime_handle._save_cache_model(dirname, **configs)

918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981
    @is_non_distributed_check
    @inited_runtime_handler
    def check_save_pre_patch_done(self):
        return self._runtime_handle._check_save_pre_patch_done()

    @is_non_distributed_check
    @inited_runtime_handler
    def save_one_table(self, table_id, path, mode):
        """
        save fleet one table from path


        Returns:
            None

        Examples:

            .. code-block:: python

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

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

                fleet.save_one_table(0, "path", mode=0)

        """
        self._runtime_handle._save_one_table(table_id, path, mode)

    @is_non_distributed_check
    @inited_runtime_handler
    def save_dense_params(self,
                          executor,
                          dirname,
                          scope,
                          program,
                          var_names=None):
        """
        save fleet one table from path


        Returns:
            None

        Examples:

            .. code-block:: python

                import paddle.distributed.fleet as fleet
                fleet.init()
                import paddle
                place = paddle.fluid.CPUPlace()
                exe = paddle.fluid.Executor(place)

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

                fleet.save_dense_params(exe, "path", scope=paddle.static.global_scope(), program=paddle.static.default_main_program())

        """
        self._runtime_handle._save_dense_params(executor, dirname, scope,
                                                program, var_names)

982
    def shrink(self, threshold=None):
983 984
        self._runtime_handle._shrink(threshold)

985
    def distributed_optimizer(self, optimizer, strategy=None):
986
        """
987 988 989 990 991 992 993
        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.
994
            strategy(DistributedStrategy): Extra properties for distributed optimizer.
995
                It is recommended to use DistributedStrategy in fleet.init(). The strategy
996 997
                here is for compatibility. If the strategy in fleet.distributed_optimizer()
                is not None, then it will overwrite the DistributedStrategy in fleet.init(),
998
                which will take effect in distributed training.
999

1000
        Returns:
1001
            Fleet: instance of fleet.
1002 1003

        Examples:
1004

1005
            .. code-block:: python
1006

1
123malin 已提交
1007
                import paddle
1008
                import paddle.distributed.fleet as fleet
1
123malin 已提交
1009
                fleet.init(is_collective=True)
1010 1011 1012 1013
                strategy = fleet.DistributedStrategy()
                optimizer = paddle.optimizer.SGD(learning_rate=0.001)
                optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)

1014 1015
        """
        self.user_defined_optimizer = optimizer
1016

1017
        if strategy is not None:
T
tangwei12 已提交
1018 1019 1020 1021 1022 1023 1024
            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.")
1025
            self._user_defined_strategy = copy.deepcopy(strategy)
D
Dong Daxiang 已提交
1026 1027

        self._context = {}
S
ShenLiang 已提交
1028

1029 1030
        return self

1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047
    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):
1048 1049
        """Return the real-time loss scaling factor.
        """
1050 1051 1052
        amp_optimizer = self._get_amp_optimizer()
        return amp_optimizer.get_loss_scaling()

H
huangxu96 已提交
1053 1054 1055 1056 1057 1058 1059
    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.
1060

H
huangxu96 已提交
1061
        Args:
1062
            place(CUDAPlace): place is used to initialize
H
huangxu96 已提交
1063 1064 1065 1066
                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.
1067

H
huangxu96 已提交
1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088
        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
1089
                    # or the slow convergence in a way.
H
huangxu96 已提交
1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108
                    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())
1109

H
huangxu96 已提交
1110
                if paddle.is_compiled_with_cuda() and len(paddle.static.cuda_places()) > 0:
1111
                    run_example_code()
H
huangxu96 已提交
1112
        """
1113
        amp_optimizer = self._get_amp_optimizer()
1114
        return amp_optimizer.amp_init(place, scope, test_program, use_fp16_test)
H
huangxu96 已提交
1115

D
Dong Daxiang 已提交
1116 1117 1118 1119 1120 1121 1122 1123 1124
    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"]

1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142
    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"]

1143 1144 1145 1146 1147 1148 1149 1150 1151
    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
1152
            loss (Tensor): A ``Tensor`` containing the value to minimize.
1153 1154 1155
            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
1156
            parameter_list (Iterable, optional): Iterable of ``Tensor`` or ``Tensor.name`` to update
1157 1158
                to minimize ``loss``. The default value is None, at this time all parameters
                will be updated.
1
123malin 已提交
1159
            no_grad_set (set, optional): Set of ``Tensor``  or ``Tensor.name`` that don't need
1160 1161 1162 1163
                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
1164
            by minimize and a list of (param, grad) tensor pairs, param is
1165
            ``Parameter``, grad is the gradient value corresponding to the parameter.
1166 1167
            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
1168 1169 1170
            ``fetch_list`` before run, see details in ``Executor``.

        Examples:
1
123malin 已提交
1171

1172
            .. code-block:: python
1173

1174
                import paddle
1
123malin 已提交
1175
                paddle.enable_static()
1176
                import paddle.distributed.fleet as fleet
1
123malin 已提交
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187
                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)
1188

1
123malin 已提交
1189
                fleet.init(is_collective=True)
1190 1191 1192 1193
                strategy = fleet.DistributedStrategy()
                optimizer = paddle.optimizer.SGD(learning_rate=0.001)
                optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
                optimizer.minimize(avg_cost)
1194

1195
                # for more examples, please reference https://github.com/PaddlePaddle/PaddleFleetX
1196 1197

        """
1198 1199 1200 1201
        if not isinstance(loss, list):
            return self._minimize_impl(loss, startup_program, parameter_list,
                                       no_grad_set)
        else:
J
Jiabin Yang 已提交
1202
            if paddle.fluid.framework._non_static_mode(
1203 1204 1205 1206 1207 1208 1209 1210 1211 1212
            ) or self._role_maker._is_non_distributed() or self._is_collective:
                raise ValueError("loss can be list only in PS mode")
            return self._minimize_losses_impl(loss, startup_program,
                                              parameter_list, no_grad_set)

    def _minimize_impl(self,
                       loss,
                       startup_program=None,
                       parameter_list=None,
                       no_grad_set=None):
D
Dong Daxiang 已提交
1213 1214 1215
        context = {}
        context["user_defined_strategy"] = copy.deepcopy(
            self._user_defined_strategy)
J
Jiabin Yang 已提交
1216
        if paddle.fluid.framework._non_static_mode():
1217 1218
            # imitate target optimizer retrieval
            target_opt = self.user_defined_optimizer
D
Dong Daxiang 已提交
1219
            self._context = context
1220 1221
            return target_opt.minimize(loss)

1222 1223
        # cache original feed forward program
        self.origin_main_program = loss.block.program
B
Baibaifan 已提交
1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239
        # add distributed attr
        if not hasattr(self.origin_main_program, "distributed_info_"):
            setattr(self.origin_main_program, "distributed_info_", dict())
            self.origin_main_program.distributed_info_[
                "dp_degree"] = self._user_defined_strategy.sharding_configs[
                    "dp_degree"]
            self.origin_main_program.distributed_info_[
                "mp_degree"] = self._user_defined_strategy.sharding_configs[
                    "mp_degree"]
            self.origin_main_program.distributed_info_[
                "pp_degree"] = self._user_defined_strategy.sharding_configs[
                    "pp_degree"]
            self.origin_main_program.distributed_info_[
                "sharding_degree"] = self._user_defined_strategy.sharding_configs[
                    "sharding_degree"]

1240
        context["origin_main_program"] = self.origin_main_program
1241
        context["origin_main_programs"] = [self.origin_main_program]
1242
        context["loss"] = loss
1243 1244
        if startup_program == None:
            self.origin_startup_program = \
1245 1246
                paddle.static.default_startup_program().clone(for_test=False)
            startup_program = paddle.static.default_startup_program()
1247 1248 1249
        else:
            self.origin_startup_program = \
                startup_program.clone(for_test=False)
1250

1251
        context["origin_startup_program"] = startup_program
1252
        context["origin_startup_programs"] = [startup_program]
1253
        context["role_maker"] = self._role_maker
1254

1255
        # Use the auto-parallel's routines instead
1256
        if self._user_defined_strategy.semi_auto or self._user_defined_strategy.auto_search:
W
wuhuachaocoding 已提交
1257
            from ..auto_parallel.parallelizer import AutoParallelizer
1258 1259 1260
            auto_parallelizer = AutoParallelizer(self)
            optimize_ops, params_grads, dist_startup_prog, dist_main_prog = auto_parallelizer.parallelize(
                loss, startup_program, parameter_list, no_grad_set)
1261

1262 1263
            return optimize_ops, params_grads, dist_startup_prog, dist_main_prog

1264 1265 1266 1267
        # compile time
        distributed_optimizer_list = \
            MetaOptimizerFactory()._get_valid_meta_optimizers(
                self.user_defined_optimizer)
D
Dong Daxiang 已提交
1268

D
Dong Daxiang 已提交
1269 1270 1271
        context["user_defined_strategy"] = copy.deepcopy(
            self._user_defined_strategy)
        copy_user_defined_strategy = copy.deepcopy(self._user_defined_strategy)
1272 1273 1274 1275 1276 1277

        # 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
1278
        if copy_user_defined_strategy._is_strict_auto():
1279 1280
            # turn on all the strategy for each optimizer
            for opt in distributed_optimizer_list:
D
Dong Daxiang 已提交
1281
                opt._enable_strategy(copy_user_defined_strategy, context)
1282

1283 1284
        valid_optimizer_list = []
        valid_graph_optimizer_list = []
D
Dong Daxiang 已提交
1285
        can_not_apply_optimizer_list = []
1286 1287 1288 1289
        # 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
1290
                                copy_user_defined_strategy)
1291 1292
            if opt._can_apply() and not opt._is_graph_out():
                valid_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1293
            elif opt._can_apply() and opt._is_graph_out():
1294
                valid_graph_optimizer_list.append(opt)
D
Dong Daxiang 已提交
1295 1296
            else:
                can_not_apply_optimizer_list.append(opt)
1297
        # combine recalled meta optimizers to be a valid meta optimizer
D
Dong Daxiang 已提交
1298
        meta_optimizer, graph_optimizer = \
1299 1300
            self.strategy_compiler.generate_optimizer(
                loss, self._role_maker, self.user_defined_optimizer,
D
Dong Daxiang 已提交
1301
                copy_user_defined_strategy, valid_optimizer_list,
1302
                valid_graph_optimizer_list)
D
Dong Daxiang 已提交
1303

D
Dong Daxiang 已提交
1304
        valid_strategy = self.strategy_compiler._get_valid_strategy(
D
Dong Daxiang 已提交
1305 1306 1307
            copy_user_defined_strategy, can_not_apply_optimizer_list)

        context["valid_strategy"] = copy.deepcopy(valid_strategy)
1308 1309
        # print("valid_strategy:", context["valid_strategy"])
        # print("user_defined_strategy:", context["user_defined_strategy"])
1310

1311 1312 1313 1314 1315 1316
        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
1317
        self._context = context
1318

D
Dong Daxiang 已提交
1319
        self.valid_strategy = valid_strategy
1320
        self.valid_strategy._enable_env()
D
Dong Daxiang 已提交
1321

1322 1323
        optimize_ops = []
        params_grads = []
1324

1325 1326 1327 1328 1329 1330 1331 1332
        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
1333 1334 1335 1336
            return self.user_defined_optimizer.minimize(loss,
                                                        startup_program,
                                                        parameter_list,
                                                        no_grad_set=no_grad_set)
1337

1338
        if meta_optimizer:
1339
            # print("before minimize program id:", id(loss.block.program))
1340
            optimize_ops, params_grads = meta_optimizer.minimize(
M
MRXLT 已提交
1341
                loss, startup_program, parameter_list, no_grad_set=no_grad_set)
1342
            # print("after minimize program id:", id(loss.block.program))
1343

1344
            default_program = paddle.static.default_main_program()
1345
            # print("default program id:", id(default_program))
1346 1347 1348

            if id(default_program) != id(loss.block.program):
                paddle.fluid.framework.switch_main_program(loss.block.program)
1349
            # print("default program id after switch:", id(default_program))
1350

1351 1352
        else:
            optimize_ops, params_grads = self.user_defined_optimizer.minimize(
M
MRXLT 已提交
1353
                loss, startup_program, parameter_list, no_grad_set=no_grad_set)
1354

1355 1356
        context["program_optimize_ops"] = optimize_ops
        context["program_params_grads"] = params_grads
1357

1358
        if graph_optimizer:
1359
            # print("before graph minimize program id:", id(loss.block.program))
D
Dong Daxiang 已提交
1360
            optimize_ops, params_grads = graph_optimizer.minimize(
M
MRXLT 已提交
1361
                loss, startup_program, parameter_list, no_grad_set=no_grad_set)
1362 1363 1364 1365
            # 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
1366 1367
            context["graph_optimize_ops"] = optimize_ops
            context["graph_optimize_grads"] = params_grads
1368 1369
        else:
            apply_ir_passes(loss.block.program, startup_program, self)
1370

1371 1372
        if not self._role_maker._is_heter_parameter_server_mode:
            program = paddle.static.default_main_program()
1373 1374 1375 1376 1377
            opt_info = {} if program._fleet_opt is None else program._fleet_opt
            opt_info["mpi_size"] = self.worker_num()
            opt_info["mpi_rank"] = self.worker_index()
            for k, v in self._user_defined_strategy.trainer_desc_configs.items(
            ):
1378
                if v or k not in opt_info:
1379
                    opt_info[k] = v
1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431
            program._fleet_opt = opt_info

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

        import paddle.distributed.fleet as fleet
        fleet.util._set_strategy(context["valid_strategy"])

        return optimize_ops, params_grads

    def _minimize_losses_impl(self,
                              losses,
                              startup_programs=None,
                              parameter_list=None,
                              no_grad_set=None):
        context = {}

        # cache original feed forward program
        self.origin_main_program = losses[0].block.program
        context["origin_main_program"] = self.origin_main_program
        context["origin_main_programs"] = []
        for loss in losses:
            context["origin_main_programs"].append(loss.block.program)
        context["loss"] = losses

        if startup_programs is None:
            if len(losses) == 1:
                startup_programs = [paddle.static.default_startup_program()]
            else:
                raise ValueError(
                    "startup_program can't be None when loss is list.")
        self.origin_startup_program = startup_programs[0].clone(for_test=False)
        context["origin_startup_program"] = startup_programs[0]
        context["origin_startup_programs"] = []
        for program in startup_programs:
            context["origin_startup_programs"].append(program)

        context["role_maker"] = self._role_maker

        context["user_defined_strategy"] = copy.deepcopy(
            self._user_defined_strategy)

        context["valid_strategy"] = copy.deepcopy(self._user_defined_strategy)

        self._context = context

        self.valid_strategy = context["valid_strategy"]
        self.valid_strategy._enable_env()

        optimize_ops = []
        params_grads = []

W
wuhuachaocoding 已提交
1432
        from .meta_optimizers import ParameterServerOptimizer
1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450
        ps_optimizer = ParameterServerOptimizer(self.user_defined_optimizer)
        ps_optimizer._set_basic_info(losses, self._role_maker,
                                     self.user_defined_optimizer,
                                     self._user_defined_strategy)
        optimize_ops, params_grads = ps_optimizer.minimize_losses_impl(
            losses, startup_programs, parameter_list, no_grad_set=no_grad_set)

        # default_program = paddle.static.default_main_program()

        # if id(default_program) != id(losses[0].block.program):
        #     paddle.fluid.framework.switch_main_program(losses[0].block.program)

        context["program_optimize_ops"] = optimize_ops
        context["program_params_grads"] = params_grads

        for loss in losses:
            program = loss.block.program
            opt_info = {} if program._fleet_opt is None else program._fleet_opt
1451 1452 1453 1454
            opt_info["mpi_size"] = self.worker_num()
            opt_info["mpi_rank"] = self.worker_index()
            for k, v in self._user_defined_strategy.trainer_desc_configs.items(
            ):
1455
                if v or k not in opt_info:
1456
                    opt_info[k] = v
1457
            program._fleet_opt = opt_info
1458
            # print("fleet base opt info:", id(program), program._fleet_opt)
1459

1460
        if self._runtime_handle is None:
1461
            self._runtime_handle = RuntimeFactory()._create_runtime(context)
1462

1463 1464
        import paddle.distributed.fleet as fleet
        fleet.util._set_strategy(context["valid_strategy"])
1465 1466

        return optimize_ops, params_grads