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
30
import paddle.fluid.dygraph as dygraph
31
import paddle.fluid.incubate.fleet.base.role_maker as role_maker
32 33
from paddle.fluid import compiler
from paddle.fluid.incubate.fleet.collective import DistributedStrategy, fleet
34

Y
Yan Xu 已提交
35
RUN_STEP = 5
36
DEFAULT_BATCH_SIZE = 2
37
DIST_UT_PORT = 0
38

T
typhoonzero 已提交
39

40
def print_to_out(out_losses):
T
tianshuo78520a 已提交
41
    sys.stdout.buffer.write(pickle.dumps(out_losses))
42 43 44


def print_to_err(class_name, log_str):
45 46
    localtime = time.asctime(time.localtime(time.time()))
    print_str = localtime + "\t" + class_name + "\t" + log_str
T
tianshuo78520a 已提交
47
    sys.stderr.buffer.write(pickle.dumps(print_str))
G
guru4elephant 已提交
48 49


50 51 52 53
def eprint(*args, **kwargs):
    print(*args, file=sys.stderr, **kwargs)


54
class TestDistRunnerBase:
55 56 57 58 59 60 61 62
    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
63
        raise NotImplementedError(
64 65
            "get_model should be implemented by child classes."
        )
T
typhoonzero 已提交
66

67
    @staticmethod
68 69 70 71 72 73 74 75 76 77 78
    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
79
        # NOTE: import fluid until runtime, or else forking processes will cause error.
80
        config = fluid.DistributeTranspilerConfig()
W
Wu Yi 已提交
81
        config.enable_dc_asgd = dc_asgd
82
        config.sync_mode = sync_mode
T
tangwei12 已提交
83 84
        config.runtime_split_send_recv = hogwild_mode

85 86
        if nccl_comm_num > 1:
            config.nccl_comm_num = nccl_comm_num
87
        # config.runtime_split_send_recv = True
88
        t = fluid.DistributeTranspiler(config=config)
89 90 91 92 93 94 95 96
        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
97 98
        return t

99 100 101 102 103
    @staticmethod
    def get_lr_scheduler(program):
        lr_sheduler = None
        if hasattr(program, 'lr_sheduler'):
            from paddle.optimizer.lr import LRScheduler
104

105 106 107 108
            lr_sheduler = program.lr_sheduler
            assert isinstance(lr_sheduler, LRScheduler), "must be LRScheduler"
        return lr_sheduler

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

114 115 116 117 118 119 120 121 122
        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
123
        pserver_prog = t.get_pserver_program(args.current_endpoint)
124 125 126
        startup_prog = t.get_startup_program(
            args.current_endpoint, pserver_prog
        )
Y
Yancey1989 已提交
127

T
typhoonzero 已提交
128 129 130
        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        exe.run(startup_prog)
131
        print_to_err(type(self).__name__, "run pserver startup program done.")
T
typhoonzero 已提交
132
        exe.run(pserver_prog)
133
        print_to_err(type(self).__name__, "run pserver main program done.")
T
typhoonzero 已提交
134

135 136 137 138
    def run_pipeline_trainer(self, args):
        self.lr = args.lr

        dist_strategy = DistributedStrategy()
139 140 141 142 143 144 145 146 147 148 149
        (
            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
        )
150 151 152 153 154 155 156 157 158 159 160 161 162

        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 = []
163 164 165

        main_program = fluid.default_main_program()
        lr_sheduler = self.get_lr_scheduler(main_program)
166
        for i in range(RUN_STEP):
167
            loss = exe.run(main_program, fetch_list=[avg_cost])
168 169 170
            loss = loss[0] if loss else None
            out_losses.append(loss)
            print_to_err(type(self).__name__, "run step %d finished" % i)
171 172 173
            if lr_sheduler is not None:
                lr_sheduler.step()

174
        data_loader.reset()
175 176
        print_to_err(type(self).__name__, "trainer run finished")

T
tianshuo78520a 已提交
177
        sys.stdout.buffer.write(pickle.dumps(out_losses))
178

179 180 181 182 183 184 185 186 187 188 189
    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:")

190 191 192 193 194 195 196 197
        (
            test_program,
            avg_cost,
            train_reader,
            test_reader,
            batch_acc,
            predict,
        ) = self.get_model(batch_size=args.batch_size)
198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214

        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 = [
215 216
            var
            for var in fluid.default_main_program().global_block().vars.values()
217 218 219 220 221 222 223 224 225 226 227 228 229
            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)
230 231 232 233
            if (
                paddle.distributed.get_world_size() == 1
                and args.update_method == 'gloo'
            ):  # Gloo single mode
X
xiongkun 已提交
234 235 236
                return origin_batch

            elif args.update_method != "local" and args.use_reader_alloc:
237 238 239 240 241 242 243 244 245 246
                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 = []
247
        for i in range(RUN_STEP):
248 249 250 251 252
            (loss,) = exe.run(
                fluid.default_main_program(),
                fetch_list=[avg_cost.name],
                feed=feeder.feed(get_data()),
            )
253 254 255 256 257
            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
258
        sys.stdout.buffer.write(pickle.dumps(out_losses))
259

260 261
    def run_use_fleet_api_trainer(self, args):
        assert args.update_method == "nccl2" or "bkcl"
262 263 264 265 266 267 268 269

        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
270
        dist_strategy.fuse_memory_size = 1  # MB
271
        dist_strategy.fuse_laryer_size = 1
272 273 274 275
        if args.use_local_sgd:
            dist_strategy.use_local_sgd = True
        if args.ut4grad_allreduce:
            dist_strategy._ut4grad_allreduce = True
276 277
        if args.sync_batch_norm:
            dist_strategy.sync_batch_norm = True
278 279 280

        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
        fleet.init(role)
281
        print_to_err("use_fleet", "fleet.node_num:")
T
tangwei12 已提交
282 283
        # "fleet.node_id:", fleet.node_id(),
        # "fleet.trainer_num:", fleet.worker_num())
284

285 286 287 288 289 290 291 292 293 294
        (
            test_program,
            avg_cost,
            train_reader,
            test_reader,
            batch_acc,
            predict,
        ) = self.get_model(
            batch_size=args.batch_size, dist_strategy=dist_strategy
        )
295 296 297 298

        trainer_prog = fleet._origin_program
        dist_prog = fleet.main_program

299 300 301 302 303 304 305 306 307 308
        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."
            )
309 310 311 312 313 314

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

        feed_var_list = [
315 316
            var
            for var in trainer_prog.global_block().vars.values()
317 318 319
            if var.is_data
        ]

320 321 322 323 324 325 326
        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]

327 328 329 330 331 332 333 334 335 336 337 338 339 340
        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

341
        print_to_err(type(self).__name__, "begin to train on trainer")
342
        out_losses = []
343
        for i in range(RUN_STEP):
344 345 346 347 348
            (loss,) = exe.run(
                dist_prog,
                fetch_list=[avg_cost.name],
                feed=feeder.feed(get_data()),
            )
349
            out_losses.append(loss[0])
350 351
            print_to_err(type(self).__name__, "run step %d finished" % i)
        print_to_err(type(self).__name__, "trainer run finished")
352

T
tianshuo78520a 已提交
353
        sys.stdout.buffer.write(pickle.dumps(out_losses))
354

355 356 357
        if args.save_model:
            model_save_dir = "/tmp"
            if fleet.worker_index() == 0:
358 359 360 361 362 363 364 365 366 367 368 369
                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"
                )
370
            else:
371 372 373 374 375 376 377 378 379 380 381 382
                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"
                )
383
            paddle.distributed.io.save_persistables(
384 385
                exe, model_save_dir_fluid, fleet._origin_program
            )
386 387
            fleet.save_persistables(executor=exe, dirname=model_save_dir_fleet)
            feeded_var_names = [var.name for var in feed_var_list]
388 389 390 391 392 393 394 395 396 397
            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]
            )
398

399
    def run_trainer(self, args):
W
Wu Yi 已提交
400
        self.lr = args.lr
W
Wu Yi 已提交
401
        if args.nccl2_reduce_layer_local_run:
402 403 404 405 406 407 408 409
            (
                test_program,
                avg_cost,
                train_reader,
                test_reader,
                batch_acc,
                predict,
            ) = self.get_model(batch_size=args.batch_size, single_device=True)
410
        elif args.use_dgc:
411 412 413 414 415 416 417 418
            (
                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
419
        else:
420 421 422 423 424 425 426 427
            (
                test_program,
                avg_cost,
                train_reader,
                test_reader,
                batch_acc,
                predict,
            ) = self.get_model(batch_size=args.batch_size)
428

W
Wu Yi 已提交
429
        if args.update_method == "pserver":
430
            print_to_err(
431
                type(self).__name__,
432 433 434 435 436 437 438 439 440 441 442
                "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
443

T
typhoonzero 已提交
444
            trainer_prog = t.get_trainer_program()
445
            print_to_err(
446
                type(self).__name__,
447 448 449 450 451 452
                "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
453 454 455
            # transpile for nccl2
            config = fluid.DistributeTranspilerConfig()
            config.mode = "nccl2"
456
            config.nccl_comm_num = args.nccl_comm_num
457 458
            if args.use_hallreduce:
                config.use_hierarchical_allreduce = True
459 460 461
                config.hierarchical_allreduce_inter_nranks = (
                    args.hallreduce_inter_nranks
                )
462
            print_to_err(
463
                type(self).__name__,
464 465
                "begin to run transpile on trainer with nccl2 mode",
            )
W
Wu Yi 已提交
466
            nccl2_t = fluid.DistributeTranspiler(config=config)
467 468 469 470 471 472 473
            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,
            )
474
            print_to_err(
475 476
                type(self).__name__, "get trainer program done. with nccl2 mode"
            )
W
Wu Yi 已提交
477
            trainer_prog = fluid.default_main_program()
T
typhoonzero 已提交
478
        else:
479
            print_to_err(
480
                type(self).__name__,
481 482
                "do nothing about main program, just use it",
            )
T
typhoonzero 已提交
483
            trainer_prog = fluid.default_main_program()
484
            print_to_err(type(self).__name__, "use main program done.")
T
typhoonzero 已提交
485

486 487 488
        # FIXME(gongwb):wait pserver initialization.
        time.sleep(1)

489
        if args.use_cuda:
490 491
            device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
            place = fluid.CUDAPlace(device_id)
492 493 494
        else:
            place = fluid.CPUPlace()

495 496
        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
497
        print_to_err(type(self).__name__, "run worker startup program done.")
T
typhoonzero 已提交
498

W
Wu Yi 已提交
499 500
        exec_strategy = fluid.ExecutionStrategy()
        exec_strategy.num_threads = 1
501

W
Wu Yi 已提交
502
        build_stra = fluid.BuildStrategy()
503 504 505
        # FIXME force disable enable_inplace and memory_optimize
        build_stra.enable_inplace = False
        build_stra.memory_optimize = False
W
Wu Yi 已提交
506

507 508 509 510
        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
511 512 513
        if args.hogwild:
            build_stra.async_mode = True

514 515 516
        if args.enable_backward_deps:
            build_stra.enable_backward_optimizer_op_deps = True

W
Wu Yi 已提交
517
        if args.use_reduce:
518 519 520
            build_stra.reduce_strategy = (
                fluid.BuildStrategy.ReduceStrategy.Reduce
            )
W
Wu Yi 已提交
521
        else:
522 523 524
            build_stra.reduce_strategy = (
                fluid.BuildStrategy.ReduceStrategy.AllReduce
            )
W
Wu Yi 已提交
525

W
Wu Yi 已提交
526
        pass_builder = None
X
Xin Pan 已提交
527
        if args.batch_merge_repeat > 1:
X
fix  
Xin Pan 已提交
528
            pass_builder = build_stra._finalize_strategy_and_create_passes()
529
            mypass = pass_builder.insert_pass(0, "multi_batch_merge_pass")
530
            mypass.set("num_repeats", args.batch_merge_repeat)
X
Xin Pan 已提交
531

532 533 534 535
        if (
            args.update_method == "nccl2"
            or args.update_method == "nccl2_reduce_layer"
        ):
536 537
            build_stra.num_trainers = len(args.endpoints.split(","))
            build_stra.trainer_id = args.trainer_id
W
Wu Yi 已提交
538
        else:
W
Wu Yi 已提交
539
            # case args.update_method == "nccl2_reduce_layer":
540 541
            build_stra.num_trainers = 1
            build_stra.trainer_id = 0
W
Wu Yi 已提交
542

543
        print_to_err(type(self).__name__, "begin to compile with data parallel")
X
Xin Pan 已提交
544
        binary = compiler.CompiledProgram(trainer_prog).with_data_parallel(
W
Wu Yi 已提交
545
            loss_name=avg_cost.name,
W
Wu Yi 已提交
546
            build_strategy=build_stra,
547 548
            exec_strategy=exec_strategy,
        )
549
        print_to_err(type(self).__name__, "program compiled with data parallel")
T
typhoonzero 已提交
550 551

        feed_var_list = [
552 553
            var
            for var in trainer_prog.global_block().vars.values()
T
typhoonzero 已提交
554 555 556 557
            if var.is_data
        ]

        feeder = fluid.DataFeeder(feed_var_list, place)
558
        reader_generator = train_reader()
T
typhoonzero 已提交
559

560 561
        def get_data():
            origin_batch = next(reader_generator)
W
Wu Yi 已提交
562
            if args.update_method != "local" and args.use_reader_alloc:
563 564 565 566 567 568 569
                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
570

571
        lr_scheduler = self.get_lr_scheduler(trainer_prog)
572
        print_to_err(type(self).__name__, "begin to train on trainer")
W
Wu Yi 已提交
573
        out_losses = []
574
        for i in range(RUN_STEP):
575 576 577
            (loss,) = exe.run(
                binary, fetch_list=[avg_cost.name], feed=feeder.feed(get_data())
            )
W
Wu Yi 已提交
578
            out_losses.append(loss[0])
579
            print_to_err(type(self).__name__, "run step %d finished" % i)
580 581 582
            if lr_scheduler is not None:
                lr_scheduler.step()

583
        print_to_err(type(self).__name__, "trainer run finished")
584

585
        print_to_out(out_losses)
T
typhoonzero 已提交
586 587


588
class TestParallelDyGraphRunnerBase:
589 590
    def get_model(self):
        raise NotImplementedError(
591 592
            "get_model should be implemented by child classes."
        )
593 594 595

    def run_one_loop(self, model, opt, data):
        raise NotImplementedError(
596 597
            "train_one_loop should be implemented by the child classes."
        )
598

599
    def _get_data(self, batch, args):
600 601 602 603
        if (
            paddle.distributed.get_world_size() == 1
            and args.update_method == 'gloo'
        ):  # Gloo single mode
X
xiongkun 已提交
604 605
            return batch
        elif args.update_method != "local":
606
            new_batch = []
607

608 609 610
            # 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].
611 612
            # this function is for test sparse_embedding_differ_length
            if hasattr(args, "diff_batch") and args.diff_batch:
613 614 615
                assert (
                    len(batch) > 2
                ), "in differ_batch mode, len(batch) must > 2."
616 617 618 619 620 621 622 623 624 625 626 627 628 629
                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
630 631 632
        else:
            return batch

633 634
    def run_trainer(self, args):
        seed = 90
X
xiongkun 已提交
635 636 637
        if args.update_method == 'gloo':
            place = fluid.CPUPlace()
        elif fluid.core.is_compiled_with_cuda():
638 639 640 641 642
            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)
643 644 645
        elif fluid.core.is_compiled_with_npu():
            device_id = int(os.getenv("FLAGS_selected_npus", "0"))
            place = fluid.NPUPlace(device_id)
646 647 648
        elif fluid.core.is_compiled_with_mlu():
            device_id = int(os.getenv("FLAGS_selected_mlus", "0"))
            place = fluid.MLUPlace(device_id)
649
        else:
650
            assert "Only support CUDAPlace or XPUPlace or CPU(Gloo) for now."
651 652 653 654

        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
655 656
            np.random.seed(seed)
            import random
657

658
            random.seed(seed)
659 660
            model, train_reader, opt = self.get_model()
            nranks = len(args.endpoints.split(",")) if args.endpoints else 1
Y
Yan Xu 已提交
661

662 663 664 665 666 667 668
            # 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"
            ):
669 670 671 672 673
                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
674
                paddle.distributed.init_parallel_env()
675
                print_to_err(
676
                    type(self).__name__,
677 678
                    "begin to prepare context in dygraph with nccl2",
                )
679 680
                if not args.find_unused_parameters:
                    model = dygraph.parallel.DataParallel(
681 682
                        model, strategy, find_unused_parameters=False
                    )
683 684
                else:
                    model = dygraph.parallel.DataParallel(
685 686
                        model, strategy, find_unused_parameters=True
                    )
687
                print_to_err(type(self).__name__, "model built in dygraph")
X
xiongkun 已提交
688 689 690 691 692

            elif args.update_method == "gloo":
                paddle.distributed.init_parallel_env()
                if not args.find_unused_parameters:
                    model = dygraph.parallel.DataParallel(
693 694
                        model, find_unused_parameters=False
                    )
X
xiongkun 已提交
695 696
                else:
                    model = dygraph.parallel.DataParallel(
697 698
                        model, find_unused_parameters=True
                    )
X
xiongkun 已提交
699

700
            out_losses = []
701
            print_to_err(type(self).__name__, "begin to run dygraph training")
702
            for step_id, data in enumerate(train_reader()):
703
                data = self._get_data(data, args)
704 705 706
                if step_id == RUN_STEP:
                    break
                loss = self.run_one_loop(model, opt, data)
G
guru4elephant 已提交
707
                if step_id % 10 == 0:
708
                    print_to_err(
709
                        type(self).__name__,
710 711
                        "loss at step %d: %f" % (step_id, loss.numpy()),
                    )
Y
Yan Xu 已提交
712
                out_losses.append(loss.numpy())
713 714 715 716

                loss.backward()

                opt.minimize(loss)
717 718
                if not args.accumulate_gradient:
                    model.clear_gradients()
719
        print_to_out(out_losses)
720

721 722 723 724 725 726 727 728 729
    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)
730
        random.seed(seed)
731
        # get trainer id
L
LiYuRio 已提交
732 733
        paddle.distributed.parallel._get_global_parallel_env()
        args.trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
734 735

        # 3. init parallel env
X
xiongkun 已提交
736
        if args.update_method in ["nccl2", "gloo"]:
737 738 739 740
            paddle.distributed.init_parallel_env()

        # 4. train model
        model, train_reader, opt = self.get_model()
X
xiongkun 已提交
741
        if args.update_method in ["nccl2", "gloo"]:
742
            model = paddle.DataParallel(
743 744
                model, find_unused_parameters=args.find_unused_parameters
            )
745 746 747 748 749 750 751 752 753 754 755 756 757 758 759

        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

760
    def run_use_fleet_api_trainer(self, args):
761
        import paddle.distributed.fleet as fleet
762

763 764 765 766 767 768 769 770
        # 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)
771
        random.seed(seed)
772
        # get trainer id
L
LiYuRio 已提交
773 774
        paddle.distributed.parallel._get_global_parallel_env()
        args.trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
775

776 777
        # set strategy
        strategy = fleet.DistributedStrategy()
778 779
        if args.find_unused_parameters:
            strategy.find_unused_parameters = True
780

781
        # 3. init parallel env
782
        if args.update_method == "nccl2" or "bkcl" or "hccl":
783
            fleet.init(is_collective=True, strategy=strategy)
784 785 786

        # 4. train model
        model, train_reader, opt = self.get_model()
787
        if args.update_method == "nccl2" or "bkcl" or "hccl":
788 789 790 791 792 793 794 795 796 797 798 799 800 801
            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()
802 803
            if not args.accumulate_gradient:
                opt.clear_grad()
804 805
        print_to_out(out_losses)

806

T
typhoonzero 已提交
807
def runtime_main(test_class):
W
Wu Yi 已提交
808
    parser = argparse.ArgumentParser(description='Run dist test.')
809 810 811
    parser.add_argument(
        '--role', type=str, required=True, choices=['pserver', 'trainer']
    )
W
Wu Yi 已提交
812
    parser.add_argument('--endpoints', type=str, required=False, default="")
813 814 815 816 817 818 819 820 821 822 823 824 825 826 827
    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
828 829
    parser.add_argument('--trainer_id', type=int, required=False, default=0)
    parser.add_argument('--trainers', type=int, required=False, default=1)
830
    parser.add_argument('--nccl_comm_num', type=int, required=False, default=1)
831 832
    parser.add_argument('--enable_backward_deps', action='store_true')
    parser.add_argument('--use_hallreduce', action='store_true')
833
    parser.add_argument('--use_pipeline', action='store_true')
834
    parser.add_argument('--use_fleet_api', action='store_true')
835
    parser.add_argument('--use_fleet_api_20', action='store_true')
836
    parser.add_argument('--use_local_sgd', action='store_true')
837
    parser.add_argument('--diff_batch', action='store_true')
838
    parser.add_argument('--ut4grad_allreduce', action='store_true')
839 840 841 842 843 844
    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
845
    parser.add_argument('--sync_mode', action='store_true')
846
    parser.add_argument('--use_cuda', action='store_true')
X
xiongkun 已提交
847
    parser.add_argument('--use_cpu', action='store_true')
848
    parser.add_argument('--use_xpu', action='store_true')
849
    parser.add_argument('--use_dgc', action='store_true')
850
    parser.add_argument('--use_npu', action='store_true')
851
    parser.add_argument('--use_mlu', action='store_true')
852
    parser.add_argument('--accumulate_gradient', action='store_true')
853
    parser.add_argument('--find_unused_parameters', action='store_true')
W
Wu Yi 已提交
854
    parser.add_argument('--use_reduce', action='store_true')
W
Wu Yi 已提交
855
    parser.add_argument('--dc_asgd', action='store_true')
T
tangwei12 已提交
856
    parser.add_argument('--hogwild', action='store_true')
857
    parser.add_argument('--save_model', action='store_true')
858 859 860
    parser.add_argument(
        '--use_reader_alloc', action='store_true', required=False
    )
861
    parser.add_argument('--batch_size', required=False, type=int, default=2)
W
Wu Yi 已提交
862
    parser.add_argument('--lr', required=False, type=float, default=0.001)
863 864 865 866 867 868 869 870 871
    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,
    )
872
    parser.add_argument('--sync_batch_norm', action='store_true')
873 874 875
    parser.add_argument(
        '--fuse_all_reduce', required=False, type=ast.literal_eval, default=None
    )
W
Wu Yi 已提交
876 877

    args = parser.parse_args()
T
typhoonzero 已提交
878

X
xiongkun 已提交
879 880 881
    if args.update_method == 'gloo':
        paddle.set_device("cpu")

T
typhoonzero 已提交
882
    model = test_class()
W
Wu Yi 已提交
883
    if args.role == "pserver" and args.update_method == "pserver":
W
Wu Yi 已提交
884
        model.run_pserver(args)
885 886
    elif args.use_fleet_api:
        model.run_use_fleet_api_trainer(args)
887 888
    elif args.use_fleet_api_20:
        model.run_use_fleet_api_20_trainer(args)
889 890
    elif args.use_pipeline:
        model.run_pipeline_trainer(args)
T
typhoonzero 已提交
891
    else:
892
        model.run_trainer(args)
X
Xin Pan 已提交
893

M
minqiyang 已提交
894

Y
Yancey1989 已提交
895 896
import socket
from contextlib import closing
M
minqiyang 已提交
897

X
Xin Pan 已提交
898 899

class TestDistBase(unittest.TestCase):
W
Wu Yi 已提交
900 901 902
    def _setup_config(self):
        raise NotImplementedError("tests should have _setup_config implemented")

903 904 905
    def _after_setup_config(self):
        if self._enforce_place == "CPU":
            self.__use_cuda = False
906
            self.__use_xpu = False
907
            self._use_dgc = False
908
            self.__use_npu = False
909
            self._use_mlu = False
910 911
        elif self._enforce_place == "GPU":
            self.__use_cuda = True
912
            self.__use_xpu = False
913
            self.__use_npu = False
914
            self._use_mlu = False
915 916 917 918
        elif self._enforce_place == "XPU":
            self.__use_cuda = False
            self.__use_xpu = True
            self._use_dgc = False
919
            self.__use_npu = False
920
            self._use_mlu = False
921 922 923 924 925
        elif self._enforce_place == "NPU":
            self.__use_cuda = False
            self.__use_xpu = False
            self._use_dgc = False
            self.__use_npu = True
926 927 928 929 930 931 932
            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
933 934 935 936 937
        else:
            if fluid.core.is_compiled_with_cuda():
                self.__use_cuda = True
            else:
                self.__use_cuda = False
938 939 940 941
                self._use_dgc = False

        if self._use_reduce:
            assert not self._use_dgc
942

X
Xin Pan 已提交
943 944 945
    def setUp(self):
        self._trainers = 2
        self._pservers = 2
Y
Yancey1989 已提交
946
        self._port_set = set()
M
minqiyang 已提交
947
        self._python_interp = sys.executable
W
Wu Yi 已提交
948
        self._sync_mode = True
T
tangwei12 已提交
949
        self._hogwild_mode = False
950
        self._enforce_place = None
W
Wu Yi 已提交
951
        self._use_reduce = False
W
Wu Yi 已提交
952
        self._dc_asgd = False  # must use with async mode
953
        self._use_reader_alloc = True
W
Wu Yi 已提交
954
        self._nccl2_mode = False
955
        self._bkcl_mode = False
X
xiongkun 已提交
956
        self._gloo_mode = False  # now, support gloo backend
957
        self._hccl_mode = False
958
        self._cncl_mode = False
959
        self._pipeline_mode = False
960
        self._mp_mode = False
961
        self._diff_batch = False
W
Wu Yi 已提交
962 963 964 965 966
        # 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
967
        self._lr = 0.001
968
        self._use_dgc = False
969
        self._dygraph = False
970
        self._nccl_comm_num = 1
971
        self._enable_backward_deps = False
972
        self._use_fleet_api = False
973
        self._use_fleet_api_20 = False
974 975
        self._use_local_sgd = False
        self._ut4grad_allreduce = False
976
        self._use_hallreduce = False
977
        self._save_model = False
978
        self._fuse_all_reduce = None
979
        self._accumulate_gradient = False
980
        self._find_unused_parameters = False
W
Wu Yi 已提交
981
        self._setup_config()
982 983 984 985 986 987 988

        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" % (
989 990 991
                self._find_free_port(),
                self._find_free_port(),
            )
992 993
        else:
            self._ps_endpoints = "127.0.0.1:%s,127.0.0.1:%s" % (
994 995 996
                DIST_UT_PORT,
                DIST_UT_PORT + 1,
            )
997
            DIST_UT_PORT += 2
998
            self._dist_port = DIST_UT_PORT
999

1000
        self._after_setup_config()
X
Xin Pan 已提交
1001

1002 1003 1004 1005 1006
        self.temp_dir = tempfile.TemporaryDirectory()

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

Y
Yancey1989 已提交
1007
    def _find_free_port(self):
Y
Yancey1989 已提交
1008
        def __free_port():
1009 1010 1011
            with closing(
                socket.socket(socket.AF_INET, socket.SOCK_STREAM)
            ) as s:
Y
Yancey1989 已提交
1012
                s.bind(('', 0))
1013
                print_to_err(
1014 1015
                    type(self).__name__, "socket name: %s" % s.getsockname()[1]
                )
Y
Yancey1989 已提交
1016 1017 1018 1019 1020 1021 1022
                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
1023

1024 1025 1026
    def start_pserver(
        self, model_file, check_error_log, required_envs, log_name=""
    ):
X
Xin Pan 已提交
1027
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
1028 1029 1030 1031 1032 1033 1034 1035
        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"

1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049
        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
1050 1051 1052 1053

        if self._sync_mode:
            ps0_cmd += " --sync_mode"
            ps1_cmd += " --sync_mode"
X
Xin Pan 已提交
1054

1055 1056
        print(ps0_cmd)
        print(ps1_cmd)
1057 1058 1059 1060
        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
1061

1062
        print_to_err(type(self).__name__, "going to start pserver process 0")
1063 1064 1065 1066 1067 1068
        ps0_proc = subprocess.Popen(
            ps0_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps0_pipe,
            env=required_envs,
        )
1069
        print_to_err(type(self).__name__, "going to start pserver process 1")
1070 1071 1072 1073 1074 1075
        ps1_proc = subprocess.Popen(
            ps1_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps1_pipe,
            env=required_envs,
        )
G
gongweibao 已提交
1076

1077
        return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe
X
Xin Pan 已提交
1078

1079 1080 1081 1082 1083 1084 1085 1086 1087 1088
    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
1089

1090 1091 1092 1093 1094 1095
        cmd = self._python_interp

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

1096 1097 1098 1099
        cmd += " %s --role trainer --update_method local --lr %f" % (
            model,
            self._lr,
        )
1100

1101 1102 1103 1104
        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
1105 1106
        if self._nccl2_reduce_layer:
            cmd += " --nccl2_reduce_layer_local_run 1"
1107

1108
        if self.__use_cuda:
1109
            cmd += " --use_cuda"
W
Wu Yi 已提交
1110
            env_local = {
1111 1112
                "CUDA_VISIBLE_DEVICES": devices,
                "PADDLE_TRAINERS_NUM": "1",
1113
                "PADDLE_TRAINER_ID": "0",
1114 1115 1116 1117 1118
            }
        elif self.__use_xpu:
            cmd += " --use_xpu"
            env_local = {
                "FLAGS_selected_xpus": devices,
W
Wu Yi 已提交
1119
                "PADDLE_TRAINERS_NUM": "1",
1120
                "PADDLE_TRAINER_ID": "0",
W
Wu Yi 已提交
1121
            }
1122 1123 1124 1125 1126
        elif self.__use_npu:
            cmd += " --use_npu"
            env_local = {
                "FLAGS_selected_npus": devices,
                "PADDLE_TRAINERS_NUM": "1",
1127
                "PADDLE_TRAINER_ID": "0",
1128
            }
1129 1130 1131
        else:
            env_local = {'CPU_NUM': '1'}

1132
        # not use dgc in single card
1133
        if len(devices) > 1 and self._use_dgc:
1134 1135
            cmd += " --use_dgc"

1136 1137 1138
        if self._accumulate_gradient:
            cmd += " --accumulate_gradient"

1139 1140 1141
        if self._find_unused_parameters:
            cmd += " --find_unused_parameters"

W
Wu Yi 已提交
1142 1143
        env_local.update(envs)
        print("local_cmd: {}, env: {}".format(cmd, env_local))
G
gongweibao 已提交
1144

1145
        if check_error_log:
1146 1147
            path = os.path.join(self.temp_dir.name, log_name + "_local.log")
            err_log = open(path, "wb")
1148 1149 1150 1151 1152 1153
            local_proc = subprocess.Popen(
                cmd.split(" "),
                stdout=subprocess.PIPE,
                stderr=err_log,
                env=env_local,
            )
G
gongweibao 已提交
1154
        else:
1155 1156 1157 1158 1159 1160
            local_proc = subprocess.Popen(
                cmd.split(" "),
                stdout=subprocess.PIPE,
                stderr=subprocess.PIPE,
                env=env_local,
            )
G
gongweibao 已提交
1161

1162 1163 1164 1165 1166 1167
        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
1168
        sys.stderr.write('local_stdout: %s\n' % pickle.loads(local_out))
X
Xin Pan 已提交
1169

W
Wu Yi 已提交
1170
        return pickle.loads(local_out)
1171

1172 1173 1174 1175 1176 1177 1178 1179 1180 1181
    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
1182 1183
        saved_endpoints = self._ps_endpoints
        self._ps_endpoints = self._ps_endpoints.split(',')[0]
1184 1185 1186
        result = self._run_cluster_gloo(
            model, envs, 'gloo', check_error_log, log_name
        )
X
xiongkun 已提交
1187 1188 1189
        self._ps_endpoints = saved_endpoints
        return result

1190
    def _run_cluster(self, model, envs, check_error_log, log_name):
X
Xin Pan 已提交
1191
        # Run dist train to compare with local results
1192 1193 1194
        ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(
            model, check_error_log, envs, log_name=log_name
        )
W
Wu Yi 已提交
1195

X
Xin Pan 已提交
1196
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
1197

1198 1199 1200 1201 1202 1203 1204 1205
        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"

1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223
        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
1224 1225 1226 1227

        if self._sync_mode:
            tr0_cmd += " --sync_mode"
            tr1_cmd += " --sync_mode"
T
tangwei12 已提交
1228 1229 1230
        if self._hogwild_mode:
            tr0_cmd += " --hogwild"
            tr1_cmd += " --hogwild"
W
Wu Yi 已提交
1231 1232 1233
        if self._use_reduce:
            tr0_cmd += " --use_reduce"
            tr1_cmd += " --use_reduce"
1234 1235 1236
        if self._use_reader_alloc:
            tr0_cmd += " --use_reader_alloc"
            tr1_cmd += " --use_reader_alloc"
1237
        if self.__use_cuda:
1238 1239 1240 1241 1242 1243 1244 1245 1246 1247
            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
1248

W
Wu Yi 已提交
1249 1250
        print("tr0_cmd: {}, env: {}".format(tr0_cmd, env0))
        print("tr1_cmd: {}, env: {}".format(tr1_cmd, env1))
1251 1252 1253 1254 1255

        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
1256

1257
        print_to_err(type(self).__name__, "going to start trainer process 0")
1258 1259 1260 1261 1262 1263
        tr0_proc = subprocess.Popen(
            tr0_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=tr0_pipe,
            env=env0,
        )
1264
        print_to_err(type(self).__name__, "going to start trainer process 1")
1265 1266 1267 1268 1269 1270
        tr1_proc = subprocess.Popen(
            tr1_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=tr1_pipe,
            env=env1,
        )
X
Xin Pan 已提交
1271

1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283
        # 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

1284 1285
        tr0_out, tr0_err = tr0_proc.communicate()
        tr1_out, tr1_err = tr1_proc.communicate()
X
Xin Pan 已提交
1286

G
gongweibao 已提交
1287
        # close trainer file
1288 1289 1290 1291
        tr0_pipe.close()
        tr1_pipe.close()
        ps0_pipe.close()
        ps1_pipe.close()
W
Wu Yi 已提交
1292

W
Wu Yi 已提交
1293 1294
        ps0.terminate()
        ps1.terminate()
T
typhoonzero 已提交
1295

W
Wu Yi 已提交
1296 1297
        return pickle.loads(tr0_out), pickle.loads(tr1_out)

1298 1299 1300
    def _get_gloo_trainer_cmd(
        self, model, ep, update_method, trainer_id, trainer_num
    ):
X
xiongkun 已提交
1301 1302 1303 1304 1305 1306 1307 1308
        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"

1309 1310 1311 1312 1313 1314 1315 1316 1317
        tr_cmd = tr_cmd % (
            self._python_interp,
            model,
            self._ps_endpoints,
            trainer_id,
            ep,
            update_method,
            self._lr,
        )
X
xiongkun 已提交
1318 1319 1320 1321 1322

        if self._use_reduce:
            tr_cmd += " --use_reduce"
        if self._use_reader_alloc:
            tr_cmd += " --use_reader_alloc"
1323 1324
        # 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
1325 1326
        if self._save_model:
            tr_cmd += " --save_model"
1327 1328
        if self._diff_batch:
            tr_cmd += " --diff_batch"
X
xiongkun 已提交
1329 1330
        self.__use_cuda = False
        self.__use_xpu = False
1331 1332
        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
1333
        tr_cmd += " --use_cpu"
1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344
        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
1345

1346
        assert not self._use_dgc, "gloo not support use dgc"
1347

X
xiongkun 已提交
1348 1349 1350 1351 1352 1353
        if self._accumulate_gradient:
            tr_cmd += " --accumulate_gradient"

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

1354
        assert not self._pipeline_mode, "gloo not support use pipeline"
X
xiongkun 已提交
1355 1356 1357 1358 1359 1360 1361

        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)

1362 1363
        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
1364 1365
        return tr_cmd, env

1366 1367 1368
    def _get_nccl2_trainer_cmd(
        self, model, ep, update_method, trainer_id, trainer_num
    ):
1369
        env = {}
1370 1371 1372 1373 1374 1375 1376
        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"

1377 1378 1379 1380 1381 1382 1383 1384 1385
        tr_cmd = tr_cmd % (
            self._python_interp,
            model,
            self._ps_endpoints,
            trainer_id,
            ep,
            update_method,
            self._lr,
        )
W
Wu Yi 已提交
1386 1387

        if self._use_reduce:
1388
            tr_cmd += " --use_reduce"
W
Wu Yi 已提交
1389
        if self._use_reader_alloc:
1390
            tr_cmd += " --use_reader_alloc"
1391 1392
        if self._save_model:
            tr_cmd += " --save_model"
W
Wu Yi 已提交
1393
        if self.__use_cuda:
1394
            tr_cmd += " --use_cuda"
1395 1396 1397 1398 1399 1400 1401 1402 1403 1404
            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,
                }
            )
1405 1406 1407 1408
        # 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"
1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419
            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",
                }
            )
1420 1421
        elif self.__use_npu:
            tr_cmd += " --use_npu"
1422 1423 1424 1425 1426 1427 1428 1429 1430 1431
            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",
                }
            )
1432 1433
        elif self._use_mlu:
            tr_cmd += " --use_mlu"
1434 1435 1436 1437 1438 1439 1440 1441 1442 1443
            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
1444
        else:
1445
            env.update({'CPU_NUM': '1'})
W
Wu Yi 已提交
1446

1447
        if self._use_dgc:
1448 1449
            tr_cmd += " --use_dgc"

1450 1451 1452
        if self._accumulate_gradient:
            tr_cmd += " --accumulate_gradient"

1453 1454 1455
        if self._find_unused_parameters:
            tr_cmd += " --find_unused_parameters"

1456 1457
        if self._pipeline_mode:
            tr_cmd += " --use_pipeline"
1458
        if self._mp_mode:
W
WangXi 已提交
1459
            env = {"FLAGS_selected_gpus": "{}".format(trainer_id)}
1460 1461

        if self._nccl_comm_num > 1:
1462
            tr_cmd += " --nccl_comm_num {}".format(self._nccl_comm_num)
1463

1464 1465
        if self._use_hallreduce:
            tr_cmd += " --use_hallreduce --hallreduce_inter_nranks 2"
1466

1467
        if self._enable_backward_deps:
1468
            tr_cmd += " --enable_backward_deps"
1469

1470 1471 1472
        if self._fuse_all_reduce is not None:
            tr_cmd += " --fuse_all_reduce {}".format(self._fuse_all_reduce)

1473
        if self._use_fleet_api:
1474 1475 1476 1477 1478
            tr_cmd += (
                " --use_fleet_api_20"
                if self._use_fleet_api_20
                else " --use_fleet_api"
            )
1479 1480 1481 1482
            if self._use_local_sgd:
                tr_cmd += " --use_local_sgd"
            if self._ut4grad_allreduce:
                tr_cmd += " --ut4grad_allreduce"
1483 1484
            if hasattr(self, '_sync_batch_norm') and self._sync_batch_norm:
                tr_cmd += " --sync_batch_norm"
1485

1486 1487 1488
        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
            env['COVERAGE_FILE'] = os.getenv('COVERAGE_FILE', '')

1489
        return tr_cmd, env
W
Wu Yi 已提交
1490

1491 1492 1493 1494 1495 1496 1497 1498 1499 1500
    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
1501 1502 1503 1504 1505 1506 1507 1508

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

        trainer_num = len(worker_endpoints)

        procs = []
        pipes = []
        for i in range(0, trainer_num):
1509 1510 1511
            tr_cmd, tr_env = self._get_gloo_trainer_cmd(
                model, worker_endpoints[i], update_method, i, trainer_num
            )
X
xiongkun 已提交
1512 1513 1514
            tr_env.update(envs)
            tr_env["GLOG_vmodule"] = 'gloo_context=4'
            tr_env["GLOG_v"] = '3'
1515 1516 1517 1518 1519
            print(
                "use_hallreduce:{} tr_cmd:{}, env: {}".format(
                    self._use_hallreduce, tr_cmd, tr_env
                )
            )
X
xiongkun 已提交
1520

1521 1522 1523
            path = os.path.join(
                self.temp_dir.name, log_name + "_tr{}_err.log".format(i)
            )
1524
            tr_pipe = open(path, "wb")
X
xiongkun 已提交
1525 1526 1527

            print_to_err(
                type(self).__name__,
1528 1529 1530 1531 1532 1533 1534 1535
                "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
1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547

            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:
1548 1549
            if check_error_log:
                print("outs[0]:", outs[0])
X
xiongkun 已提交
1550 1551 1552 1553 1554 1555 1556 1557
            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])

1558 1559 1560
    def _run_cluster_nccl2(
        self, model, envs, update_method, check_error_log, log_name
    ):
1561 1562
        if self._use_hallreduce:
            self._ps_endpoints = ""
1563 1564 1565

            global DIST_UT_PORT
            if DIST_UT_PORT == 0:
W
WangXi 已提交
1566
                # NOTE(wangxi). hallreduce test must use 4cards after nccl>=2.7
1567 1568
                for i in range(0, 4):
                    self._ps_endpoints += "127.0.0.1:%s," % (
1569 1570
                        self._find_free_port()
                    )
1571 1572 1573 1574
            else:
                for i in range(0, 4):
                    self._ps_endpoints += "127.0.0.1:%s," % (DIST_UT_PORT + i)
                DIST_UT_PORT += 4
1575
            self._ps_endpoints = self._ps_endpoints[:-1]
W
Wu Yi 已提交
1576

1577 1578
        # NOTE: we reuse ps_endpoints as nccl2 worker endpoints
        worker_endpoints = self._ps_endpoints.split(",")
W
Wu Yi 已提交
1579

1580
        trainer_num = len(worker_endpoints)
W
Wu Yi 已提交
1581

1582 1583 1584 1585
        procs = []
        pipes = []
        for i in range(0, trainer_num):
            tr_cmd, tr_env = self._get_nccl2_trainer_cmd(
1586 1587
                model, worker_endpoints[i], update_method, i, trainer_num
            )
1588
            tr_env.update(envs)
1589 1590 1591 1592 1593
            print(
                "use_hallreduce:{} tr_cmd:{}, env: {}".format(
                    self._use_hallreduce, tr_cmd, tr_env
                )
            )
W
Wu Yi 已提交
1594

1595 1596 1597
            path = os.path.join(
                self.temp_dir.name, log_name + "_tr{}_err.log".format(i)
            )
1598
            tr_pipe = open(path, "wb")
W
Wu Yi 已提交
1599

1600
            print_to_err(
1601
                type(self).__name__,
1602 1603 1604 1605 1606 1607 1608 1609
                "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,
            )
1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620

            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))

1621 1622 1623
        if check_error_log:
            print("outs[0]:", outs[0])
            print("outs[1]:", outs[1])
1624

1625
        return pickle.loads(outs[0]), pickle.loads(outs[1])
1626

1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637
    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(
1638 1639
                model, worker_endpoints[i], update_method, i, trainer_num
            )
1640 1641 1642 1643 1644 1645 1646
            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))

1647 1648
            path = os.path.join(self.temp_dir.name + "tr{}_err.log".format(i))
            tr_pipe = open(path, "wb")
1649 1650 1651

            print_to_err(
                type(self).__name__,
1652 1653 1654 1655 1656 1657 1658 1659
                "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,
            )
1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675

            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])

1676
    def _get_required_envs(self, check_error_log=False, need_envs={}):
1677 1678 1679 1680 1681 1682
        # 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
1683
            "FLAGS_rpc_deadline": "30000",  # 5sec to fail fast
1684
            "FLAGS_rpc_retry_bind_port": "50",
1685
            "FLAGS_cudnn_deterministic": "1",
1686
            "FLAGS_rpc_disable_reuse_port": "1",
W
Wu Yi 已提交
1687
            "http_proxy": "",
1688
            "NCCL_P2P_DISABLE": "1",
1689
            "NCCL_SHM_DISABLE": "1",
1690
            "FLAGS_CONVERT_GRAPH_TO_PROGRAM": "1",
1691 1692 1693
        }

        if check_error_log:
1694 1695 1696 1697
            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,"
1698
                "grpc_server=10,request_handler_impl=10,section_worker=10"
1699
            )
1700 1701
            required_envs["GLOG_logtostderr"] = "1"

1702 1703
        if os.getenv('NVIDIA_TF32_OVERRIDE', '') is not None:
            required_envs['NVIDIA_TF32_OVERRIDE'] = os.getenv(
1704 1705
                'NVIDIA_TF32_OVERRIDE', ''
            )
1706

1707 1708 1709
        required_envs.update(need_envs)
        return required_envs

1710 1711 1712 1713 1714 1715 1716 1717
    def check_with_place(
        self,
        model_file,
        delta=1e-3,
        check_error_log=False,
        need_envs={},
        log_name="",
    ):
1718
        if self._dygraph and (self._gloo_mode or self._nccl2_mode):
1719 1720 1721 1722 1723 1724 1725
            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,
            )
1726
        else:
1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742
            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="",
    ):
1743 1744
        required_envs = self._get_required_envs(check_error_log, need_envs)

X
xiongkun 已提交
1745
        if self._gloo_mode:
1746 1747 1748
            local_losses = self._run_local_gloo(
                model_file, required_envs, check_error_log, log_name=log_name
            )
X
xiongkun 已提交
1749
        else:
1750 1751 1752
            local_losses = self._run_local(
                model_file, required_envs, check_error_log, log_name=log_name
            )
1753

W
Wu Yi 已提交
1754
        if self._nccl2_mode:
W
Wu Yi 已提交
1755 1756
            if self._nccl2_reduce_layer:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
1757 1758
                    model_file,
                    required_envs,
1759 1760
                    update_method="nccl2_reduce_layer",
                    check_error_log=check_error_log,
1761 1762
                    log_name=log_name,
                )
W
Wu Yi 已提交
1763 1764
            else:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
1765 1766
                    model_file,
                    required_envs,
1767 1768
                    update_method='nccl2',
                    check_error_log=check_error_log,
1769 1770
                    log_name=log_name,
                )
1771 1772 1773 1774 1775 1776
        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,
1777 1778
                log_name=log_name,
            )
X
xiongkun 已提交
1779 1780 1781 1782 1783 1784 1785
        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,
1786 1787
                log_name=log_name,
            )
1788 1789 1790 1791 1792 1793
        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,
1794 1795
                log_name=log_name,
            )
1796 1797 1798 1799 1800 1801
        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,
1802 1803
                log_name=log_name,
            )
1804
        elif self._pipeline_mode:
1805 1806 1807
            tr0_losses, tr1_losses = self._run_pipeline(
                model_file, required_envs, check_error_log, log_name=log_name
            )
W
Wu Yi 已提交
1808
        else:
1809 1810 1811
            tr0_losses, tr1_losses = self._run_cluster(
                model_file, required_envs, check_error_log, log_name=log_name
            )
1812 1813

        for step_id in range(RUN_STEP):
W
Wu Yi 已提交
1814 1815 1816
            local_loss = local_losses[step_id]
            tr0_loss = tr0_losses[step_id]
            tr1_loss = tr1_losses[step_id]
1817 1818 1819 1820
            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
1821 1822
            print("=======", local_loss, ":", dist_loss[0], "=======")
            self.assertAlmostEqual(local_loss, dist_loss[0], delta=delta)
1823

1824 1825 1826 1827 1828 1829 1830 1831
    def check_with_place_multi_cards(
        self,
        model_file,
        delta=1e-3,
        check_error_log=False,
        need_envs={},
        log_name="",
    ):
1832

1833 1834 1835 1836 1837 1838
        # 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:
1839 1840 1841 1842 1843 1844 1845
            multi_cards_losses = self._run_local(
                model_file,
                required_envs,
                check_error_log,
                log_name=log_name + "_dgc_2cards",
                devices="0,1",
            )
1846 1847

            self._use_dgc = False
1848 1849 1850 1851 1852 1853 1854
            base_losses = self._run_local(
                model_file,
                required_envs,
                check_error_log,
                log_name=log_name + "_base_2cards",
                devices="0,1",
            )
1855 1856 1857 1858 1859 1860 1861 1862

            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)