test_dist_base.py 40.5 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

from __future__ import print_function
X
Xin Pan 已提交
16 17 18 19 20 21 22
import time

import unittest
import os
import sys
import signal
import subprocess
23
import six
W
Wu Yi 已提交
24
import argparse
W
Wu Yi 已提交
25
import pickle
26
import random
W
Wu Yi 已提交
27
import numpy as np
28
import time
29 30

import paddle
31
import paddle.fluid as fluid
32
from paddle.fluid import compiler
33 34 35
import paddle.fluid.dygraph as dygraph
from paddle.fluid.dygraph.base import to_variable
from paddle.fluid.dygraph.parallel import DataParallel
36

37 38 39
from paddle.fluid.incubate.fleet.collective import fleet, DistributedStrategy
import paddle.fluid.incubate.fleet.base.role_maker as role_maker

Y
Yan Xu 已提交
40
RUN_STEP = 5
41
DEFAULT_BATCH_SIZE = 2
42
DIST_UT_PORT = 0
43

T
typhoonzero 已提交
44

45 46 47 48 49 50 51 52
def print_to_out(out_losses):
    if six.PY2:
        print(pickle.dumps(out_losses))
    else:
        sys.stdout.buffer.write(pickle.dumps(out_losses))


def print_to_err(class_name, log_str):
53 54
    localtime = time.asctime(time.localtime(time.time()))
    print_str = localtime + "\t" + class_name + "\t" + log_str
G
guru4elephant 已提交
55
    if six.PY2:
56
        sys.stderr.write(pickle.dumps(print_str))
G
guru4elephant 已提交
57
    else:
58
        sys.stderr.buffer.write(pickle.dumps(print_str))
G
guru4elephant 已提交
59 60


61 62 63 64
def eprint(*args, **kwargs):
    print(*args, file=sys.stderr, **kwargs)


T
typhoonzero 已提交
65
class TestDistRunnerBase(object):
W
Wu Yi 已提交
66 67 68
    def get_model(self,
                  batch_size=DEFAULT_BATCH_SIZE,
                  lr=0.1,
69 70
                  single_device=False,
                  use_dgc=False):
T
typhoonzero 已提交
71 72 73
        raise NotImplementedError(
            "get_model should be implemented by child classes.")

74
    @staticmethod
W
Wu Yi 已提交
75 76 77 78 79
    def get_transpiler(trainer_id,
                       main_program,
                       pserver_endpoints,
                       trainers,
                       sync_mode,
80
                       dc_asgd=False,
81
                       current_endpoint=None,
T
tangwei12 已提交
82 83
                       nccl_comm_num=1,
                       hogwild_mode=False):
T
typhoonzero 已提交
84
        # NOTE: import fluid until runtime, or else forking processes will cause error.
85
        config = fluid.DistributeTranspilerConfig()
W
Wu Yi 已提交
86
        config.enable_dc_asgd = dc_asgd
87
        config.sync_mode = sync_mode
T
tangwei12 已提交
88 89
        config.runtime_split_send_recv = hogwild_mode

90 91
        if nccl_comm_num > 1:
            config.nccl_comm_num = nccl_comm_num
92
        # config.runtime_split_send_recv = True
93
        t = fluid.DistributeTranspiler(config=config)
T
typhoonzero 已提交
94 95 96 97
        t.transpile(
            trainer_id=trainer_id,
            program=main_program,
            pservers=pserver_endpoints,
W
Wu Yi 已提交
98
            trainers=trainers,
T
tangwei12 已提交
99
            sync_mode=sync_mode,
100
            current_endpoint=current_endpoint)
T
typhoonzero 已提交
101 102
        return t

W
Wu Yi 已提交
103
    def run_pserver(self, args):
W
Wu Yi 已提交
104
        self.lr = args.lr
105
        self.get_model(batch_size=args.batch_size)
106
        # NOTE: pserver should not call memory optimize
T
tangwei12 已提交
107 108 109 110 111 112 113 114 115

        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
116 117 118
        pserver_prog = t.get_pserver_program(args.current_endpoint)
        startup_prog = t.get_startup_program(args.current_endpoint,
                                             pserver_prog)
Y
Yancey1989 已提交
119

T
typhoonzero 已提交
120 121 122
        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        exe.run(startup_prog)
123
        print_to_err(type(self).__name__, "run pserver startup program done.")
T
typhoonzero 已提交
124
        exe.run(pserver_prog)
125
        print_to_err(type(self).__name__, "run pserver main program done.")
T
typhoonzero 已提交
126

127 128 129 130 131 132 133 134 135 136
    def run_gpu_fleet_api_trainer(self, args):
        assert args.update_method == "nccl2"

        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
137
        dist_strategy.fuse_memory_size = 1  # MB
138
        dist_strategy.fuse_laryer_size = 1
139 140 141 142
        if args.use_local_sgd:
            dist_strategy.use_local_sgd = True
        if args.ut4grad_allreduce:
            dist_strategy._ut4grad_allreduce = True
143 144
        if args.sync_batch_norm:
            dist_strategy.sync_batch_norm = True
145 146 147

        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
        fleet.init(role)
148
        print_to_err("gpu_fleet", "fleet.node_num:")
T
tangwei12 已提交
149 150
        # "fleet.node_id:", fleet.node_id(),
        # "fleet.trainer_num:", fleet.worker_num())
151 152

        test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
T
tangwei12 已提交
153
            self.get_model(batch_size=args.batch_size, dist_strategy=dist_strategy)
154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169

        trainer_prog = fleet._origin_program
        dist_prog = fleet.main_program

        device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
        place = fluid.CUDAPlace(device_id)

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

        feed_var_list = [
            var for var in trainer_prog.global_block().vars.values()
            if var.is_data
        ]

170 171 172 173 174 175 176
        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]

177 178 179 180 181 182 183 184 185 186 187 188 189 190
        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

191
        print_to_err(type(self).__name__, "begin to train on trainer")
192 193 194 195 196 197
        out_losses = []
        for i in six.moves.xrange(RUN_STEP):
            loss, = exe.run(dist_prog,
                            fetch_list=[avg_cost.name],
                            feed=feeder.feed(get_data()))
            out_losses.append(loss[0])
198 199
            print_to_err(type(self).__name__, "run step %d finished" % i)
        print_to_err(type(self).__name__, "trainer run finished")
200 201 202 203 204 205

        if six.PY2:
            print(pickle.dumps(out_losses))
        else:
            sys.stdout.buffer.write(pickle.dumps(out_losses))

206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
        if args.save_model:
            model_save_dir = "/tmp"
            if fleet.worker_index() == 0:
                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")
            else:
                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")
            fluid.io.save_persistables(exe, model_save_dir_fluid,
                                       fleet._origin_program)
            fleet.save_persistables(executor=exe, dirname=model_save_dir_fleet)
            feeded_var_names = [var.name for var in feed_var_list]
            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])

236
    def run_trainer(self, args):
W
Wu Yi 已提交
237
        self.lr = args.lr
W
Wu Yi 已提交
238 239 240
        if args.nccl2_reduce_layer_local_run:
            test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
                self.get_model(batch_size=args.batch_size, single_device=True)
241 242 243
        elif args.use_dgc:
            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
244 245 246
        else:
            test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
                self.get_model(batch_size=args.batch_size)
247

W
Wu Yi 已提交
248
        if args.update_method == "pserver":
249
            print_to_err(
250 251
                type(self).__name__,
                "begin to run transpile on trainer with pserver mode")
T
tangwei12 已提交
252 253 254 255 256 257 258 259 260
            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
typhoonzero 已提交
261
            trainer_prog = t.get_trainer_program()
262
            print_to_err(
263 264
                type(self).__name__,
                "get trainer program done with pserver mode.")
W
Wu Yi 已提交
265
        elif args.update_method == "nccl2" or args.update_method == "nccl2_reduce_layer":
W
Wu Yi 已提交
266 267 268
            # transpile for nccl2
            config = fluid.DistributeTranspilerConfig()
            config.mode = "nccl2"
269
            config.nccl_comm_num = args.nccl_comm_num
270 271 272
            if args.use_hallreduce:
                config.use_hierarchical_allreduce = True
                config.hierarchical_allreduce_inter_nranks = args.hallreduce_inter_nranks
273
            print_to_err(
274 275
                type(self).__name__,
                "begin to run transpile on trainer with nccl2 mode")
W
Wu Yi 已提交
276 277 278 279 280 281 282
            nccl2_t = fluid.DistributeTranspiler(config=config)
            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)
283
            print_to_err(
284 285
                type(self).__name__,
                "get trainer program done. with nccl2 mode")
W
Wu Yi 已提交
286
            trainer_prog = fluid.default_main_program()
T
typhoonzero 已提交
287
        else:
288
            print_to_err(
289 290
                type(self).__name__,
                "do nothing about main program, just use it")
T
typhoonzero 已提交
291
            trainer_prog = fluid.default_main_program()
292
            print_to_err(type(self).__name__, "use main program done.")
T
typhoonzero 已提交
293

294 295 296
        # FIXME(gongwb):wait pserver initialization.
        time.sleep(1)

297
        if args.use_cuda:
298 299
            device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
            place = fluid.CUDAPlace(device_id)
300 301 302
        else:
            place = fluid.CPUPlace()

303 304
        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
305
        print_to_err(type(self).__name__, "run worker startup program done.")
T
typhoonzero 已提交
306

W
Wu Yi 已提交
307 308
        exec_strategy = fluid.ExecutionStrategy()
        exec_strategy.num_threads = 1
309

W
Wu Yi 已提交
310
        build_stra = fluid.BuildStrategy()
311 312 313
        # FIXME force disable enable_inplace and memory_optimize
        build_stra.enable_inplace = False
        build_stra.memory_optimize = False
W
Wu Yi 已提交
314

T
tangwei12 已提交
315 316 317
        if args.hogwild:
            build_stra.async_mode = True

318 319 320
        if args.enable_backward_deps:
            build_stra.enable_backward_optimizer_op_deps = True

W
Wu Yi 已提交
321 322 323 324 325
        if args.use_reduce:
            build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
        else:
            build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce

W
Wu Yi 已提交
326
        pass_builder = None
X
Xin Pan 已提交
327
        if args.batch_merge_repeat > 1:
X
fix  
Xin Pan 已提交
328
            pass_builder = build_stra._finalize_strategy_and_create_passes()
329
            mypass = pass_builder.insert_pass(0, "multi_batch_merge_pass")
330
            mypass.set("num_repeats", args.batch_merge_repeat)
X
Xin Pan 已提交
331

W
Wu Yi 已提交
332
        if args.update_method == "nccl2" or args.update_method == "nccl2_reduce_layer":
333 334
            build_stra.num_trainers = len(args.endpoints.split(","))
            build_stra.trainer_id = args.trainer_id
W
Wu Yi 已提交
335
        else:
W
Wu Yi 已提交
336
            # case args.update_method == "nccl2_reduce_layer":
337 338
            build_stra.num_trainers = 1
            build_stra.trainer_id = 0
W
Wu Yi 已提交
339

340
        print_to_err(type(self).__name__, "begin to compile with data parallel")
X
Xin Pan 已提交
341
        binary = compiler.CompiledProgram(trainer_prog).with_data_parallel(
W
Wu Yi 已提交
342
            loss_name=avg_cost.name,
W
Wu Yi 已提交
343
            build_strategy=build_stra,
W
Wu Yi 已提交
344
            exec_strategy=exec_strategy)
345
        print_to_err(type(self).__name__, "program compiled with data parallel")
T
typhoonzero 已提交
346 347 348 349 350 351 352

        feed_var_list = [
            var for var in trainer_prog.global_block().vars.values()
            if var.is_data
        ]

        feeder = fluid.DataFeeder(feed_var_list, place)
353
        reader_generator = train_reader()
T
typhoonzero 已提交
354

355 356
        def get_data():
            origin_batch = next(reader_generator)
W
Wu Yi 已提交
357
            if args.update_method != "local" and args.use_reader_alloc:
358 359 360 361 362 363 364
                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
365

366
        print_to_err(type(self).__name__, "begin to train on trainer")
W
Wu Yi 已提交
367
        out_losses = []
368
        for i in six.moves.xrange(RUN_STEP):
369 370
            loss, = exe.run(binary,
                            fetch_list=[avg_cost.name],
371
                            feed=feeder.feed(get_data()))
W
Wu Yi 已提交
372
            out_losses.append(loss[0])
373 374
            print_to_err(type(self).__name__, "run step %d finished" % i)
        print_to_err(type(self).__name__, "trainer run finished")
375

376
        print_to_out(out_losses)
T
typhoonzero 已提交
377 378


379 380 381 382 383 384 385 386 387
class TestParallelDyGraphRunnerBase(object):
    def get_model(self):
        raise NotImplementedError(
            "get_model should be implemented by child classes.")

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

388 389 390 391 392 393 394 395 396 397
    def _get_data(self, batch, args):
        if args.update_method != "local":
            new_batch = []
            for offset, item in enumerate(batch):
                if offset % 2 == args.trainer_id:
                    new_batch.append(item)
            return new_batch
        else:
            return batch

398
    def run_trainer(self, args):
Y
Yan Xu 已提交
399

400 401 402 403 404 405 406
        seed = 90
        device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
        place = fluid.CUDAPlace(device_id)

        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
407 408
            np.random.seed(seed)
            import random
409
            random.seed = seed
410 411
            model, train_reader, opt = self.get_model()
            nranks = len(args.endpoints.split(",")) if args.endpoints else 1
Y
Yan Xu 已提交
412

413 414 415 416 417 418
            if args.update_method == "nccl2":
                strategy = dygraph.parallel.ParallelStrategy()
                strategy.nranks = nranks
                strategy.local_rank = args.trainer_id
                strategy.trainer_endpoints = args.endpoints.split(",")
                strategy.current_endpoint = args.current_endpoint
419
                print_to_err(
420 421
                    type(self).__name__,
                    "begin to prepare context in dygraph with nccl2")
422
                dygraph.parallel.prepare_context(strategy)
Y
Yan Xu 已提交
423
                model = dygraph.parallel.DataParallel(model, strategy)
424
                print_to_err(type(self).__name__, "model built in dygraph")
425
            out_losses = []
426
            print_to_err(type(self).__name__, "begin to run dygraph training")
427
            for step_id, data in enumerate(train_reader()):
428
                data = self._get_data(data, args)
429 430 431
                if step_id == RUN_STEP:
                    break
                loss = self.run_one_loop(model, opt, data)
G
guru4elephant 已提交
432
                if step_id % 10 == 0:
433
                    print_to_err(
434
                        type(self).__name__,
435
                        "loss at step %d: %f" % (step_id, loss.numpy()))
Y
Yan Xu 已提交
436
                out_losses.append(loss.numpy())
437

Y
Yan Xu 已提交
438 439 440
                # FIXME(Yancey1989): scale the loss inplace
                if args.update_method == "nccl2":
                    loss = model.scale_loss(loss)
441 442

                loss.backward()
Y
Yan Xu 已提交
443 444
                if args.update_method == "nccl2":
                    model.apply_collective_grads()
445 446 447

                opt.minimize(loss)
                model.clear_gradients()
448
        print_to_out(out_losses)
449

450 451 452 453 454 455 456 457 458
    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)
459
        random.seed = seed
460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490
        # get trainer id
        args.trainer_id = paddle.distributed.get_rank()

        # 3. init parallel env
        if args.update_method == "nccl2":
            paddle.distributed.init_parallel_env()

        # 4. train model
        model, train_reader, opt = self.get_model()
        if args.update_method == "nccl2":
            model = paddle.DataParallel(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())

            if args.update_method == "nccl2":
                loss = model.scale_loss(loss)

            loss.backward()
            if args.update_method == "nccl2":
                model.apply_collective_grads()

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

491 492 493 494 495 496 497 498 499 500 501
    def run_gpu_fleet_api_trainer(self, args):
        import paddle.distributed.fleet as fleet
        import paddle.distributed.fleet.base.role_maker as role_maker
        # 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)
502
        random.seed = seed
503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534
        # get trainer id
        args.trainer_id = paddle.distributed.get_rank()

        # 3. init parallel env
        if args.update_method == "nccl2":
            fleet.init(is_collective=True)

        # 4. train model
        model, train_reader, opt = self.get_model()
        if args.update_method == "nccl2":
            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())

            if args.update_method == "nccl2":
                loss = model.scale_loss(loss)

            loss.backward()
            if args.update_method == "nccl2":
                model.apply_collective_grads()

            opt.step()
            opt.clear_grad()
        print_to_out(out_losses)

535

T
typhoonzero 已提交
536
def runtime_main(test_class):
W
Wu Yi 已提交
537 538 539 540
    parser = argparse.ArgumentParser(description='Run dist test.')
    parser.add_argument(
        '--role', type=str, required=True, choices=['pserver', 'trainer'])
    parser.add_argument('--endpoints', type=str, required=False, default="")
W
Wu Yi 已提交
541 542 543 544
    parser.add_argument(
        '--update_method',
        type=str,
        default="local",
W
Wu Yi 已提交
545
        choices=["pserver", "nccl2", "local", "nccl2_reduce_layer"])
W
Wu Yi 已提交
546 547
    parser.add_argument('--trainer_id', type=int, required=False, default=0)
    parser.add_argument('--trainers', type=int, required=False, default=1)
548
    parser.add_argument('--nccl_comm_num', type=int, required=False, default=1)
549 550
    parser.add_argument('--enable_backward_deps', action='store_true')
    parser.add_argument('--use_hallreduce', action='store_true')
551
    parser.add_argument('--gpu_fleet_api', action='store_true')
552 553
    parser.add_argument('--use_local_sgd', action='store_true')
    parser.add_argument('--ut4grad_allreduce', action='store_true')
554
    parser.add_argument(
555
        '--hallreduce_inter_nranks', type=int, required=False, default=2)
W
Wu Yi 已提交
556 557 558
    parser.add_argument(
        '--current_endpoint', type=str, required=False, default="")
    parser.add_argument('--sync_mode', action='store_true')
559
    parser.add_argument('--use_cuda', action='store_true')
560
    parser.add_argument('--use_dgc', action='store_true')
W
Wu Yi 已提交
561
    parser.add_argument('--use_reduce', action='store_true')
W
Wu Yi 已提交
562
    parser.add_argument('--dc_asgd', action='store_true')
T
tangwei12 已提交
563
    parser.add_argument('--hogwild', action='store_true')
564
    parser.add_argument('--save_model', action='store_true')
565
    parser.add_argument(
W
Wu Yi 已提交
566
        '--use_reader_alloc', action='store_true', required=False)
567
    parser.add_argument('--batch_size', required=False, type=int, default=2)
W
Wu Yi 已提交
568
    parser.add_argument('--lr', required=False, type=float, default=0.001)
569 570
    parser.add_argument(
        '--batch_merge_repeat', required=False, type=int, default=1)
W
Wu Yi 已提交
571 572 573 574 575
    parser.add_argument(
        '--nccl2_reduce_layer_local_run',
        required=False,
        type=bool,
        default=False)
576
    parser.add_argument('--sync_batch_norm', action='store_true')
W
Wu Yi 已提交
577 578

    args = parser.parse_args()
T
typhoonzero 已提交
579 580

    model = test_class()
W
Wu Yi 已提交
581
    if args.role == "pserver" and args.update_method == "pserver":
W
Wu Yi 已提交
582
        model.run_pserver(args)
583 584
    elif args.gpu_fleet_api:
        model.run_gpu_fleet_api_trainer(args)
T
typhoonzero 已提交
585
    else:
586
        model.run_trainer(args)
X
Xin Pan 已提交
587

M
minqiyang 已提交
588

M
minqiyang 已提交
589
import paddle.compat as cpt
Y
Yancey1989 已提交
590 591
import socket
from contextlib import closing
M
minqiyang 已提交
592

X
Xin Pan 已提交
593 594

class TestDistBase(unittest.TestCase):
W
Wu Yi 已提交
595 596 597
    def _setup_config(self):
        raise NotImplementedError("tests should have _setup_config implemented")

598 599 600
    def _after_setup_config(self):
        if self._enforce_place == "CPU":
            self.__use_cuda = False
601
            self._use_dgc = False
602 603 604 605 606 607 608
        elif self._enforce_place == "GPU":
            self.__use_cuda = True
        else:
            if fluid.core.is_compiled_with_cuda():
                self.__use_cuda = True
            else:
                self.__use_cuda = False
609 610 611 612
                self._use_dgc = False

        if self._use_reduce:
            assert not self._use_dgc
613

X
Xin Pan 已提交
614 615 616
    def setUp(self):
        self._trainers = 2
        self._pservers = 2
Y
Yancey1989 已提交
617
        self._port_set = set()
M
minqiyang 已提交
618
        self._python_interp = sys.executable
W
Wu Yi 已提交
619
        self._sync_mode = True
T
tangwei12 已提交
620
        self._hogwild_mode = False
621
        self._enforce_place = None
W
Wu Yi 已提交
622
        self._use_reduce = False
W
Wu Yi 已提交
623
        self._dc_asgd = False  # must use with async mode
624
        self._use_reader_alloc = True
W
Wu Yi 已提交
625
        self._nccl2_mode = False
626
        self._mp_mode = False
W
Wu Yi 已提交
627 628 629 630 631
        # 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
632
        self._lr = 0.001
633
        self._use_dgc = False
634
        self._dygraph = False
635
        self._nccl_comm_num = 1
636
        self._enable_backward_deps = False
637
        self._gpu_fleet_api = False
638 639
        self._use_local_sgd = False
        self._ut4grad_allreduce = False
640
        self._use_hallreduce = False
641
        self._save_model = False
W
Wu Yi 已提交
642
        self._setup_config()
643 644 645 646 647 648 649 650 651 652 653 654 655 656

        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" % (
                self._find_free_port(), self._find_free_port())
        else:
            print("set begin_port:", DIST_UT_PORT)
            self._ps_endpoints = "127.0.0.1:%s,127.0.0.1:%s" % (
                DIST_UT_PORT, DIST_UT_PORT + 1)
            DIST_UT_PORT += 2

657
        self._after_setup_config()
X
Xin Pan 已提交
658

Y
Yancey1989 已提交
659
    def _find_free_port(self):
Y
Yancey1989 已提交
660 661 662 663
        def __free_port():
            with closing(socket.socket(socket.AF_INET,
                                       socket.SOCK_STREAM)) as s:
                s.bind(('', 0))
664
                print_to_err(
665
                    type(self).__name__, "socket name: %s" % s.getsockname()[1])
Y
Yancey1989 已提交
666 667 668 669 670 671 672
                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
673

674 675 676 677 678
    def start_pserver(self,
                      model_file,
                      check_error_log,
                      required_envs,
                      log_name=""):
X
Xin Pan 已提交
679
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
680 681 682 683 684 685 686 687
        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"

W
Wu Yi 已提交
688
        ps0_cmd = ps_cmd % \
689 690
                  (self._python_interp, model_file, self._ps_endpoints, ps0_ep,
                   self._trainers)
W
Wu Yi 已提交
691
        ps1_cmd = ps_cmd % \
692 693
                  (self._python_interp, model_file, self._ps_endpoints, ps1_ep,
                   self._trainers)
W
Wu Yi 已提交
694 695 696 697

        if self._sync_mode:
            ps0_cmd += " --sync_mode"
            ps1_cmd += " --sync_mode"
X
Xin Pan 已提交
698

699 700
        print(ps0_cmd)
        print(ps1_cmd)
701 702
        ps0_pipe = open(log_name + "_ps0_err.log", "wb")
        ps1_pipe = open(log_name + "_ps1_err.log", "wb")
G
gongweibao 已提交
703

704
        print_to_err(type(self).__name__, "going to start pserver process 0")
X
Xin Pan 已提交
705
        ps0_proc = subprocess.Popen(
706 707 708 709
            ps0_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps0_pipe,
            env=required_envs)
710
        print_to_err(type(self).__name__, "going to start pserver process 1")
X
Xin Pan 已提交
711
        ps1_proc = subprocess.Popen(
712 713 714 715
            ps1_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps1_pipe,
            env=required_envs)
G
gongweibao 已提交
716

717
        return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe
X
Xin Pan 已提交
718

719 720 721 722 723
    def _run_local(self,
                   model,
                   envs,
                   check_error_log=False,
                   batch_size=DEFAULT_BATCH_SIZE,
724
                   batch_merge_repeat=1,
725 726
                   log_name="",
                   gpus="0"):
G
gongweibao 已提交
727

728 729 730 731 732 733
        cmd = self._python_interp

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

734 735
        cmd += " %s --role trainer --update_method local --lr %f" % (model,
                                                                     self._lr)
736

737 738 739 740
        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
741 742
        if self._nccl2_reduce_layer:
            cmd += " --nccl2_reduce_layer_local_run 1"
743

744
        if self.__use_cuda:
745
            cmd += " --use_cuda"
W
Wu Yi 已提交
746
            env_local = {
747
                "CUDA_VISIBLE_DEVICES": gpus,
W
Wu Yi 已提交
748 749 750
                "PADDLE_TRAINERS_NUM": "1",
                "PADDLE_TRAINER_ID": "0"
            }
751 752 753
        else:
            env_local = {'CPU_NUM': '1'}

754 755 756 757
        # not use dgc in single card
        if len(gpus) > 1 and self._use_dgc:
            cmd += " --use_dgc"

W
Wu Yi 已提交
758 759
        env_local.update(envs)
        print("local_cmd: {}, env: {}".format(cmd, env_local))
G
gongweibao 已提交
760

761
        if check_error_log:
762
            err_log = open(log_name + "_local.log", "wb")
G
gongweibao 已提交
763
            local_proc = subprocess.Popen(
764
                cmd.split(" "),
G
gongweibao 已提交
765
                stdout=subprocess.PIPE,
766
                stderr=err_log,
W
Wu Yi 已提交
767
                env=env_local)
G
gongweibao 已提交
768 769
        else:
            local_proc = subprocess.Popen(
770
                cmd.split(" "),
G
gongweibao 已提交
771
                stdout=subprocess.PIPE,
772
                stderr=subprocess.PIPE,
W
Wu Yi 已提交
773
                env=env_local)
G
gongweibao 已提交
774

775 776 777 778 779 780
        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
781
        sys.stderr.write('local_stdout: %s\n' % pickle.loads(local_out))
X
Xin Pan 已提交
782

W
Wu Yi 已提交
783
        return pickle.loads(local_out)
784

785
    def _run_cluster(self, model, envs, check_error_log, log_name):
X
Xin Pan 已提交
786
        # Run dist train to compare with local results
787 788
        ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(
            model, check_error_log, envs, log_name=log_name)
W
Wu Yi 已提交
789

X
Xin Pan 已提交
790
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
791

792 793 794 795 796 797 798 799
        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"

W
Wu Yi 已提交
800
        tr0_cmd = tr_cmd % \
801
                  (self._python_interp, model, self._ps_endpoints,
W
Wu Yi 已提交
802
                   0, ps0_ep, self._trainers, self._lr)
W
Wu Yi 已提交
803
        tr1_cmd = tr_cmd % \
804
                  (self._python_interp, model, self._ps_endpoints,
W
Wu Yi 已提交
805
                   1, ps1_ep, self._trainers, self._lr)
W
Wu Yi 已提交
806 807 808 809

        if self._sync_mode:
            tr0_cmd += " --sync_mode"
            tr1_cmd += " --sync_mode"
T
tangwei12 已提交
810 811 812
        if self._hogwild_mode:
            tr0_cmd += " --hogwild"
            tr1_cmd += " --hogwild"
W
Wu Yi 已提交
813 814 815
        if self._use_reduce:
            tr0_cmd += " --use_reduce"
            tr1_cmd += " --use_reduce"
816 817 818
        if self._use_reader_alloc:
            tr0_cmd += " --use_reader_alloc"
            tr1_cmd += " --use_reader_alloc"
819
        if self.__use_cuda:
820 821 822 823 824 825 826 827 828 829
            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
830

W
Wu Yi 已提交
831 832
        print("tr0_cmd: {}, env: {}".format(tr0_cmd, env0))
        print("tr1_cmd: {}, env: {}".format(tr1_cmd, env1))
833 834
        tr0_pipe = open(log_name + "_tr0_err.log", "wb")
        tr1_pipe = open(log_name + "_tr1_err.log", "wb")
G
gongweibao 已提交
835

836
        print_to_err(type(self).__name__, "going to start trainer process 0")
X
Xin Pan 已提交
837
        tr0_proc = subprocess.Popen(
W
Wu Yi 已提交
838
            tr0_cmd.strip().split(" "),
X
Xin Pan 已提交
839
            stdout=subprocess.PIPE,
G
gongweibao 已提交
840
            stderr=tr0_pipe,
X
Xin Pan 已提交
841
            env=env0)
842
        print_to_err(type(self).__name__, "going to start trainer process 1")
X
Xin Pan 已提交
843
        tr1_proc = subprocess.Popen(
W
Wu Yi 已提交
844
            tr1_cmd.strip().split(" "),
X
Xin Pan 已提交
845
            stdout=subprocess.PIPE,
G
gongweibao 已提交
846
            stderr=tr1_pipe,
X
Xin Pan 已提交
847 848
            env=env1)

849 850 851 852 853 854 855 856 857 858 859 860
        # 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

861 862
        tr0_out, tr0_err = tr0_proc.communicate()
        tr1_out, tr1_err = tr1_proc.communicate()
X
Xin Pan 已提交
863

G
gongweibao 已提交
864
        # close trainer file
865 866 867 868
        tr0_pipe.close()
        tr1_pipe.close()
        ps0_pipe.close()
        ps1_pipe.close()
W
Wu Yi 已提交
869

W
Wu Yi 已提交
870 871
        ps0.terminate()
        ps1.terminate()
T
typhoonzero 已提交
872

W
Wu Yi 已提交
873 874
        return pickle.loads(tr0_out), pickle.loads(tr1_out)

875 876 877
    def _get_nccl2_trainer_cmd(self, model, ep, update_method, trainer_id,
                               trainer_num):
        env = {}
878 879 880 881 882 883 884
        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"

885
        tr_cmd = tr_cmd % \
T
tangwei12 已提交
886 887
                 (self._python_interp, model, self._ps_endpoints,
                  trainer_id, ep, update_method, self._lr)
W
Wu Yi 已提交
888 889

        if self._use_reduce:
890
            tr_cmd += " --use_reduce"
W
Wu Yi 已提交
891
        if self._use_reader_alloc:
892
            tr_cmd += " --use_reader_alloc"
893 894
        if self._save_model:
            tr_cmd += " --save_model"
W
Wu Yi 已提交
895
        if self.__use_cuda:
896 897
            tr_cmd += " --use_cuda"
            env.update({
898
                "FLAGS_selected_gpus": "{}".format(0),
899
                "CUDA_VISIBLE_DEVICES": "{}".format(trainer_id % 2),
900
                "PADDLE_TRAINERS_NUM": "{}".format(trainer_num),
901 902 903
                "PADDLE_TRAINER_ID": "{}".format(trainer_id),
                "PADDLE_TRAINER_ENDPOINTS": self._ps_endpoints,
                "PADDLE_CURRENT_ENDPOINT": ep,
904
            })
W
Wu Yi 已提交
905
        else:
906
            env.update({'CPU_NUM': '1'})
W
Wu Yi 已提交
907

908
        if self._use_dgc:
909 910 911
            tr_cmd += " --use_dgc"

        if self._mp_mode:
912
            env = {"FLAGS_selected_gpus": "{}".format(trainer_id % 2)}
913 914

        if self._nccl_comm_num > 1:
915
            tr_cmd += " --nccl_comm_num {}".format(self._nccl_comm_num)
916

917 918
        if self._use_hallreduce:
            tr_cmd += " --use_hallreduce --hallreduce_inter_nranks 2"
919

920
        if self._enable_backward_deps:
921
            tr_cmd += " --enable_backward_deps"
922

923 924
        if self._gpu_fleet_api:
            tr_cmd += " --gpu_fleet_api"
925 926 927 928
            if self._use_local_sgd:
                tr_cmd += " --use_local_sgd"
            if self._ut4grad_allreduce:
                tr_cmd += " --ut4grad_allreduce"
929 930
            if hasattr(self, '_sync_batch_norm') and self._sync_batch_norm:
                tr_cmd += " --sync_batch_norm"
931

932 933 934
        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            env['COVERAGE_FILE'] = os.getenv('COVERAGE_FILE', '')

935
        return tr_cmd, env
W
Wu Yi 已提交
936

937
    def _run_cluster_nccl2(self, model, envs, nccl2_reduce_layer,
938
                           check_error_log, log_name):
939 940
        if self._use_hallreduce:
            self._ps_endpoints = ""
941 942 943 944 945 946 947 948 949 950

            global DIST_UT_PORT
            if DIST_UT_PORT == 0:
                for i in range(0, 4):
                    self._ps_endpoints += "127.0.0.1:%s," % (
                        self._find_free_port())
            else:
                for i in range(0, 4):
                    self._ps_endpoints += "127.0.0.1:%s," % (DIST_UT_PORT + i)
                DIST_UT_PORT += 4
951
            self._ps_endpoints = self._ps_endpoints[:-1]
W
Wu Yi 已提交
952

953 954 955 956 957 958
        # NOTE: we reuse ps_endpoints as nccl2 worker endpoints
        worker_endpoints = self._ps_endpoints.split(",")
        if nccl2_reduce_layer:
            update_method = "nccl2_reduce_layer"
        else:
            update_method = "nccl2"
W
Wu Yi 已提交
959

960
        trainer_num = len(worker_endpoints)
W
Wu Yi 已提交
961

962 963 964 965 966 967 968 969
        procs = []
        pipes = []
        for i in range(0, trainer_num):
            tr_cmd, tr_env = self._get_nccl2_trainer_cmd(
                model, worker_endpoints[i], update_method, i, trainer_num)
            tr_env.update(envs)
            print("use_hallreduce:{} tr_cmd:{}, env: {}".format(
                self._use_hallreduce, tr_cmd, tr_env))
W
Wu Yi 已提交
970

971
            tr_pipe = open(log_name + "_tr{}_err.log".format(i), "wb")
W
Wu Yi 已提交
972

973
            print_to_err(
974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991
                type(self).__name__,
                "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)

            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))

992 993 994
        if check_error_log:
            print("outs[0]:", outs[0])
            print("outs[1]:", outs[1])
995
        return pickle.loads(outs[0]), pickle.loads(outs[1])
996

997
    def _get_required_envs(self, check_error_log=False, need_envs={}):
998 999 1000 1001 1002 1003
        # 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
1004
            "FLAGS_rpc_deadline": "30000",  # 5sec to fail fast
1005
            "FLAGS_rpc_retry_bind_port": "50",
1006
            "FLAGS_cudnn_deterministic": "1",
1007
            "FLAGS_rpc_disable_reuse_port": "1",
W
Wu Yi 已提交
1008
            "http_proxy": "",
1009 1010
            "NCCL_P2P_DISABLE": "1",
            "NCCL_SHM_DISABLE": "1"
1011 1012 1013
        }

        if check_error_log:
1014
            required_envs["GLOG_vmodule"] = \
1015 1016
                "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," \
1017
                "sparse_all_reduce_op_handle=10,gen_nccl_id_op=10,nccl_helper=10,grpc_client=10,grpc_server=10,request_handler_impl=10"
1018 1019
            required_envs["GLOG_logtostderr"] = "1"

1020 1021 1022 1023 1024 1025 1026 1027 1028
        required_envs.update(need_envs)
        return required_envs

    def check_with_place(self,
                         model_file,
                         delta=1e-3,
                         check_error_log=False,
                         need_envs={},
                         log_name=""):
1029

1030 1031
        required_envs = self._get_required_envs(check_error_log, need_envs)

T
tangwei12 已提交
1032
        local_losses \
1033
            = self._run_local(model_file, required_envs,
1034 1035
                              check_error_log, log_name=log_name)

W
Wu Yi 已提交
1036
        if self._nccl2_mode:
W
Wu Yi 已提交
1037 1038
            if self._nccl2_reduce_layer:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
1039 1040 1041 1042 1043
                    model_file,
                    required_envs,
                    True,
                    check_error_log,
                    log_name=log_name)
W
Wu Yi 已提交
1044 1045
            else:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
1046 1047 1048 1049 1050
                    model_file,
                    required_envs,
                    False,
                    check_error_log,
                    log_name=log_name)
W
Wu Yi 已提交
1051 1052
        else:
            tr0_losses, tr1_losses = self._run_cluster(
1053
                model_file, required_envs, check_error_log, log_name=log_name)
1054 1055

        for step_id in range(RUN_STEP):
W
Wu Yi 已提交
1056 1057 1058
            local_loss = local_losses[step_id]
            tr0_loss = tr0_losses[step_id]
            tr1_loss = tr1_losses[step_id]
Y
Yan Xu 已提交
1059
            dist_loss = (np.array([tr0_loss]) + np.array([tr1_loss])) / 2
W
Wu Yi 已提交
1060 1061
            print("=======", local_loss, ":", dist_loss[0], "=======")
            self.assertAlmostEqual(local_loss, dist_loss[0], delta=delta)
1062 1063 1064 1065 1066 1067 1068

    def check_with_place_multi_cards(self,
                                     model_file,
                                     delta=1e-3,
                                     check_error_log=False,
                                     need_envs={},
                                     log_name=""):
1069

1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097
        # 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:
            multi_cards_losses = self._run_local(
                model_file,
                required_envs,
                check_error_log,
                log_name=log_name + "_dgc_2cards",
                gpus="0,1")

            self._use_dgc = False
            base_losses = self._run_local(
                model_file,
                required_envs,
                check_error_log,
                log_name=log_name + "_base_2cards",
                gpus="0,1")

            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)