test_dist_base.py 63.6 KB
Newer Older
X
Xin Pan 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
#   Copyright (c) 2018 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.
14

15
import argparse
16
import ast
X
Xin Pan 已提交
17
import os
W
Wu Yi 已提交
18
import pickle
19
import random
20 21 22
import subprocess
import sys
import tempfile
23
import time
24 25 26
import unittest

import numpy as np
27 28

import paddle
29
import paddle.fluid as fluid
meteor135's avatar
meteor135 已提交
30
import paddle.incubate.distributed.fleet.role_maker as role_maker
31
from paddle.fluid import compiler
meteor135's avatar
meteor135 已提交
32 33 34 35
from paddle.incubate.distributed.fleet.collective import (
    DistributedStrategy,
    fleet,
)
36

Y
Yan Xu 已提交
37
RUN_STEP = 5
38
DEFAULT_BATCH_SIZE = 2
39
DIST_UT_PORT = 0
40

T
typhoonzero 已提交
41

42
def print_to_out(out_losses):
T
tianshuo78520a 已提交
43
    sys.stdout.buffer.write(pickle.dumps(out_losses))
44 45 46


def print_to_err(class_name, log_str):
47 48
    localtime = time.asctime(time.localtime(time.time()))
    print_str = localtime + "\t" + class_name + "\t" + log_str
T
tianshuo78520a 已提交
49
    sys.stderr.buffer.write(pickle.dumps(print_str))
G
guru4elephant 已提交
50 51


52 53 54 55
def eprint(*args, **kwargs):
    print(*args, file=sys.stderr, **kwargs)


56
class TestDistRunnerBase:
57 58 59 60 61 62 63 64
    def get_model(
        self,
        batch_size=DEFAULT_BATCH_SIZE,
        lr=0.1,
        single_device=False,
        use_dgc=False,
        dist_strategy=None,
    ):
T
typhoonzero 已提交
65
        raise NotImplementedError(
66 67
            "get_model should be implemented by child classes."
        )
T
typhoonzero 已提交
68

69
    @staticmethod
70 71 72 73 74 75 76 77 78 79 80
    def get_transpiler(
        trainer_id,
        main_program,
        pserver_endpoints,
        trainers,
        sync_mode,
        dc_asgd=False,
        current_endpoint=None,
        nccl_comm_num=1,
        hogwild_mode=False,
    ):
T
typhoonzero 已提交
81
        # NOTE: import fluid until runtime, or else forking processes will cause error.
82
        config = paddle.distributed.transpiler.DistributeTranspilerConfig()
W
Wu Yi 已提交
83
        config.enable_dc_asgd = dc_asgd
84
        config.sync_mode = sync_mode
T
tangwei12 已提交
85 86
        config.runtime_split_send_recv = hogwild_mode

87 88
        if nccl_comm_num > 1:
            config.nccl_comm_num = nccl_comm_num
89
        # config.runtime_split_send_recv = True
90
        t = paddle.distributed.transpiler.DistributeTranspiler(config=config)
91 92 93 94 95 96 97 98
        t.transpile(
            trainer_id=trainer_id,
            program=main_program,
            pservers=pserver_endpoints,
            trainers=trainers,
            sync_mode=sync_mode,
            current_endpoint=current_endpoint,
        )
T
typhoonzero 已提交
99 100
        return t

101 102 103 104 105
    @staticmethod
    def get_lr_scheduler(program):
        lr_sheduler = None
        if hasattr(program, 'lr_sheduler'):
            from paddle.optimizer.lr import LRScheduler
106

107 108 109 110
            lr_sheduler = program.lr_sheduler
            assert isinstance(lr_sheduler, LRScheduler), "must be LRScheduler"
        return lr_sheduler

W
Wu Yi 已提交
111
    def run_pserver(self, args):
W
Wu Yi 已提交
112
        self.lr = args.lr
113
        self.get_model(batch_size=args.batch_size)
114
        # NOTE: pserver should not call memory optimize
T
tangwei12 已提交
115

116 117 118 119 120 121 122 123 124
        t = self.get_transpiler(
            trainer_id=args.trainer_id,
            main_program=fluid.default_main_program(),
            pserver_endpoints=args.endpoints,
            trainers=args.trainers,
            sync_mode=args.sync_mode,
            dc_asgd=args.dc_asgd,
            hogwild_mode=args.hogwild,
        )
W
Wu Yi 已提交
125
        pserver_prog = t.get_pserver_program(args.current_endpoint)
126 127 128
        startup_prog = t.get_startup_program(
            args.current_endpoint, pserver_prog
        )
Y
Yancey1989 已提交
129

T
typhoonzero 已提交
130 131 132
        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        exe.run(startup_prog)
133
        print_to_err(type(self).__name__, "run pserver startup program done.")
T
typhoonzero 已提交
134
        exe.run(pserver_prog)
135
        print_to_err(type(self).__name__, "run pserver main program done.")
T
typhoonzero 已提交
136

137 138 139 140
    def run_pipeline_trainer(self, args):
        self.lr = args.lr

        dist_strategy = DistributedStrategy()
141 142 143 144 145 146 147 148 149 150 151
        (
            test_program,
            avg_cost,
            train_reader,
            test_reader,
            batch_acc,
            predict,
            data_loader,
        ) = self.get_model(
            batch_size=args.batch_size, dist_strategy=dist_strategy
        )
152 153 154 155 156 157 158 159 160 161 162 163 164

        device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
        eprint(type(self).__name__, "device_id: %d." % device_id)
        place = fluid.CUDAPlace(device_id)

        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
        eprint(type(self).__name__, "run worker startup program done.")

        data_loader.set_sample_list_generator(train_reader, place)
        data_loader.start()
        print_to_err(type(self).__name__, "begin to train on trainer")
        out_losses = []
165 166 167

        main_program = fluid.default_main_program()
        lr_sheduler = self.get_lr_scheduler(main_program)
168
        for i in range(RUN_STEP):
169
            loss = exe.run(main_program, fetch_list=[avg_cost])
170 171 172
            loss = loss[0] if loss else None
            out_losses.append(loss)
            print_to_err(type(self).__name__, "run step %d finished" % i)
173 174 175
            if lr_sheduler is not None:
                lr_sheduler.step()

176
        data_loader.reset()
177 178
        print_to_err(type(self).__name__, "trainer run finished")

T
tianshuo78520a 已提交
179
        sys.stdout.buffer.write(pickle.dumps(out_losses))
180

181 182 183 184 185 186 187 188 189 190 191
    def run_use_fleet_api_20_trainer(self, args):
        """
        1. remove codes for DistributedStrategy and leave the DistributedStrategy part to get_model()
        2. to run with fleet 2.0 api, set flags _use_fleet_api and _use_fleet_api_20 to True
        3. for now, not support test for model save
        """
        assert args.update_method == "nccl2" or "bkcl"

        self.lr = args.lr
        print_to_err("use_fleet 2.0", "fleet.node_num:")

192 193 194 195 196 197 198 199
        (
            test_program,
            avg_cost,
            train_reader,
            test_reader,
            batch_acc,
            predict,
        ) = self.get_model(batch_size=args.batch_size)
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216

        if fluid.core.is_compiled_with_cuda():
            device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
            place = fluid.CUDAPlace(device_id)
        elif fluid.core.is_compiled_with_xpu():
            device_id = int(os.getenv("FLAGS_selected_xpus", "0"))
            place = fluid.XPUPlace(device_id)
        else:
            raise ValueError(
                "fleet dygraph api must in paddlepaddle-xpu or paddlepaddle-gpu."
            )

        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
        eprint(type(self).__name__, "run worker startup program done.")

        feed_var_list = [
217 218
            var
            for var in fluid.default_main_program().global_block().vars.values()
219 220 221 222 223 224 225 226 227 228 229 230 231
            if var.is_data
        ]

        eprint("feed_var_list:", feed_var_list)

        if feed_var_list[0].name == 'label':
            feed_var_list = feed_var_list[::-1]

        feeder = fluid.DataFeeder(feed_var_list, place)
        reader_generator = train_reader()

        def get_data():
            origin_batch = next(reader_generator)
232 233 234 235
            if (
                paddle.distributed.get_world_size() == 1
                and args.update_method == 'gloo'
            ):  # Gloo single mode
X
xiongkun 已提交
236 237 238
                return origin_batch

            elif args.update_method != "local" and args.use_reader_alloc:
239 240 241 242 243 244 245 246 247 248
                new_batch = []
                for offset, item in enumerate(origin_batch):
                    if offset % 2 == args.trainer_id:
                        new_batch.append(item)
                return new_batch
            else:
                return origin_batch

        print_to_err(type(self).__name__, "begin to train on trainer")
        out_losses = []
249
        for i in range(RUN_STEP):
250 251 252 253 254
            (loss,) = exe.run(
                fluid.default_main_program(),
                fetch_list=[avg_cost.name],
                feed=feeder.feed(get_data()),
            )
255 256 257 258 259
            out_losses.append(loss[0])
            print_to_err(type(self).__name__, "run step %d finished" % i)
        print_to_err(type(self).__name__, "trainer run finished")
        print_to_err(type(self).__name__, "dist losses: {}".format(out_losses))

T
tianshuo78520a 已提交
260
        sys.stdout.buffer.write(pickle.dumps(out_losses))
261

262 263
    def run_use_fleet_api_trainer(self, args):
        assert args.update_method == "nccl2" or "bkcl"
264 265 266 267 268 269 270 271

        self.lr = args.lr

        exec_strategy = fluid.ExecutionStrategy()
        exec_strategy.num_threads = 1

        dist_strategy = DistributedStrategy()
        dist_strategy.exec_strategy = exec_strategy
T
tangwei12 已提交
272
        dist_strategy.fuse_memory_size = 1  # MB
273
        dist_strategy.fuse_laryer_size = 1
274 275 276 277
        if args.use_local_sgd:
            dist_strategy.use_local_sgd = True
        if args.ut4grad_allreduce:
            dist_strategy._ut4grad_allreduce = True
278 279
        if args.sync_batch_norm:
            dist_strategy.sync_batch_norm = True
280 281 282

        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
        fleet.init(role)
283
        print_to_err("use_fleet", "fleet.node_num:")
T
tangwei12 已提交
284 285
        # "fleet.node_id:", fleet.node_id(),
        # "fleet.trainer_num:", fleet.worker_num())
286

287 288 289 290 291 292 293 294 295 296
        (
            test_program,
            avg_cost,
            train_reader,
            test_reader,
            batch_acc,
            predict,
        ) = self.get_model(
            batch_size=args.batch_size, dist_strategy=dist_strategy
        )
297 298 299 300

        trainer_prog = fleet._origin_program
        dist_prog = fleet.main_program

301 302 303 304 305 306 307 308 309 310
        if fluid.core.is_compiled_with_cuda():
            device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
            place = fluid.CUDAPlace(device_id)
        elif fluid.core.is_compiled_with_xpu():
            device_id = int(os.getenv("FLAGS_selected_xpus", "0"))
            place = fluid.XPUPlace(device_id)
        else:
            raise ValueError(
                "fleet dygraph api must in paddlepaddle-xpu or paddlepaddle-gpu."
            )
311 312 313 314 315 316

        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
        eprint(type(self).__name__, "run worker startup program done.")

        feed_var_list = [
317 318
            var
            for var in trainer_prog.global_block().vars.values()
319 320 321
            if var.is_data
        ]

322 323 324 325 326 327 328
        eprint("feed_var_list:", feed_var_list)

        # tmp add this code to pass python35 gcc8 CI
        # Fixme(gongweibao, wangxi), need fix fleet api program order
        if feed_var_list[0].name == 'label':
            feed_var_list = feed_var_list[::-1]

329 330 331 332 333 334 335 336 337 338 339 340 341 342
        feeder = fluid.DataFeeder(feed_var_list, place)
        reader_generator = train_reader()

        def get_data():
            origin_batch = next(reader_generator)
            if args.update_method != "local" and args.use_reader_alloc:
                new_batch = []
                for offset, item in enumerate(origin_batch):
                    if offset % 2 == args.trainer_id:
                        new_batch.append(item)
                return new_batch
            else:
                return origin_batch

343
        print_to_err(type(self).__name__, "begin to train on trainer")
344
        out_losses = []
345
        for i in range(RUN_STEP):
346 347 348 349 350
            (loss,) = exe.run(
                dist_prog,
                fetch_list=[avg_cost.name],
                feed=feeder.feed(get_data()),
            )
351
            out_losses.append(loss[0])
352 353
            print_to_err(type(self).__name__, "run step %d finished" % i)
        print_to_err(type(self).__name__, "trainer run finished")
354

T
tianshuo78520a 已提交
355
        sys.stdout.buffer.write(pickle.dumps(out_losses))
356

357 358 359
        if args.save_model:
            model_save_dir = "/tmp"
            if fleet.worker_index() == 0:
360 361 362 363 364 365 366 367 368 369 370 371
                model_save_dir_fluid = os.path.join(
                    model_save_dir, "fluid_persistables"
                )
                model_save_dir_fleet = os.path.join(
                    model_save_dir, "fleet_persistables"
                )
                infer_save_dir_fluid = os.path.join(
                    model_save_dir, "fluid_infer"
                )
                infer_save_dir_fleet = os.path.join(
                    model_save_dir, "fleet_infer"
                )
372
            else:
373 374 375 376 377 378 379 380 381 382 383 384
                model_save_dir_fluid = os.path.join(
                    model_save_dir, "fluid_persistables_2"
                )
                model_save_dir_fleet = os.path.join(
                    model_save_dir, "fleet_persistables_2"
                )
                infer_save_dir_fluid = os.path.join(
                    model_save_dir, "fluid_infer_2"
                )
                infer_save_dir_fleet = os.path.join(
                    model_save_dir, "fleet_infer_2"
                )
385
            paddle.distributed.io.save_persistables(
386 387
                exe, model_save_dir_fluid, fleet._origin_program
            )
388 389
            fleet.save_persistables(executor=exe, dirname=model_save_dir_fleet)
            feeded_var_names = [var.name for var in feed_var_list]
390 391 392 393 394 395 396 397 398 399
            fluid.io.save_inference_model(
                infer_save_dir_fluid,
                feeded_var_names,
                [avg_cost],
                exe,
                fleet._origin_program,
            )
            fleet.save_inference_model(
                exe, infer_save_dir_fleet, feeded_var_names, [avg_cost]
            )
400

401
    def run_trainer(self, args):
W
Wu Yi 已提交
402
        self.lr = args.lr
W
Wu Yi 已提交
403
        if args.nccl2_reduce_layer_local_run:
404 405 406 407 408 409 410 411
            (
                test_program,
                avg_cost,
                train_reader,
                test_reader,
                batch_acc,
                predict,
            ) = self.get_model(batch_size=args.batch_size, single_device=True)
412
        elif args.use_dgc:
413 414 415 416 417 418 419 420
            (
                test_program,
                avg_cost,
                train_reader,
                test_reader,
                batch_acc,
                predict,
            ) = self.get_model(batch_size=args.batch_size, use_dgc=args.use_dgc)
W
Wu Yi 已提交
421
        else:
422 423 424 425 426 427 428 429
            (
                test_program,
                avg_cost,
                train_reader,
                test_reader,
                batch_acc,
                predict,
            ) = self.get_model(batch_size=args.batch_size)
430

W
Wu Yi 已提交
431
        if args.update_method == "pserver":
432
            print_to_err(
433
                type(self).__name__,
434 435 436 437 438 439 440 441 442 443 444
                "begin to run transpile on trainer with pserver mode",
            )
            t = self.get_transpiler(
                trainer_id=args.trainer_id,
                main_program=fluid.default_main_program(),
                pserver_endpoints=args.endpoints,
                trainers=args.trainers,
                sync_mode=args.sync_mode,
                dc_asgd=args.dc_asgd,
                hogwild_mode=args.hogwild,
            )
T
tangwei12 已提交
445

T
typhoonzero 已提交
446
            trainer_prog = t.get_trainer_program()
447
            print_to_err(
448
                type(self).__name__,
449 450 451 452 453 454
                "get trainer program done with pserver mode.",
            )
        elif (
            args.update_method == "nccl2"
            or args.update_method == "nccl2_reduce_layer"
        ):
W
Wu Yi 已提交
455
            # transpile for nccl2
456
            config = paddle.distributed.transpiler.DistributeTranspilerConfig()
W
Wu Yi 已提交
457
            config.mode = "nccl2"
458
            config.nccl_comm_num = args.nccl_comm_num
459 460
            if args.use_hallreduce:
                config.use_hierarchical_allreduce = True
461 462 463
                config.hierarchical_allreduce_inter_nranks = (
                    args.hallreduce_inter_nranks
                )
464
            print_to_err(
465
                type(self).__name__,
466 467
                "begin to run transpile on trainer with nccl2 mode",
            )
468 469 470
            nccl2_t = paddle.distributed.transpiler.DistributeTranspiler(
                config=config
            )
471 472 473 474 475 476 477
            nccl2_t.transpile(
                args.trainer_id,
                program=fluid.default_main_program(),
                startup_program=fluid.default_startup_program(),
                trainers=args.endpoints,
                current_endpoint=args.current_endpoint,
            )
478
            print_to_err(
479 480
                type(self).__name__, "get trainer program done. with nccl2 mode"
            )
W
Wu Yi 已提交
481
            trainer_prog = fluid.default_main_program()
T
typhoonzero 已提交
482
        else:
483
            print_to_err(
484
                type(self).__name__,
485 486
                "do nothing about main program, just use it",
            )
T
typhoonzero 已提交
487
            trainer_prog = fluid.default_main_program()
488
            print_to_err(type(self).__name__, "use main program done.")
T
typhoonzero 已提交
489

490 491 492
        # FIXME(gongwb):wait pserver initialization.
        time.sleep(1)

493
        if args.use_cuda:
494 495
            device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
            place = fluid.CUDAPlace(device_id)
496 497 498
        else:
            place = fluid.CPUPlace()

499 500
        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
501
        print_to_err(type(self).__name__, "run worker startup program done.")
T
typhoonzero 已提交
502

W
Wu Yi 已提交
503 504
        exec_strategy = fluid.ExecutionStrategy()
        exec_strategy.num_threads = 1
505

W
Wu Yi 已提交
506
        build_stra = fluid.BuildStrategy()
507 508 509
        # FIXME force disable enable_inplace and memory_optimize
        build_stra.enable_inplace = False
        build_stra.memory_optimize = False
W
Wu Yi 已提交
510

511 512 513 514
        if args.fuse_all_reduce is not None:
            sys.stderr.write('fuse_all_reduce={}'.format(args.fuse_all_reduce))
            build_stra.fuse_all_reduce_ops = args.fuse_all_reduce

T
tangwei12 已提交
515 516 517
        if args.hogwild:
            build_stra.async_mode = True

518 519 520
        if args.enable_backward_deps:
            build_stra.enable_backward_optimizer_op_deps = True

W
Wu Yi 已提交
521
        if args.use_reduce:
522 523 524
            build_stra.reduce_strategy = (
                fluid.BuildStrategy.ReduceStrategy.Reduce
            )
W
Wu Yi 已提交
525
        else:
526 527 528
            build_stra.reduce_strategy = (
                fluid.BuildStrategy.ReduceStrategy.AllReduce
            )
W
Wu Yi 已提交
529

W
Wu Yi 已提交
530
        pass_builder = None
X
Xin Pan 已提交
531
        if args.batch_merge_repeat > 1:
X
fix  
Xin Pan 已提交
532
            pass_builder = build_stra._finalize_strategy_and_create_passes()
533
            mypass = pass_builder.insert_pass(0, "multi_batch_merge_pass")
534
            mypass.set("num_repeats", args.batch_merge_repeat)
X
Xin Pan 已提交
535

536 537 538 539
        if (
            args.update_method == "nccl2"
            or args.update_method == "nccl2_reduce_layer"
        ):
540 541
            build_stra.num_trainers = len(args.endpoints.split(","))
            build_stra.trainer_id = args.trainer_id
W
Wu Yi 已提交
542
        else:
W
Wu Yi 已提交
543
            # case args.update_method == "nccl2_reduce_layer":
544 545
            build_stra.num_trainers = 1
            build_stra.trainer_id = 0
W
Wu Yi 已提交
546

547
        print_to_err(type(self).__name__, "begin to compile with data parallel")
X
Xin Pan 已提交
548
        binary = compiler.CompiledProgram(trainer_prog).with_data_parallel(
W
Wu Yi 已提交
549
            loss_name=avg_cost.name,
W
Wu Yi 已提交
550
            build_strategy=build_stra,
551 552
            exec_strategy=exec_strategy,
        )
553
        print_to_err(type(self).__name__, "program compiled with data parallel")
T
typhoonzero 已提交
554 555

        feed_var_list = [
556 557
            var
            for var in trainer_prog.global_block().vars.values()
T
typhoonzero 已提交
558 559 560 561
            if var.is_data
        ]

        feeder = fluid.DataFeeder(feed_var_list, place)
562
        reader_generator = train_reader()
T
typhoonzero 已提交
563

564 565
        def get_data():
            origin_batch = next(reader_generator)
W
Wu Yi 已提交
566
            if args.update_method != "local" and args.use_reader_alloc:
567 568 569 570 571 572 573
                new_batch = []
                for offset, item in enumerate(origin_batch):
                    if offset % 2 == args.trainer_id:
                        new_batch.append(item)
                return new_batch
            else:
                return origin_batch
T
typhoonzero 已提交
574

575
        lr_scheduler = self.get_lr_scheduler(trainer_prog)
576
        print_to_err(type(self).__name__, "begin to train on trainer")
W
Wu Yi 已提交
577
        out_losses = []
578
        for i in range(RUN_STEP):
579 580 581
            (loss,) = exe.run(
                binary, fetch_list=[avg_cost.name], feed=feeder.feed(get_data())
            )
W
Wu Yi 已提交
582
            out_losses.append(loss[0])
583
            print_to_err(type(self).__name__, "run step %d finished" % i)
584 585 586
            if lr_scheduler is not None:
                lr_scheduler.step()

587
        print_to_err(type(self).__name__, "trainer run finished")
588

589
        print_to_out(out_losses)
T
typhoonzero 已提交
590 591


592
class TestParallelDyGraphRunnerBase:
593 594
    def get_model(self):
        raise NotImplementedError(
595 596
            "get_model should be implemented by child classes."
        )
597 598 599

    def run_one_loop(self, model, opt, data):
        raise NotImplementedError(
600 601
            "train_one_loop should be implemented by the child classes."
        )
602

603
    def _get_data(self, batch, args):
604 605 606 607
        if (
            paddle.distributed.get_world_size() == 1
            and args.update_method == 'gloo'
        ):  # Gloo single mode
X
xiongkun 已提交
608 609
            return batch
        elif args.update_method != "local":
610
            new_batch = []
611

612 613 614
            # NOTE(@xiongkun03) args.diff_batch means batch length is different:
            # such as : batch = [2,3,4,5], then the first rank will get [2]  and
            # the second rank will get [3,4,5].
615 616
            # this function is for test sparse_embedding_differ_length
            if hasattr(args, "diff_batch") and args.diff_batch:
617 618 619
                assert (
                    len(batch) > 2
                ), "in differ_batch mode, len(batch) must > 2."
620 621 622 623 624 625 626 627 628 629 630 631 632 633
                if paddle.distributed.get_rank() == 0:
                    new_batch.append(batch[0])
                elif paddle.distributed.get_rank() == 1:
                    new_batch.extend([_ for _ in batch[1:]])
                else:
                    raise NotImplementedError(
                        "Current TestParallelDyGraphRunnerBase don't support world_size > 2"
                    )
                return new_batch
            else:
                for offset, item in enumerate(batch):
                    if offset % 2 == args.trainer_id:
                        new_batch.append(item)
                return new_batch
634 635 636
        else:
            return batch

637 638
    def run_trainer(self, args):
        seed = 90
X
xiongkun 已提交
639 640 641
        if args.update_method == 'gloo':
            place = fluid.CPUPlace()
        elif fluid.core.is_compiled_with_cuda():
642 643 644 645 646
            device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
            place = fluid.CUDAPlace(device_id)
        elif fluid.core.is_compiled_with_xpu():
            device_id = int(os.getenv("FLAGS_selected_xpus", "0"))
            place = fluid.XPUPlace(device_id)
647 648 649
        elif fluid.core.is_compiled_with_npu():
            device_id = int(os.getenv("FLAGS_selected_npus", "0"))
            place = fluid.NPUPlace(device_id)
650 651 652
        elif fluid.core.is_compiled_with_mlu():
            device_id = int(os.getenv("FLAGS_selected_mlus", "0"))
            place = fluid.MLUPlace(device_id)
653
        else:
654
            assert "Only support CUDAPlace or XPUPlace or CPU(Gloo) for now."
655 656 657 658

        with fluid.dygraph.guard(place):
            fluid.default_startup_program().random_seed = seed
            fluid.default_main_program().random_seed = seed
Y
Yan Xu 已提交
659 660
            np.random.seed(seed)
            import random
661

662
            random.seed(seed)
663 664
            model, train_reader, opt = self.get_model()
            nranks = len(args.endpoints.split(",")) if args.endpoints else 1
Y
Yan Xu 已提交
665

666 667 668 669 670 671 672
            # if args.update_method == "nccl2":
            if (
                args.update_method == "nccl2"
                or args.update_method == "bkcl"
                or args.update_method == "hccl"
                or args.update_method == "cncl"
            ):
Q
qizhaoaoe 已提交
673
                strategy = paddle.distributed.parallel.ParallelStrategy()
674 675 676 677
                strategy.nranks = nranks
                strategy.local_rank = args.trainer_id
                strategy.trainer_endpoints = args.endpoints.split(",")
                strategy.current_endpoint = args.current_endpoint
678
                paddle.distributed.init_parallel_env()
679
                print_to_err(
680
                    type(self).__name__,
681 682
                    "begin to prepare context in dygraph with nccl2",
                )
683
                if not args.find_unused_parameters:
Q
qizhaoaoe 已提交
684
                    model = paddle.DataParallel(
685 686
                        model, strategy, find_unused_parameters=False
                    )
687
                else:
Q
qizhaoaoe 已提交
688
                    model = paddle.DataParallel(
689 690
                        model, strategy, find_unused_parameters=True
                    )
691
                print_to_err(type(self).__name__, "model built in dygraph")
X
xiongkun 已提交
692 693 694 695

            elif args.update_method == "gloo":
                paddle.distributed.init_parallel_env()
                if not args.find_unused_parameters:
Q
qizhaoaoe 已提交
696
                    model = paddle.DataParallel(
697 698
                        model, find_unused_parameters=False
                    )
X
xiongkun 已提交
699
                else:
Q
qizhaoaoe 已提交
700
                    model = paddle.DataParallel(
701 702
                        model, find_unused_parameters=True
                    )
X
xiongkun 已提交
703

704
            out_losses = []
705
            print_to_err(type(self).__name__, "begin to run dygraph training")
706
            for step_id, data in enumerate(train_reader()):
707
                data = self._get_data(data, args)
708 709 710
                if step_id == RUN_STEP:
                    break
                loss = self.run_one_loop(model, opt, data)
G
guru4elephant 已提交
711
                if step_id % 10 == 0:
712
                    print_to_err(
713
                        type(self).__name__,
714 715
                        "loss at step %d: %f" % (step_id, loss.numpy()),
                    )
Y
Yan Xu 已提交
716
                out_losses.append(loss.numpy())
717 718 719 720

                loss.backward()

                opt.minimize(loss)
721 722
                if not args.accumulate_gradient:
                    model.clear_gradients()
723
        print_to_out(out_losses)
724

725 726 727 728 729 730 731 732 733
    def run_trainer_with_spawn(self, args):
        # 1. enable dygraph
        paddle.disable_static()

        # 2. init seed
        seed = 90
        paddle.static.default_startup_program().random_seed = seed
        paddle.static.default_main_program().random_seed = seed
        np.random.seed(seed)
734
        random.seed(seed)
735
        # get trainer id
L
LiYuRio 已提交
736 737
        paddle.distributed.parallel._get_global_parallel_env()
        args.trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
738 739

        # 3. init parallel env
X
xiongkun 已提交
740
        if args.update_method in ["nccl2", "gloo"]:
741 742 743 744
            paddle.distributed.init_parallel_env()

        # 4. train model
        model, train_reader, opt = self.get_model()
X
xiongkun 已提交
745
        if args.update_method in ["nccl2", "gloo"]:
746
            model = paddle.DataParallel(
747 748
                model, find_unused_parameters=args.find_unused_parameters
            )
749 750 751 752 753 754 755 756 757 758 759 760 761 762 763

        out_losses = []
        for step_id, data in enumerate(train_reader()):
            data = self._get_data(data, args)
            if step_id == RUN_STEP:
                break
            loss = self.run_one_loop(model, opt, data)
            out_losses.append(loss.numpy())

            loss.backward()

            opt.minimize(loss)
            model.clear_gradients()
        return out_losses

764
    def run_use_fleet_api_trainer(self, args):
765
        import paddle.distributed.fleet as fleet
766

767 768 769 770 771 772 773 774
        # 1. enable dygraph
        paddle.disable_static()

        # 2. init seed
        seed = 90
        paddle.static.default_startup_program().random_seed = seed
        paddle.static.default_main_program().random_seed = seed
        np.random.seed(seed)
775
        random.seed(seed)
776
        # get trainer id
L
LiYuRio 已提交
777 778
        paddle.distributed.parallel._get_global_parallel_env()
        args.trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
779

780 781
        # set strategy
        strategy = fleet.DistributedStrategy()
782 783
        if args.find_unused_parameters:
            strategy.find_unused_parameters = True
784

785
        # 3. init parallel env
786
        if args.update_method == "nccl2" or "bkcl" or "hccl":
787
            fleet.init(is_collective=True, strategy=strategy)
788 789 790

        # 4. train model
        model, train_reader, opt = self.get_model()
791
        if args.update_method == "nccl2" or "bkcl" or "hccl":
792 793 794 795 796 797 798 799 800 801 802 803 804 805
            opt = fleet.distributed_optimizer(opt)
            model = fleet.distributed_model(model)

        out_losses = []
        for step_id, data in enumerate(train_reader()):
            data = self._get_data(data, args)
            if step_id == RUN_STEP:
                break
            loss = self.run_one_loop(model, opt, data)
            out_losses.append(loss.numpy())

            loss.backward()

            opt.step()
806 807
            if not args.accumulate_gradient:
                opt.clear_grad()
808 809
        print_to_out(out_losses)

810

T
typhoonzero 已提交
811
def runtime_main(test_class):
W
Wu Yi 已提交
812
    parser = argparse.ArgumentParser(description='Run dist test.')
813 814 815
    parser.add_argument(
        '--role', type=str, required=True, choices=['pserver', 'trainer']
    )
W
Wu Yi 已提交
816
    parser.add_argument('--endpoints', type=str, required=False, default="")
817 818 819 820 821 822 823 824 825 826 827 828 829 830 831
    parser.add_argument(
        '--update_method',
        type=str,
        default="local",
        choices=[
            "pserver",
            "nccl2",
            "bkcl",
            "local",
            "nccl2_reduce_layer",
            "gloo",
            "hccl",
            "cncl",
        ],
    )
W
Wu Yi 已提交
832 833
    parser.add_argument('--trainer_id', type=int, required=False, default=0)
    parser.add_argument('--trainers', type=int, required=False, default=1)
834
    parser.add_argument('--nccl_comm_num', type=int, required=False, default=1)
835 836
    parser.add_argument('--enable_backward_deps', action='store_true')
    parser.add_argument('--use_hallreduce', action='store_true')
837
    parser.add_argument('--use_pipeline', action='store_true')
838
    parser.add_argument('--use_fleet_api', action='store_true')
839
    parser.add_argument('--use_fleet_api_20', action='store_true')
840
    parser.add_argument('--use_local_sgd', action='store_true')
841
    parser.add_argument('--diff_batch', action='store_true')
842
    parser.add_argument('--ut4grad_allreduce', action='store_true')
843 844 845 846 847 848
    parser.add_argument(
        '--hallreduce_inter_nranks', type=int, required=False, default=2
    )
    parser.add_argument(
        '--current_endpoint', type=str, required=False, default=""
    )
W
Wu Yi 已提交
849
    parser.add_argument('--sync_mode', action='store_true')
850
    parser.add_argument('--use_cuda', action='store_true')
X
xiongkun 已提交
851
    parser.add_argument('--use_cpu', action='store_true')
852
    parser.add_argument('--use_xpu', action='store_true')
853
    parser.add_argument('--use_dgc', action='store_true')
854
    parser.add_argument('--use_npu', action='store_true')
855
    parser.add_argument('--use_mlu', action='store_true')
856
    parser.add_argument('--accumulate_gradient', action='store_true')
857
    parser.add_argument('--find_unused_parameters', action='store_true')
W
Wu Yi 已提交
858
    parser.add_argument('--use_reduce', action='store_true')
W
Wu Yi 已提交
859
    parser.add_argument('--dc_asgd', action='store_true')
T
tangwei12 已提交
860
    parser.add_argument('--hogwild', action='store_true')
861
    parser.add_argument('--save_model', action='store_true')
862 863 864
    parser.add_argument(
        '--use_reader_alloc', action='store_true', required=False
    )
865
    parser.add_argument('--batch_size', required=False, type=int, default=2)
W
Wu Yi 已提交
866
    parser.add_argument('--lr', required=False, type=float, default=0.001)
867 868 869 870 871 872 873 874 875
    parser.add_argument(
        '--batch_merge_repeat', required=False, type=int, default=1
    )
    parser.add_argument(
        '--nccl2_reduce_layer_local_run',
        required=False,
        type=bool,
        default=False,
    )
876
    parser.add_argument('--sync_batch_norm', action='store_true')
877 878 879
    parser.add_argument(
        '--fuse_all_reduce', required=False, type=ast.literal_eval, default=None
    )
W
Wu Yi 已提交
880 881

    args = parser.parse_args()
T
typhoonzero 已提交
882

X
xiongkun 已提交
883 884 885
    if args.update_method == 'gloo':
        paddle.set_device("cpu")

T
typhoonzero 已提交
886
    model = test_class()
W
Wu Yi 已提交
887
    if args.role == "pserver" and args.update_method == "pserver":
W
Wu Yi 已提交
888
        model.run_pserver(args)
889 890
    elif args.use_fleet_api:
        model.run_use_fleet_api_trainer(args)
891 892
    elif args.use_fleet_api_20:
        model.run_use_fleet_api_20_trainer(args)
893 894
    elif args.use_pipeline:
        model.run_pipeline_trainer(args)
T
typhoonzero 已提交
895
    else:
896
        model.run_trainer(args)
X
Xin Pan 已提交
897

M
minqiyang 已提交
898

Y
Yancey1989 已提交
899 900
import socket
from contextlib import closing
M
minqiyang 已提交
901

X
Xin Pan 已提交
902 903

class TestDistBase(unittest.TestCase):
W
Wu Yi 已提交
904 905 906
    def _setup_config(self):
        raise NotImplementedError("tests should have _setup_config implemented")

907 908 909
    def _after_setup_config(self):
        if self._enforce_place == "CPU":
            self.__use_cuda = False
910
            self.__use_xpu = False
911
            self._use_dgc = False
912
            self.__use_npu = False
913
            self._use_mlu = False
914 915
        elif self._enforce_place == "GPU":
            self.__use_cuda = True
916
            self.__use_xpu = False
917
            self.__use_npu = False
918
            self._use_mlu = False
919 920 921 922
        elif self._enforce_place == "XPU":
            self.__use_cuda = False
            self.__use_xpu = True
            self._use_dgc = False
923
            self.__use_npu = False
924
            self._use_mlu = False
925 926 927 928 929
        elif self._enforce_place == "NPU":
            self.__use_cuda = False
            self.__use_xpu = False
            self._use_dgc = False
            self.__use_npu = True
930 931 932 933 934 935 936
            self._use_mlu = False
        elif self._enforce_place == "MLU":
            self.__use_cuda = False
            self.__use_xpu = False
            self._use_dgc = False
            self.__use_npu = False
            self._use_mlu = True
937 938 939 940 941
        else:
            if fluid.core.is_compiled_with_cuda():
                self.__use_cuda = True
            else:
                self.__use_cuda = False
942 943 944 945
                self._use_dgc = False

        if self._use_reduce:
            assert not self._use_dgc
946

X
Xin Pan 已提交
947 948 949
    def setUp(self):
        self._trainers = 2
        self._pservers = 2
Y
Yancey1989 已提交
950
        self._port_set = set()
M
minqiyang 已提交
951
        self._python_interp = sys.executable
W
Wu Yi 已提交
952
        self._sync_mode = True
T
tangwei12 已提交
953
        self._hogwild_mode = False
954
        self._enforce_place = None
W
Wu Yi 已提交
955
        self._use_reduce = False
W
Wu Yi 已提交
956
        self._dc_asgd = False  # must use with async mode
957
        self._use_reader_alloc = True
W
Wu Yi 已提交
958
        self._nccl2_mode = False
959
        self._bkcl_mode = False
X
xiongkun 已提交
960
        self._gloo_mode = False  # now, support gloo backend
961
        self._hccl_mode = False
962
        self._cncl_mode = False
963
        self._pipeline_mode = False
964
        self._mp_mode = False
965
        self._diff_batch = False
W
Wu Yi 已提交
966 967 968 969 970
        # FIXME(typhoonzero): I added this stupid argument to enable
        # testing allreduce layers, which users can call layers.allreduce
        # to accumulate tensors at anywhere. Find a better way to do this
        # test, reduce check this argument everywhere.
        self._nccl2_reduce_layer = False
W
Wu Yi 已提交
971
        self._lr = 0.001
972
        self._use_dgc = False
973
        self._dygraph = False
974
        self._nccl_comm_num = 1
975
        self._enable_backward_deps = False
976
        self._use_fleet_api = False
977
        self._use_fleet_api_20 = False
978 979
        self._use_local_sgd = False
        self._ut4grad_allreduce = False
980
        self._use_hallreduce = False
981
        self._save_model = False
982
        self._fuse_all_reduce = None
983
        self._accumulate_gradient = False
984
        self._find_unused_parameters = False
W
Wu Yi 已提交
985
        self._setup_config()
986 987 988 989 990 991 992

        global DIST_UT_PORT
        if DIST_UT_PORT == 0 and os.getenv("PADDLE_DIST_UT_PORT"):
            DIST_UT_PORT = int(os.getenv("PADDLE_DIST_UT_PORT"))

        if DIST_UT_PORT == 0:
            self._ps_endpoints = "127.0.0.1:%s,127.0.0.1:%s" % (
993 994 995
                self._find_free_port(),
                self._find_free_port(),
            )
996 997
        else:
            self._ps_endpoints = "127.0.0.1:%s,127.0.0.1:%s" % (
998 999 1000
                DIST_UT_PORT,
                DIST_UT_PORT + 1,
            )
1001
            DIST_UT_PORT += 2
1002
            self._dist_port = DIST_UT_PORT
1003

1004
        self._after_setup_config()
X
Xin Pan 已提交
1005

1006 1007 1008 1009 1010
        self.temp_dir = tempfile.TemporaryDirectory()

    def tearDown(self):
        self.temp_dir.cleanup()

Y
Yancey1989 已提交
1011
    def _find_free_port(self):
Y
Yancey1989 已提交
1012
        def __free_port():
1013 1014 1015
            with closing(
                socket.socket(socket.AF_INET, socket.SOCK_STREAM)
            ) as s:
Y
Yancey1989 已提交
1016
                s.bind(('', 0))
1017
                print_to_err(
1018 1019
                    type(self).__name__, "socket name: %s" % s.getsockname()[1]
                )
Y
Yancey1989 已提交
1020 1021 1022 1023 1024 1025 1026
                return s.getsockname()[1]

        while True:
            port = __free_port()
            if port not in self._port_set:
                self._port_set.add(port)
                return port
Y
Yancey1989 已提交
1027

1028 1029 1030
    def start_pserver(
        self, model_file, check_error_log, required_envs, log_name=""
    ):
X
Xin Pan 已提交
1031
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
1032 1033 1034 1035 1036 1037 1038 1039
        ps_cmd = "%s"

        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            required_envs['COVERAGE_FILE'] = os.getenv('COVERAGE_FILE', '')
            ps_cmd += " -m coverage run --branch -p"

        ps_cmd += " %s --role pserver --endpoints %s --trainer_id 0 --current_endpoint %s --trainers %d --update_method pserver"

1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053
        ps0_cmd = ps_cmd % (
            self._python_interp,
            model_file,
            self._ps_endpoints,
            ps0_ep,
            self._trainers,
        )
        ps1_cmd = ps_cmd % (
            self._python_interp,
            model_file,
            self._ps_endpoints,
            ps1_ep,
            self._trainers,
        )
W
Wu Yi 已提交
1054 1055 1056 1057

        if self._sync_mode:
            ps0_cmd += " --sync_mode"
            ps1_cmd += " --sync_mode"
X
Xin Pan 已提交
1058

1059 1060
        print(ps0_cmd)
        print(ps1_cmd)
1061 1062 1063 1064
        path0 = os.path.join(self.temp_dir.name, log_name + "_ps0_err.log")
        path1 = os.path.join(self.temp_dir.name, log_name + "_ps1_err.log")
        ps0_pipe = open(path0, "wb")
        ps1_pipe = open(path1, "wb")
G
gongweibao 已提交
1065

1066
        print_to_err(type(self).__name__, "going to start pserver process 0")
1067 1068 1069 1070 1071 1072
        ps0_proc = subprocess.Popen(
            ps0_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps0_pipe,
            env=required_envs,
        )
1073
        print_to_err(type(self).__name__, "going to start pserver process 1")
1074 1075 1076 1077 1078 1079
        ps1_proc = subprocess.Popen(
            ps1_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps1_pipe,
            env=required_envs,
        )
G
gongweibao 已提交
1080

1081
        return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe
X
Xin Pan 已提交
1082

1083 1084 1085 1086 1087 1088 1089 1090 1091 1092
    def _run_local(
        self,
        model,
        envs,
        check_error_log=False,
        batch_size=DEFAULT_BATCH_SIZE,
        batch_merge_repeat=1,
        log_name="",
        devices="1",
    ):
G
gongweibao 已提交
1093

1094 1095 1096 1097 1098 1099
        cmd = self._python_interp

        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            envs['COVERAGE_FILE'] = os.getenv('COVERAGE_FILE', '')
            cmd += " -m coverage run --branch -p"

1100 1101 1102 1103
        cmd += " %s --role trainer --update_method local --lr %f" % (
            model,
            self._lr,
        )
1104

1105 1106 1107 1108
        if batch_size != DEFAULT_BATCH_SIZE:
            cmd += " --batch_size %d" % batch_size
        if batch_merge_repeat > 1:
            cmd += " --batch_merge_repeat %d" % batch_merge_repeat
W
Wu Yi 已提交
1109 1110
        if self._nccl2_reduce_layer:
            cmd += " --nccl2_reduce_layer_local_run 1"
1111

1112
        if self.__use_cuda:
1113
            cmd += " --use_cuda"
W
Wu Yi 已提交
1114
            env_local = {
1115 1116
                "CUDA_VISIBLE_DEVICES": devices,
                "PADDLE_TRAINERS_NUM": "1",
1117
                "PADDLE_TRAINER_ID": "0",
1118 1119 1120 1121 1122
            }
        elif self.__use_xpu:
            cmd += " --use_xpu"
            env_local = {
                "FLAGS_selected_xpus": devices,
W
Wu Yi 已提交
1123
                "PADDLE_TRAINERS_NUM": "1",
1124
                "PADDLE_TRAINER_ID": "0",
W
Wu Yi 已提交
1125
            }
1126 1127 1128 1129 1130
        elif self.__use_npu:
            cmd += " --use_npu"
            env_local = {
                "FLAGS_selected_npus": devices,
                "PADDLE_TRAINERS_NUM": "1",
1131
                "PADDLE_TRAINER_ID": "0",
1132
            }
1133 1134 1135
        else:
            env_local = {'CPU_NUM': '1'}

1136
        # not use dgc in single card
1137
        if len(devices) > 1 and self._use_dgc:
1138 1139
            cmd += " --use_dgc"

1140 1141 1142
        if self._accumulate_gradient:
            cmd += " --accumulate_gradient"

1143 1144 1145
        if self._find_unused_parameters:
            cmd += " --find_unused_parameters"

W
Wu Yi 已提交
1146 1147
        env_local.update(envs)
        print("local_cmd: {}, env: {}".format(cmd, env_local))
G
gongweibao 已提交
1148

1149
        if check_error_log:
1150 1151
            path = os.path.join(self.temp_dir.name, log_name + "_local.log")
            err_log = open(path, "wb")
1152 1153 1154 1155 1156 1157
            local_proc = subprocess.Popen(
                cmd.split(" "),
                stdout=subprocess.PIPE,
                stderr=err_log,
                env=env_local,
            )
G
gongweibao 已提交
1158
        else:
1159 1160 1161 1162 1163 1164
            local_proc = subprocess.Popen(
                cmd.split(" "),
                stdout=subprocess.PIPE,
                stderr=subprocess.PIPE,
                env=env_local,
            )
G
gongweibao 已提交
1165

1166 1167 1168 1169 1170 1171
        local_out, local_err = local_proc.communicate()

        if check_error_log:
            err_log.close()

        sys.stderr.write('local_stderr: %s\n' % local_err)
W
Wu Yi 已提交
1172
        sys.stderr.write('local_stdout: %s\n' % pickle.loads(local_out))
X
Xin Pan 已提交
1173

W
Wu Yi 已提交
1174
        return pickle.loads(local_out)
1175

1176 1177 1178 1179 1180 1181 1182 1183 1184 1185
    def _run_local_gloo(
        self,
        model,
        envs,
        check_error_log=False,
        batch_size=DEFAULT_BATCH_SIZE,
        batch_merge_repeat=1,
        log_name="",
        devices="0",
    ):
X
xiongkun 已提交
1186 1187
        saved_endpoints = self._ps_endpoints
        self._ps_endpoints = self._ps_endpoints.split(',')[0]
1188 1189 1190
        result = self._run_cluster_gloo(
            model, envs, 'gloo', check_error_log, log_name
        )
X
xiongkun 已提交
1191 1192 1193
        self._ps_endpoints = saved_endpoints
        return result

1194
    def _run_cluster(self, model, envs, check_error_log, log_name):
X
Xin Pan 已提交
1195
        # Run dist train to compare with local results
1196 1197 1198
        ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(
            model, check_error_log, envs, log_name=log_name
        )
W
Wu Yi 已提交
1199

X
Xin Pan 已提交
1200
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
1201

1202 1203 1204 1205 1206 1207 1208 1209
        tr_cmd = "%s"

        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            envs['COVERAGE_FILE'] = os.getenv('COVERAGE_FILE', '')
            tr_cmd += " -m coverage run --branch -p"

        tr_cmd += " %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --update_method pserver --lr %f"

1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227
        tr0_cmd = tr_cmd % (
            self._python_interp,
            model,
            self._ps_endpoints,
            0,
            ps0_ep,
            self._trainers,
            self._lr,
        )
        tr1_cmd = tr_cmd % (
            self._python_interp,
            model,
            self._ps_endpoints,
            1,
            ps1_ep,
            self._trainers,
            self._lr,
        )
W
Wu Yi 已提交
1228 1229 1230 1231

        if self._sync_mode:
            tr0_cmd += " --sync_mode"
            tr1_cmd += " --sync_mode"
T
tangwei12 已提交
1232 1233 1234
        if self._hogwild_mode:
            tr0_cmd += " --hogwild"
            tr1_cmd += " --hogwild"
W
Wu Yi 已提交
1235 1236 1237
        if self._use_reduce:
            tr0_cmd += " --use_reduce"
            tr1_cmd += " --use_reduce"
1238 1239 1240
        if self._use_reader_alloc:
            tr0_cmd += " --use_reader_alloc"
            tr1_cmd += " --use_reader_alloc"
1241
        if self.__use_cuda:
1242 1243 1244 1245 1246 1247 1248 1249 1250 1251
            tr0_cmd += " --use_cuda"
            tr1_cmd += " --use_cuda"
            env0 = {"CUDA_VISIBLE_DEVICES": "0"}
            env1 = {"CUDA_VISIBLE_DEVICES": "1"}
        else:
            env0 = {'CPU_NUM': '1'}
            env1 = {'CPU_NUM': '1'}

        env0.update(envs)
        env1.update(envs)
X
Xin Pan 已提交
1252

W
Wu Yi 已提交
1253 1254
        print("tr0_cmd: {}, env: {}".format(tr0_cmd, env0))
        print("tr1_cmd: {}, env: {}".format(tr1_cmd, env1))
1255 1256 1257 1258 1259

        path0 = os.path.join(self.temp_dir.name, log_name + "_tr0_err.log")
        path1 = os.path.join(self.temp_dir.name, log_name + "_tr1_err.log")
        tr0_pipe = open(path0, "wb")
        tr1_pipe = open(path1, "wb")
G
gongweibao 已提交
1260

1261
        print_to_err(type(self).__name__, "going to start trainer process 0")
1262 1263 1264 1265 1266 1267
        tr0_proc = subprocess.Popen(
            tr0_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=tr0_pipe,
            env=env0,
        )
1268
        print_to_err(type(self).__name__, "going to start trainer process 1")
1269 1270 1271 1272 1273 1274
        tr1_proc = subprocess.Popen(
            tr1_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=tr1_pipe,
            env=env1,
        )
X
Xin Pan 已提交
1275

1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287
        # Wait until trainer process terminate
        while True:
            stat0 = tr0_proc.poll()
            time.sleep(0.1)
            if stat0 is not None:
                break
        while True:
            stat1 = tr1_proc.poll()
            time.sleep(0.1)
            if stat1 is not None:
                break

1288 1289
        tr0_out, tr0_err = tr0_proc.communicate()
        tr1_out, tr1_err = tr1_proc.communicate()
X
Xin Pan 已提交
1290

G
gongweibao 已提交
1291
        # close trainer file
1292 1293 1294 1295
        tr0_pipe.close()
        tr1_pipe.close()
        ps0_pipe.close()
        ps1_pipe.close()
W
Wu Yi 已提交
1296

W
Wu Yi 已提交
1297 1298
        ps0.terminate()
        ps1.terminate()
T
typhoonzero 已提交
1299

W
Wu Yi 已提交
1300 1301
        return pickle.loads(tr0_out), pickle.loads(tr1_out)

1302 1303 1304
    def _get_gloo_trainer_cmd(
        self, model, ep, update_method, trainer_id, trainer_num
    ):
X
xiongkun 已提交
1305 1306 1307 1308 1309 1310 1311 1312
        env = {}
        tr_cmd = "%s -u"

        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            tr_cmd += " -m coverage run --branch -p"

        tr_cmd += " %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --update_method %s --lr %f"

1313 1314 1315 1316 1317 1318 1319 1320 1321
        tr_cmd = tr_cmd % (
            self._python_interp,
            model,
            self._ps_endpoints,
            trainer_id,
            ep,
            update_method,
            self._lr,
        )
X
xiongkun 已提交
1322 1323 1324 1325 1326

        if self._use_reduce:
            tr_cmd += " --use_reduce"
        if self._use_reader_alloc:
            tr_cmd += " --use_reader_alloc"
1327 1328
        # assert self._use_reduce == False, "gloo not support _use_reduce"
        # assert self._use_reader_alloc == False, "gloo not support _use_reduce"
X
xiongkun 已提交
1329 1330
        if self._save_model:
            tr_cmd += " --save_model"
1331 1332
        if self._diff_batch:
            tr_cmd += " --diff_batch"
X
xiongkun 已提交
1333 1334
        self.__use_cuda = False
        self.__use_xpu = False
1335 1336
        assert not self.__use_cuda, "gloo not support use cuda"
        assert not self.__use_xpu, "gloo not support use xpu"
X
xiongkun 已提交
1337
        tr_cmd += " --use_cpu"
1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348
        env.update(
            {
                "PADDLE_TRAINERS_NUM": "{}".format(trainer_num),
                "PADDLE_TRAINER_ID": "{}".format(trainer_id),
                "PADDLE_TRAINER_ENDPOINTS": self._ps_endpoints,
                "PADDLE_CURRENT_ENDPOINT": ep,
                "PADDLE_CURRENT_ENDPOINT": ep,
                "PADDLE_DISTRI_BACKEND": "gloo",
                "GLOG_v": "2",
            }
        )
X
xiongkun 已提交
1349

1350
        assert not self._use_dgc, "gloo not support use dgc"
1351

X
xiongkun 已提交
1352 1353 1354 1355 1356 1357
        if self._accumulate_gradient:
            tr_cmd += " --accumulate_gradient"

        if self._find_unused_parameters:
            tr_cmd += " --find_unused_parameters"

1358
        assert not self._pipeline_mode, "gloo not support use pipeline"
X
xiongkun 已提交
1359 1360 1361 1362 1363 1364 1365

        if self._enable_backward_deps:  # build strategy, save it
            tr_cmd += " --enable_backward_deps"

        if self._fuse_all_reduce is not None:
            tr_cmd += " --fuse_all_reduce {}".format(self._fuse_all_reduce)

1366 1367
        assert not self._use_fleet_api, "gloo not support use fleet api"
        assert not self._use_fleet_api_20, "gloo not support use fleet api"
X
xiongkun 已提交
1368 1369
        return tr_cmd, env

1370 1371 1372
    def _get_nccl2_trainer_cmd(
        self, model, ep, update_method, trainer_id, trainer_num
    ):
1373
        env = {}
1374 1375 1376 1377 1378 1379 1380
        tr_cmd = "%s -u"

        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            tr_cmd += " -m coverage run --branch -p"

        tr_cmd += " %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --update_method %s --lr %f"

1381 1382 1383 1384 1385 1386 1387 1388 1389
        tr_cmd = tr_cmd % (
            self._python_interp,
            model,
            self._ps_endpoints,
            trainer_id,
            ep,
            update_method,
            self._lr,
        )
W
Wu Yi 已提交
1390 1391

        if self._use_reduce:
1392
            tr_cmd += " --use_reduce"
W
Wu Yi 已提交
1393
        if self._use_reader_alloc:
1394
            tr_cmd += " --use_reader_alloc"
1395 1396
        if self._save_model:
            tr_cmd += " --save_model"
W
Wu Yi 已提交
1397
        if self.__use_cuda:
1398
            tr_cmd += " --use_cuda"
1399 1400 1401 1402 1403 1404 1405 1406 1407 1408
            env.update(
                {
                    "FLAGS_selected_gpus": "{}".format(0),
                    "CUDA_VISIBLE_DEVICES": "{}".format(trainer_id),
                    "PADDLE_TRAINERS_NUM": "{}".format(trainer_num),
                    "PADDLE_TRAINER_ID": "{}".format(trainer_id),
                    "PADDLE_TRAINER_ENDPOINTS": self._ps_endpoints,
                    "PADDLE_CURRENT_ENDPOINT": ep,
                }
            )
1409 1410 1411 1412
        # TODO(liuyuhui):XPU_VISIBLE_DEVICES is not working right now,
        # will update it after Badiu Kunlun partners' support.
        elif self.__use_xpu:
            tr_cmd += " --use_xpu"
1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423
            env.update(
                {
                    "FLAGS_selected_xpus": "{}".format(trainer_id),
                    # "XPU_VISIBLE_DEVICES": "{}".format(trainer_id + 1),
                    "PADDLE_TRAINERS_NUM": "{}".format(trainer_num),
                    "PADDLE_TRAINER_ID": "{}".format(trainer_id),
                    "PADDLE_TRAINER_ENDPOINTS": self._ps_endpoints,
                    "PADDLE_CURRENT_ENDPOINT": ep,
                    "GLOG_v": "2",
                }
            )
1424 1425
        elif self.__use_npu:
            tr_cmd += " --use_npu"
1426 1427 1428 1429 1430 1431 1432 1433 1434 1435
            env.update(
                {
                    "FLAGS_selected_npus": "{}".format(trainer_id),
                    "PADDLE_TRAINERS_NUM": "{}".format(trainer_num),
                    "PADDLE_TRAINER_ID": "{}".format(trainer_id),
                    "PADDLE_TRAINER_ENDPOINTS": self._ps_endpoints,
                    "PADDLE_CURRENT_ENDPOINT": ep,
                    "GLOG_v": "2",
                }
            )
1436 1437
        elif self._use_mlu:
            tr_cmd += " --use_mlu"
1438 1439 1440 1441 1442 1443 1444 1445 1446 1447
            env.update(
                {
                    "FLAGS_selected_mlus": "{}".format(trainer_id),
                    "PADDLE_TRAINERS_NUM": "{}".format(trainer_num),
                    "PADDLE_TRAINER_ID": "{}".format(trainer_id),
                    "PADDLE_TRAINER_ENDPOINTS": self._ps_endpoints,
                    "PADDLE_CURRENT_ENDPOINT": ep,
                    "GLOG_v": "4",
                }
            )
W
Wu Yi 已提交
1448
        else:
1449
            env.update({'CPU_NUM': '1'})
W
Wu Yi 已提交
1450

1451
        if self._use_dgc:
1452 1453
            tr_cmd += " --use_dgc"

1454 1455 1456
        if self._accumulate_gradient:
            tr_cmd += " --accumulate_gradient"

1457 1458 1459
        if self._find_unused_parameters:
            tr_cmd += " --find_unused_parameters"

1460 1461
        if self._pipeline_mode:
            tr_cmd += " --use_pipeline"
1462
        if self._mp_mode:
W
WangXi 已提交
1463
            env = {"FLAGS_selected_gpus": "{}".format(trainer_id)}
1464 1465

        if self._nccl_comm_num > 1:
1466
            tr_cmd += " --nccl_comm_num {}".format(self._nccl_comm_num)
1467

1468 1469
        if self._use_hallreduce:
            tr_cmd += " --use_hallreduce --hallreduce_inter_nranks 2"
1470

1471
        if self._enable_backward_deps:
1472
            tr_cmd += " --enable_backward_deps"
1473

1474 1475 1476
        if self._fuse_all_reduce is not None:
            tr_cmd += " --fuse_all_reduce {}".format(self._fuse_all_reduce)

1477
        if self._use_fleet_api:
1478 1479 1480 1481 1482
            tr_cmd += (
                " --use_fleet_api_20"
                if self._use_fleet_api_20
                else " --use_fleet_api"
            )
1483 1484 1485 1486
            if self._use_local_sgd:
                tr_cmd += " --use_local_sgd"
            if self._ut4grad_allreduce:
                tr_cmd += " --ut4grad_allreduce"
1487 1488
            if hasattr(self, '_sync_batch_norm') and self._sync_batch_norm:
                tr_cmd += " --sync_batch_norm"
1489

1490 1491 1492
        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            env['COVERAGE_FILE'] = os.getenv('COVERAGE_FILE', '')

1493
        return tr_cmd, env
W
Wu Yi 已提交
1494

1495 1496 1497 1498 1499 1500 1501 1502 1503 1504
    def _run_cluster_gloo(
        self, model, envs, update_method, check_error_log, log_name
    ):
        assert update_method == "gloo", (
            "_run_cluster_gloo must have update_method: gloo, but get %s"
            % update_method
        )
        assert (
            not self._use_hallreduce
        ), "_run_cluster_gloo must have _use_hallreduce = false"
X
xiongkun 已提交
1505 1506 1507 1508 1509 1510 1511 1512

        worker_endpoints = self._ps_endpoints.split(",")

        trainer_num = len(worker_endpoints)

        procs = []
        pipes = []
        for i in range(0, trainer_num):
1513 1514 1515
            tr_cmd, tr_env = self._get_gloo_trainer_cmd(
                model, worker_endpoints[i], update_method, i, trainer_num
            )
X
xiongkun 已提交
1516 1517 1518
            tr_env.update(envs)
            tr_env["GLOG_vmodule"] = 'gloo_context=4'
            tr_env["GLOG_v"] = '3'
1519 1520 1521 1522 1523
            print(
                "use_hallreduce:{} tr_cmd:{}, env: {}".format(
                    self._use_hallreduce, tr_cmd, tr_env
                )
            )
X
xiongkun 已提交
1524

1525 1526 1527
            path = os.path.join(
                self.temp_dir.name, log_name + "_tr{}_err.log".format(i)
            )
1528
            tr_pipe = open(path, "wb")
X
xiongkun 已提交
1529 1530 1531

            print_to_err(
                type(self).__name__,
1532 1533 1534 1535 1536 1537 1538 1539
                "going to start process {} with nccl2".format(i),
            )
            tr_proc = subprocess.Popen(
                tr_cmd.strip().split(" "),
                stdout=subprocess.PIPE,
                stderr=tr_pipe,
                env=tr_env,
            )
X
xiongkun 已提交
1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551

            procs.append(tr_proc)
            pipes.append(tr_pipe)

        outs = []
        for i in range(0, trainer_num):
            tr_out, tr_err = procs[i].communicate()
            outs.append(tr_out)
            pipes[i].close()
            sys.stderr.write('trainer {} stderr: {}\n'.format(i, tr_err))

        if trainer_num == 1:
1552 1553
            if check_error_log:
                print("outs[0]:", outs[0])
X
xiongkun 已提交
1554 1555 1556 1557 1558 1559 1560 1561
            return pickle.loads(outs[0])

        else:
            if check_error_log:
                print("outs[0]:", outs[0])
                print("outs[1]:", outs[1])
            return pickle.loads(outs[0]), pickle.loads(outs[1])

1562 1563 1564
    def _run_cluster_nccl2(
        self, model, envs, update_method, check_error_log, log_name
    ):
1565 1566
        if self._use_hallreduce:
            self._ps_endpoints = ""
1567 1568 1569

            global DIST_UT_PORT
            if DIST_UT_PORT == 0:
W
WangXi 已提交
1570
                # NOTE(wangxi). hallreduce test must use 4cards after nccl>=2.7
1571 1572
                for i in range(0, 4):
                    self._ps_endpoints += "127.0.0.1:%s," % (
1573 1574
                        self._find_free_port()
                    )
1575 1576 1577 1578
            else:
                for i in range(0, 4):
                    self._ps_endpoints += "127.0.0.1:%s," % (DIST_UT_PORT + i)
                DIST_UT_PORT += 4
1579
            self._ps_endpoints = self._ps_endpoints[:-1]
W
Wu Yi 已提交
1580

1581 1582
        # NOTE: we reuse ps_endpoints as nccl2 worker endpoints
        worker_endpoints = self._ps_endpoints.split(",")
W
Wu Yi 已提交
1583

1584
        trainer_num = len(worker_endpoints)
W
Wu Yi 已提交
1585

1586 1587 1588 1589
        procs = []
        pipes = []
        for i in range(0, trainer_num):
            tr_cmd, tr_env = self._get_nccl2_trainer_cmd(
1590 1591
                model, worker_endpoints[i], update_method, i, trainer_num
            )
1592
            tr_env.update(envs)
1593 1594 1595 1596 1597
            print(
                "use_hallreduce:{} tr_cmd:{}, env: {}".format(
                    self._use_hallreduce, tr_cmd, tr_env
                )
            )
W
Wu Yi 已提交
1598

1599 1600 1601
            path = os.path.join(
                self.temp_dir.name, log_name + "_tr{}_err.log".format(i)
            )
1602
            tr_pipe = open(path, "wb")
W
Wu Yi 已提交
1603

1604
            print_to_err(
1605
                type(self).__name__,
1606 1607 1608 1609 1610 1611 1612 1613
                "going to start process {} with nccl2".format(i),
            )
            tr_proc = subprocess.Popen(
                tr_cmd.strip().split(" "),
                stdout=subprocess.PIPE,
                stderr=tr_pipe,
                env=tr_env,
            )
1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624

            procs.append(tr_proc)
            pipes.append(tr_pipe)

        outs = []
        for i in range(0, trainer_num):
            tr_out, tr_err = procs[i].communicate()
            outs.append(tr_out)
            pipes[i].close()
            sys.stderr.write('trainer {} stderr: {}\n'.format(i, tr_err))

1625 1626 1627
        if check_error_log:
            print("outs[0]:", outs[0])
            print("outs[1]:", outs[1])
1628

1629
        return pickle.loads(outs[0]), pickle.loads(outs[1])
1630

1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641
    def _run_pipeline(self, model, envs, check_error_log, log_name):
        # NOTE: we reuse ps_endpoints as nccl2 worker endpoints
        worker_endpoints = self._ps_endpoints.split(",")
        update_method = "nccl2"

        trainer_num = len(worker_endpoints)

        procs = []
        pipes = []
        for i in range(0, trainer_num):
            tr_cmd, tr_env = self._get_nccl2_trainer_cmd(
1642 1643
                model, worker_endpoints[i], update_method, i, trainer_num
            )
1644 1645 1646 1647 1648 1649 1650
            tr_env.update(envs)
            tr_env['CUDA_VISIBLE_DEVICES'] = "0,1"
            tr_env['NCCL_SHM_DISABLE'] = '1'
            tr_env['FLAGS_selected_gpus'] = str(i)
            tr_env['FLAGS_cudnn_deterministic'] = '0'
            print("tr_cmd:{}, env: {}".format(tr_cmd, tr_env))

1651 1652
            path = os.path.join(self.temp_dir.name + "tr{}_err.log".format(i))
            tr_pipe = open(path, "wb")
1653 1654 1655

            print_to_err(
                type(self).__name__,
1656 1657 1658 1659 1660 1661 1662 1663
                "going to start process {} with nccl2".format(i),
            )
            tr_proc = subprocess.Popen(
                tr_cmd.strip().split(" "),
                stdout=subprocess.PIPE,
                stderr=tr_pipe,
                env=tr_env,
            )
1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679

            procs.append(tr_proc)
            pipes.append(tr_pipe)

        outs = []
        for i in range(0, trainer_num):
            tr_out, tr_err = procs[i].communicate()
            outs.append(tr_out)
            pipes[i].close()
            sys.stderr.write('trainer {} stderr: {}\n'.format(i, tr_err))

        if check_error_log:
            print("outs[0]:", outs[0])
            print("outs[1]:", outs[1])
        return pickle.loads(outs[0]), pickle.loads(outs[1])

1680
    def _get_required_envs(self, check_error_log=False, need_envs={}):
1681 1682 1683 1684 1685 1686
        # TODO(typhoonzero): should auto adapt GPU count on the machine.
        required_envs = {
            "PATH": os.getenv("PATH", ""),
            "PYTHONPATH": os.getenv("PYTHONPATH", ""),
            "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""),
            "FLAGS_fraction_of_gpu_memory_to_use": "0.15",
G
guru4elephant 已提交
1687
            "FLAGS_rpc_deadline": "30000",  # 5sec to fail fast
1688
            "FLAGS_rpc_retry_bind_port": "50",
1689
            "FLAGS_cudnn_deterministic": "1",
1690
            "FLAGS_rpc_disable_reuse_port": "1",
W
Wu Yi 已提交
1691
            "http_proxy": "",
1692
            "NCCL_P2P_DISABLE": "1",
1693
            "NCCL_SHM_DISABLE": "1",
1694 1695 1696
        }

        if check_error_log:
1697 1698 1699 1700
            required_envs["GLOG_vmodule"] = (
                "fused_all_reduce_op_handle=10,all_reduce_op_handle=10,alloc_continuous_space_op=10,fuse_all_reduce_op_pass=10,"
                "alloc_continuous_space_for_grad_pass=10,fast_threaded_ssa_graph_executor=10,executor=10,operator=10,"
                "sparse_all_reduce_op_handle=10,gen_nccl_id_op=10,gen_nccl_id_op_help=10,nccl_helper=10,grpc_client=10,"
1701
                "grpc_server=10,request_handler_impl=10,section_worker=10"
1702
            )
1703 1704
            required_envs["GLOG_logtostderr"] = "1"

1705 1706
        if os.getenv('NVIDIA_TF32_OVERRIDE', '') is not None:
            required_envs['NVIDIA_TF32_OVERRIDE'] = os.getenv(
1707 1708
                'NVIDIA_TF32_OVERRIDE', ''
            )
1709

1710 1711 1712
        required_envs.update(need_envs)
        return required_envs

1713 1714 1715 1716 1717 1718 1719 1720
    def check_with_place(
        self,
        model_file,
        delta=1e-3,
        check_error_log=False,
        need_envs={},
        log_name="",
    ):
1721
        if self._dygraph and (self._gloo_mode or self._nccl2_mode):
1722 1723 1724 1725 1726 1727 1728
            self.check_with_place_func(
                model_file=model_file,
                delta=delta,
                check_error_log=check_error_log,
                need_envs=need_envs,
                log_name=log_name,
            )
1729
        else:
1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745
            self.check_with_place_func(
                model_file=model_file,
                delta=delta,
                check_error_log=check_error_log,
                need_envs=need_envs,
                log_name=log_name,
            )

    def check_with_place_func(
        self,
        model_file,
        delta=1e-3,
        check_error_log=False,
        need_envs={},
        log_name="",
    ):
1746 1747
        required_envs = self._get_required_envs(check_error_log, need_envs)

X
xiongkun 已提交
1748
        if self._gloo_mode:
1749 1750 1751
            local_losses = self._run_local_gloo(
                model_file, required_envs, check_error_log, log_name=log_name
            )
X
xiongkun 已提交
1752
        else:
1753 1754 1755
            local_losses = self._run_local(
                model_file, required_envs, check_error_log, log_name=log_name
            )
1756

W
Wu Yi 已提交
1757
        if self._nccl2_mode:
W
Wu Yi 已提交
1758 1759
            if self._nccl2_reduce_layer:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
1760 1761
                    model_file,
                    required_envs,
1762 1763
                    update_method="nccl2_reduce_layer",
                    check_error_log=check_error_log,
1764 1765
                    log_name=log_name,
                )
W
Wu Yi 已提交
1766 1767
            else:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
1768 1769
                    model_file,
                    required_envs,
1770 1771
                    update_method='nccl2',
                    check_error_log=check_error_log,
1772 1773
                    log_name=log_name,
                )
1774 1775 1776 1777 1778 1779
        elif self._bkcl_mode:
            tr0_losses, tr1_losses = self._run_cluster_nccl2(
                model_file,
                required_envs,
                update_method='bkcl',
                check_error_log=check_error_log,
1780 1781
                log_name=log_name,
            )
X
xiongkun 已提交
1782 1783 1784 1785 1786 1787 1788
        elif self._gloo_mode:
            # gloo mode, cpu only parallel train @xiongkun03
            tr0_losses, tr1_losses = self._run_cluster_gloo(
                model_file,
                required_envs,
                update_method='gloo',
                check_error_log=check_error_log,
1789 1790
                log_name=log_name,
            )
1791 1792 1793 1794 1795 1796
        elif self._hccl_mode:
            tr0_losses, tr1_losses = self._run_cluster_nccl2(
                model_file,
                required_envs,
                update_method='hccl',
                check_error_log=check_error_log,
1797 1798
                log_name=log_name,
            )
1799 1800 1801 1802 1803 1804
        elif self._cncl_mode:
            tr0_losses, tr1_losses = self._run_cluster_nccl2(
                model_file,
                required_envs,
                update_method='cncl',
                check_error_log=check_error_log,
1805 1806
                log_name=log_name,
            )
1807
        elif self._pipeline_mode:
1808 1809 1810
            tr0_losses, tr1_losses = self._run_pipeline(
                model_file, required_envs, check_error_log, log_name=log_name
            )
W
Wu Yi 已提交
1811
        else:
1812 1813 1814
            tr0_losses, tr1_losses = self._run_cluster(
                model_file, required_envs, check_error_log, log_name=log_name
            )
1815 1816

        for step_id in range(RUN_STEP):
W
Wu Yi 已提交
1817 1818 1819
            local_loss = local_losses[step_id]
            tr0_loss = tr0_losses[step_id]
            tr1_loss = tr1_losses[step_id]
1820 1821 1822 1823
            if self._pipeline_mode:
                dist_loss = np.array([tr1_loss])
            else:
                dist_loss = (np.array([tr0_loss]) + np.array([tr1_loss])) / 2
W
Wu Yi 已提交
1824 1825
            print("=======", local_loss, ":", dist_loss[0], "=======")
            self.assertAlmostEqual(local_loss, dist_loss[0], delta=delta)
1826

1827 1828 1829 1830 1831 1832 1833 1834
    def check_with_place_multi_cards(
        self,
        model_file,
        delta=1e-3,
        check_error_log=False,
        need_envs={},
        log_name="",
    ):
1835

1836 1837 1838 1839 1840 1841
        # need open p2p or shm otherwise multi cards mode will hang
        need_envs.update({"NCCL_P2P_DISABLE": "0", "NCCL_SHM_DISABLE": "0"})

        required_envs = self._get_required_envs(check_error_log, need_envs)

        if self._use_dgc:
1842 1843 1844 1845 1846 1847 1848
            multi_cards_losses = self._run_local(
                model_file,
                required_envs,
                check_error_log,
                log_name=log_name + "_dgc_2cards",
                devices="0,1",
            )
1849 1850

            self._use_dgc = False
1851 1852 1853 1854 1855 1856 1857
            base_losses = self._run_local(
                model_file,
                required_envs,
                check_error_log,
                log_name=log_name + "_base_2cards",
                devices="0,1",
            )
1858 1859 1860 1861 1862 1863 1864 1865

            self._use_dgc = True

            for step_id in range(RUN_STEP):
                base_loss = base_losses[step_id]
                multi_cards_loss = multi_cards_losses[step_id]
                print("=======", base_loss, ":", multi_cards_loss, "=======")
                self.assertAlmostEqual(base_loss, multi_cards_loss, delta=delta)