test_dist_base.py 26.8 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 26
import pickle
import numpy as np
27
import time
28
import paddle.fluid as fluid
29
from paddle.fluid import compiler
30 31 32
import paddle.fluid.dygraph as dygraph
from paddle.fluid.dygraph.base import to_variable
from paddle.fluid.dygraph.parallel import DataParallel
33

Y
Yan Xu 已提交
34
RUN_STEP = 5
35
DEFAULT_BATCH_SIZE = 2
36

T
typhoonzero 已提交
37

38 39 40
def my_print(class_name, log_str):
    localtime = time.asctime(time.localtime(time.time()))
    print_str = localtime + "\t" + class_name + "\t" + log_str
G
guru4elephant 已提交
41
    if six.PY2:
42
        sys.stderr.write(pickle.dumps(print_str))
G
guru4elephant 已提交
43
    else:
44
        sys.stderr.buffer.write(pickle.dumps(print_str))
G
guru4elephant 已提交
45 46


T
typhoonzero 已提交
47
class TestDistRunnerBase(object):
W
Wu Yi 已提交
48 49 50
    def get_model(self,
                  batch_size=DEFAULT_BATCH_SIZE,
                  lr=0.1,
51 52
                  single_device=False,
                  use_dgc=False):
T
typhoonzero 已提交
53 54 55
        raise NotImplementedError(
            "get_model should be implemented by child classes.")

56
    @staticmethod
W
Wu Yi 已提交
57 58 59 60 61
    def get_transpiler(trainer_id,
                       main_program,
                       pserver_endpoints,
                       trainers,
                       sync_mode,
62
                       dc_asgd=False,
63 64
                       current_endpoint=None,
                       nccl_comm_num=1):
T
typhoonzero 已提交
65
        # NOTE: import fluid until runtime, or else forking processes will cause error.
66
        config = fluid.DistributeTranspilerConfig()
W
Wu Yi 已提交
67
        config.enable_dc_asgd = dc_asgd
68
        config.sync_mode = sync_mode
69 70
        if nccl_comm_num > 1:
            config.nccl_comm_num = nccl_comm_num
71
        # config.runtime_split_send_recv = True
72
        t = fluid.DistributeTranspiler(config=config)
T
typhoonzero 已提交
73 74 75 76
        t.transpile(
            trainer_id=trainer_id,
            program=main_program,
            pservers=pserver_endpoints,
W
Wu Yi 已提交
77
            trainers=trainers,
78
            current_endpoint=current_endpoint)
T
typhoonzero 已提交
79 80
        return t

W
Wu Yi 已提交
81
    def run_pserver(self, args):
W
Wu Yi 已提交
82
        self.lr = args.lr
83
        self.get_model(batch_size=args.batch_size)
84
        # NOTE: pserver should not call memory optimize
W
Wu Yi 已提交
85 86
        t = self.get_transpiler(args.trainer_id,
                                fluid.default_main_program(), args.endpoints,
W
Wu Yi 已提交
87
                                args.trainers, args.sync_mode, args.dc_asgd)
W
Wu Yi 已提交
88 89 90
        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
91

T
typhoonzero 已提交
92 93 94
        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        exe.run(startup_prog)
95
        my_print(type(self).__name__, "run pserver startup program done.")
T
typhoonzero 已提交
96
        exe.run(pserver_prog)
97
        my_print(type(self).__name__, "run pserver main program done.")
T
typhoonzero 已提交
98

99
    def run_trainer(self, args):
W
Wu Yi 已提交
100
        self.lr = args.lr
W
Wu Yi 已提交
101 102 103
        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)
104 105 106
        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
107 108 109
        else:
            test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
                self.get_model(batch_size=args.batch_size)
110

W
Wu Yi 已提交
111
        if args.update_method == "pserver":
112 113 114
            my_print(
                type(self).__name__,
                "begin to run transpile on trainer with pserver mode")
W
Wu Yi 已提交
115 116 117
            t = self.get_transpiler(args.trainer_id,
                                    fluid.default_main_program(),
                                    args.endpoints, args.trainers,
W
Wu Yi 已提交
118
                                    args.sync_mode, args.dc_asgd)
T
typhoonzero 已提交
119
            trainer_prog = t.get_trainer_program()
120 121 122
            my_print(
                type(self).__name__,
                "get trainer program done with pserver mode.")
W
Wu Yi 已提交
123
        elif args.update_method == "nccl2" or args.update_method == "nccl2_reduce_layer":
W
Wu Yi 已提交
124 125 126
            # transpile for nccl2
            config = fluid.DistributeTranspilerConfig()
            config.mode = "nccl2"
127
            config.nccl_comm_num = args.nccl_comm_num
128 129 130
            if args.use_hallreduce:
                config.use_hierarchical_allreduce = True
                config.hierarchical_allreduce_inter_nranks = args.hallreduce_inter_nranks
131 132 133
            my_print(
                type(self).__name__,
                "begin to run transpile on trainer with nccl2 mode")
W
Wu Yi 已提交
134 135 136 137 138 139 140
            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)
141 142 143
            my_print(
                type(self).__name__,
                "get trainer program done. with nccl2 mode")
W
Wu Yi 已提交
144
            trainer_prog = fluid.default_main_program()
T
typhoonzero 已提交
145
        else:
146 147 148
            my_print(
                type(self).__name__,
                "do nothing about main program, just use it")
T
typhoonzero 已提交
149
            trainer_prog = fluid.default_main_program()
150
            my_print(type(self).__name__, "use main program done.")
T
typhoonzero 已提交
151

152
        if args.use_cuda:
153 154
            device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
            place = fluid.CUDAPlace(device_id)
155 156 157
        else:
            place = fluid.CPUPlace()

158 159
        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
160
        my_print(type(self).__name__, "run worker startup program done.")
T
typhoonzero 已提交
161

W
Wu Yi 已提交
162 163
        exec_strategy = fluid.ExecutionStrategy()
        exec_strategy.num_threads = 1
164

W
Wu Yi 已提交
165
        build_stra = fluid.BuildStrategy()
166 167 168
        # FIXME force disable enable_inplace and memory_optimize
        build_stra.enable_inplace = False
        build_stra.memory_optimize = False
W
Wu Yi 已提交
169

170 171 172
        if args.enable_backward_deps:
            build_stra.enable_backward_optimizer_op_deps = True

W
Wu Yi 已提交
173 174 175 176 177
        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
178
        pass_builder = None
X
Xin Pan 已提交
179
        if args.batch_merge_repeat > 1:
X
fix  
Xin Pan 已提交
180
            pass_builder = build_stra._finalize_strategy_and_create_passes()
181
            mypass = pass_builder.insert_pass(0, "multi_batch_merge_pass")
182
            mypass.set("num_repeats", args.batch_merge_repeat)
X
Xin Pan 已提交
183

W
Wu Yi 已提交
184
        if args.update_method == "nccl2" or args.update_method == "nccl2_reduce_layer":
185 186
            build_stra.num_trainers = len(args.endpoints.split(","))
            build_stra.trainer_id = args.trainer_id
W
Wu Yi 已提交
187
        else:
W
Wu Yi 已提交
188
            # case args.update_method == "nccl2_reduce_layer":
189 190
            build_stra.num_trainers = 1
            build_stra.trainer_id = 0
W
Wu Yi 已提交
191

192
        my_print(type(self).__name__, "begin to compile with data parallel")
X
Xin Pan 已提交
193
        binary = compiler.CompiledProgram(trainer_prog).with_data_parallel(
W
Wu Yi 已提交
194
            loss_name=avg_cost.name,
W
Wu Yi 已提交
195
            build_strategy=build_stra,
W
Wu Yi 已提交
196
            exec_strategy=exec_strategy)
197
        my_print(type(self).__name__, "program compiled with data parallel")
T
typhoonzero 已提交
198 199 200 201 202 203 204

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

        feeder = fluid.DataFeeder(feed_var_list, place)
205
        reader_generator = train_reader()
T
typhoonzero 已提交
206

207 208
        def get_data():
            origin_batch = next(reader_generator)
W
Wu Yi 已提交
209
            if args.update_method != "local" and args.use_reader_alloc:
210 211 212 213 214 215 216
                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
217

218
        my_print(type(self).__name__, "begin to train on trainer")
W
Wu Yi 已提交
219
        out_losses = []
220
        for i in six.moves.xrange(RUN_STEP):
221 222
            loss, = exe.run(binary,
                            fetch_list=[avg_cost.name],
223
                            feed=feeder.feed(get_data()))
W
Wu Yi 已提交
224
            out_losses.append(loss[0])
225 226 227
            my_print(type(self).__name__, "run step %d finished" % i)
        my_print(type(self).__name__, "trainer run finished")

W
Wu Yi 已提交
228 229 230 231
        if six.PY2:
            print(pickle.dumps(out_losses))
        else:
            sys.stdout.buffer.write(pickle.dumps(out_losses))
T
typhoonzero 已提交
232 233


234 235 236 237 238 239 240 241 242 243
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.")

    def run_trainer(self, args):
Y
Yan Xu 已提交
244

245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261
        seed = 90
        device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
        place = fluid.CUDAPlace(device_id)

        def _get_data(batch):
            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

        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
262 263 264
            np.random.seed(seed)
            import random
            random.seed = seed
265 266
            model, train_reader, opt = self.get_model()
            nranks = len(args.endpoints.split(",")) if args.endpoints else 1
Y
Yan Xu 已提交
267

268 269 270 271 272 273
            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
274 275 276
                my_print(
                    type(self).__name__,
                    "begin to prepare context in dygraph with nccl2")
277
                dygraph.parallel.prepare_context(strategy)
Y
Yan Xu 已提交
278
                model = dygraph.parallel.DataParallel(model, strategy)
279
                my_print(type(self).__name__, "model built in dygraph")
280
            out_losses = []
281
            my_print(type(self).__name__, "begin to run dygraph training")
282 283 284 285 286
            for step_id, data in enumerate(train_reader()):
                data = _get_data(data)
                if step_id == RUN_STEP:
                    break
                loss = self.run_one_loop(model, opt, data)
G
guru4elephant 已提交
287
                if step_id % 10 == 0:
288 289 290
                    my_print(
                        type(self).__name__,
                        "loss at step %d: %f" % (step_id, loss))
Y
Yan Xu 已提交
291
                out_losses.append(loss.numpy())
292

Y
Yan Xu 已提交
293 294 295
                # FIXME(Yancey1989): scale the loss inplace
                if args.update_method == "nccl2":
                    loss = model.scale_loss(loss)
296 297

                loss.backward()
Y
Yan Xu 已提交
298 299
                if args.update_method == "nccl2":
                    model.apply_collective_grads()
300 301 302

                opt.minimize(loss)
                model.clear_gradients()
303
            my_print(type(self).__name__, pickle.dumps(out_losses))
304 305


T
typhoonzero 已提交
306
def runtime_main(test_class):
W
Wu Yi 已提交
307 308 309 310
    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
311 312 313 314
    parser.add_argument(
        '--update_method',
        type=str,
        default="local",
W
Wu Yi 已提交
315
        choices=["pserver", "nccl2", "local", "nccl2_reduce_layer"])
W
Wu Yi 已提交
316 317
    parser.add_argument('--trainer_id', type=int, required=False, default=0)
    parser.add_argument('--trainers', type=int, required=False, default=1)
318
    parser.add_argument('--nccl_comm_num', type=int, required=False, default=1)
319 320
    parser.add_argument('--enable_backward_deps', action='store_true')
    parser.add_argument('--use_hallreduce', action='store_true')
321
    parser.add_argument(
322
        '--hallreduce_inter_nranks', type=int, required=False, default=2)
W
Wu Yi 已提交
323 324 325
    parser.add_argument(
        '--current_endpoint', type=str, required=False, default="")
    parser.add_argument('--sync_mode', action='store_true')
326
    parser.add_argument('--use_cuda', action='store_true')
327
    parser.add_argument('--use_dgc', action='store_true')
W
Wu Yi 已提交
328
    parser.add_argument('--use_reduce', action='store_true')
W
Wu Yi 已提交
329
    parser.add_argument('--dc_asgd', action='store_true')
330
    parser.add_argument(
W
Wu Yi 已提交
331
        '--use_reader_alloc', action='store_true', required=False)
332
    parser.add_argument('--batch_size', required=False, type=int, default=2)
W
Wu Yi 已提交
333
    parser.add_argument('--lr', required=False, type=float, default=0.001)
334 335
    parser.add_argument(
        '--batch_merge_repeat', required=False, type=int, default=1)
W
Wu Yi 已提交
336 337 338 339 340
    parser.add_argument(
        '--nccl2_reduce_layer_local_run',
        required=False,
        type=bool,
        default=False)
W
Wu Yi 已提交
341 342

    args = parser.parse_args()
T
typhoonzero 已提交
343 344

    model = test_class()
W
Wu Yi 已提交
345
    if args.role == "pserver" and args.update_method == "pserver":
W
Wu Yi 已提交
346
        model.run_pserver(args)
T
typhoonzero 已提交
347
    else:
348
        model.run_trainer(args)
X
Xin Pan 已提交
349

M
minqiyang 已提交
350

M
minqiyang 已提交
351
import paddle.compat as cpt
Y
Yancey1989 已提交
352 353
import socket
from contextlib import closing
M
minqiyang 已提交
354

X
Xin Pan 已提交
355 356

class TestDistBase(unittest.TestCase):
W
Wu Yi 已提交
357 358 359
    def _setup_config(self):
        raise NotImplementedError("tests should have _setup_config implemented")

360 361 362
    def _after_setup_config(self):
        if self._enforce_place == "CPU":
            self.__use_cuda = False
363
            self._use_dgc = False
364 365 366 367 368 369 370
        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
371 372 373 374
                self._use_dgc = False

        if self._use_reduce:
            assert not self._use_dgc
375

X
Xin Pan 已提交
376 377 378
    def setUp(self):
        self._trainers = 2
        self._pservers = 2
Y
Yancey1989 已提交
379
        self._port_set = set()
Y
Yancey1989 已提交
380 381
        self._ps_endpoints = "127.0.0.1:%s,127.0.0.1:%s" % (
            self._find_free_port(), self._find_free_port())
M
minqiyang 已提交
382
        self._python_interp = sys.executable
W
Wu Yi 已提交
383
        self._sync_mode = True
384
        self._enforce_place = None
W
Wu Yi 已提交
385
        self._use_reduce = False
W
Wu Yi 已提交
386
        self._dc_asgd = False  # must use with async mode
387
        self._use_reader_alloc = True
W
Wu Yi 已提交
388
        self._nccl2_mode = False
389
        self._mp_mode = False
W
Wu Yi 已提交
390 391 392 393 394
        # 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
395
        self._lr = 0.001
396
        self._use_dgc = False
397
        self._dygraph = False
398
        self._nccl_comm_num = 1
399 400
        self._enable_backward_deps = False
        self._use_hallreduce = False
W
Wu Yi 已提交
401
        self._setup_config()
402
        self._after_setup_config()
X
Xin Pan 已提交
403

Y
Yancey1989 已提交
404
    def _find_free_port(self):
Y
Yancey1989 已提交
405 406 407 408
        def __free_port():
            with closing(socket.socket(socket.AF_INET,
                                       socket.SOCK_STREAM)) as s:
                s.bind(('', 0))
409 410
                my_print(
                    type(self).__name__, "socket name: %s" % s.getsockname()[1])
Y
Yancey1989 已提交
411 412 413 414 415 416 417
                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
418

419
    def start_pserver(self, model_file, check_error_log, required_envs):
X
Xin Pan 已提交
420
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
W
Wu Yi 已提交
421
        ps_cmd = "%s %s --role pserver --endpoints %s --trainer_id 0 --current_endpoint %s --trainers %d --update_method pserver"
W
Wu Yi 已提交
422
        ps0_cmd = ps_cmd % \
423 424
                  (self._python_interp, model_file, self._ps_endpoints, ps0_ep,
                   self._trainers)
W
Wu Yi 已提交
425
        ps1_cmd = ps_cmd % \
426 427
                  (self._python_interp, model_file, self._ps_endpoints, ps1_ep,
                   self._trainers)
W
Wu Yi 已提交
428 429 430 431

        if self._sync_mode:
            ps0_cmd += " --sync_mode"
            ps1_cmd += " --sync_mode"
X
Xin Pan 已提交
432

433 434
        print(ps0_cmd)
        print(ps1_cmd)
M
minqiyang 已提交
435 436
        ps0_pipe = open("/tmp/ps0_err.log", "wb")
        ps1_pipe = open("/tmp/ps1_err.log", "wb")
G
gongweibao 已提交
437

438
        my_print(type(self).__name__, "going to start pserver process 0")
X
Xin Pan 已提交
439
        ps0_proc = subprocess.Popen(
440 441 442 443
            ps0_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps0_pipe,
            env=required_envs)
444
        my_print(type(self).__name__, "going to start pserver process 1")
X
Xin Pan 已提交
445
        ps1_proc = subprocess.Popen(
446 447 448 449
            ps1_cmd.strip().split(" "),
            stdout=subprocess.PIPE,
            stderr=ps1_pipe,
            env=required_envs)
G
gongweibao 已提交
450

451
        return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe
X
Xin Pan 已提交
452

453 454 455 456 457 458
    def _run_local(self,
                   model,
                   envs,
                   check_error_log=False,
                   batch_size=DEFAULT_BATCH_SIZE,
                   batch_merge_repeat=1):
G
gongweibao 已提交
459

W
Wu Yi 已提交
460 461
        cmd = "%s %s --role trainer --lr %f" % (self._python_interp, model,
                                                self._lr)
462 463 464 465
        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
466 467
        if self._nccl2_reduce_layer:
            cmd += " --nccl2_reduce_layer_local_run 1"
468

469
        if self.__use_cuda:
470
            cmd += " --use_cuda"
W
Wu Yi 已提交
471 472 473 474 475
            env_local = {
                "CUDA_VISIBLE_DEVICES": "0",
                "PADDLE_TRAINERS_NUM": "1",
                "PADDLE_TRAINER_ID": "0"
            }
476 477 478
        else:
            env_local = {'CPU_NUM': '1'}

W
Wu Yi 已提交
479 480
        env_local.update(envs)
        print("local_cmd: {}, env: {}".format(cmd, env_local))
G
gongweibao 已提交
481

482
        if check_error_log:
483
            err_log = open("/tmp/trainer.err.log", "wb")
G
gongweibao 已提交
484
            local_proc = subprocess.Popen(
485
                cmd.split(" "),
G
gongweibao 已提交
486
                stdout=subprocess.PIPE,
487
                stderr=err_log,
W
Wu Yi 已提交
488
                env=env_local)
G
gongweibao 已提交
489 490
        else:
            local_proc = subprocess.Popen(
491
                cmd.split(" "),
G
gongweibao 已提交
492
                stdout=subprocess.PIPE,
493
                stderr=subprocess.PIPE,
W
Wu Yi 已提交
494
                env=env_local)
G
gongweibao 已提交
495

496 497 498 499 500 501
        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
502
        sys.stderr.write('local_stdout: %s\n' % pickle.loads(local_out))
X
Xin Pan 已提交
503

W
Wu Yi 已提交
504
        return pickle.loads(local_out)
505 506

    def _run_cluster(self, model, envs, check_error_log):
X
Xin Pan 已提交
507
        # Run dist train to compare with local results
508 509
        ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(model,
                                                          check_error_log, envs)
W
Wu Yi 已提交
510

X
Xin Pan 已提交
511
        ps0_ep, ps1_ep = self._ps_endpoints.split(",")
512

W
Wu Yi 已提交
513
        tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --update_method pserver --lr %f"
W
Wu Yi 已提交
514
        tr0_cmd = tr_cmd % \
515
                  (self._python_interp, model, self._ps_endpoints,
W
Wu Yi 已提交
516
                   0, ps0_ep, self._trainers, self._lr)
W
Wu Yi 已提交
517
        tr1_cmd = tr_cmd % \
518
                  (self._python_interp, model, self._ps_endpoints,
W
Wu Yi 已提交
519
                   1, ps1_ep, self._trainers, self._lr)
W
Wu Yi 已提交
520 521 522 523 524 525 526

        if self._sync_mode:
            tr0_cmd += " --sync_mode"
            tr1_cmd += " --sync_mode"
        if self._use_reduce:
            tr0_cmd += " --use_reduce"
            tr1_cmd += " --use_reduce"
527 528 529
        if self._use_reader_alloc:
            tr0_cmd += " --use_reader_alloc"
            tr1_cmd += " --use_reader_alloc"
530
        if self.__use_cuda:
531 532 533 534 535 536 537 538 539 540
            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
541

W
Wu Yi 已提交
542 543
        print("tr0_cmd: {}, env: {}".format(tr0_cmd, env0))
        print("tr1_cmd: {}, env: {}".format(tr1_cmd, env1))
544 545
        tr0_pipe = open("/tmp/tr0_err.log", "wb")
        tr1_pipe = open("/tmp/tr1_err.log", "wb")
G
gongweibao 已提交
546

547
        my_print(type(self).__name__, "going to start trainer process 0")
X
Xin Pan 已提交
548
        tr0_proc = subprocess.Popen(
W
Wu Yi 已提交
549
            tr0_cmd.strip().split(" "),
X
Xin Pan 已提交
550
            stdout=subprocess.PIPE,
G
gongweibao 已提交
551
            stderr=tr0_pipe,
X
Xin Pan 已提交
552
            env=env0)
553
        my_print(type(self).__name__, "going to start trainer process 1")
X
Xin Pan 已提交
554
        tr1_proc = subprocess.Popen(
W
Wu Yi 已提交
555
            tr1_cmd.strip().split(" "),
X
Xin Pan 已提交
556
            stdout=subprocess.PIPE,
G
gongweibao 已提交
557
            stderr=tr1_pipe,
X
Xin Pan 已提交
558 559
            env=env1)

560 561 562 563 564 565 566 567 568 569 570 571
        # 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

572 573
        tr0_out, tr0_err = tr0_proc.communicate()
        tr1_out, tr1_err = tr1_proc.communicate()
X
Xin Pan 已提交
574

G
gongweibao 已提交
575
        # close trainer file
576 577 578 579
        tr0_pipe.close()
        tr1_pipe.close()
        ps0_pipe.close()
        ps1_pipe.close()
W
Wu Yi 已提交
580

W
Wu Yi 已提交
581 582
        ps0.terminate()
        ps1.terminate()
T
typhoonzero 已提交
583

W
Wu Yi 已提交
584 585
        return pickle.loads(tr0_out), pickle.loads(tr1_out)

586 587 588 589 590
    def _get_nccl2_trainer_cmd(self, model, ep, update_method, trainer_id,
                               trainer_num):
        env = {}
        tr_cmd = "%s -u %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --update_method %s --lr %f"
        tr_cmd = tr_cmd % \
W
Wu Yi 已提交
591
                  (self._python_interp, model, self._ps_endpoints,
592
                   trainer_id, ep, update_method, self._lr)
W
Wu Yi 已提交
593 594

        if self._use_reduce:
595
            tr_cmd += " --use_reduce"
W
Wu Yi 已提交
596
        if self._use_reader_alloc:
597
            tr_cmd += " --use_reader_alloc"
W
Wu Yi 已提交
598
        if self.__use_cuda:
599 600 601 602 603 604
            tr_cmd += " --use_cuda"
            env.update({
                "CUDA_VISIBLE_DEVICES": "{}".format(trainer_id),
                "PADDLE_TRAINERS_NUM": "{}".format(trainer_num),
                "PADDLE_TRAINER_ID": "{}".format(trainer_id)
            })
W
Wu Yi 已提交
605
        else:
606
            env.update({'CPU_NUM': '1'})
W
Wu Yi 已提交
607

608
        if self._use_dgc:
609 610 611 612
            tr_cmd += " --use_dgc"

        if self._mp_mode:
            env = {"FLAGS_selected_gpus": "{}".format(trainer_id)}
613 614

        if self._nccl_comm_num > 1:
615
            tr_cmd += " --nccl_comm_num {}".format(self._nccl_comm_num)
616

617 618
        if self._use_hallreduce:
            tr_cmd += " --use_hallreduce --hallreduce_inter_nranks 2"
619

620
        if self._enable_backward_deps:
621
            tr_cmd += " --enable_backward_deps"
622

623
        return tr_cmd, env
W
Wu Yi 已提交
624

625 626 627 628 629 630 631
    def _run_cluster_nccl2(self, model, envs, nccl2_reduce_layer,
                           check_error_log):
        if self._use_hallreduce:
            self._ps_endpoints = ""
            for i in range(0, 4):
                self._ps_endpoints += "127.0.0.1:%s," % (self._find_free_port())
            self._ps_endpoints = self._ps_endpoints[:-1]
W
Wu Yi 已提交
632

633 634 635 636 637 638
        # 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
639

640
        trainer_num = len(worker_endpoints)
W
Wu Yi 已提交
641

642 643 644 645 646 647 648 649
        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
650

651
            tr_pipe = open("/tmp/tr{}_err.log".format(i), "wb")
W
Wu Yi 已提交
652

653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
            my_print(
                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))

        return pickle.loads(outs[0]), pickle.loads(outs[1])
673 674 675 676 677 678 679 680 681 682 683 684

    def check_with_place(self,
                         model_file,
                         delta=1e-3,
                         check_error_log=False,
                         need_envs={}):
        # 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
685
            "FLAGS_rpc_deadline": "30000",  # 5sec to fail fast
686
            "FLAGS_cudnn_deterministic": "1",
W
Wu Yi 已提交
687
            "http_proxy": "",
688 689
            "NCCL_P2P_DISABLE": "1",
            "NCCL_SHM_DISABLE": "1"
690 691 692 693 694
        }

        required_envs.update(need_envs)

        if check_error_log:
695
            required_envs["GLOG_v"] = "10"
696 697 698 699
            required_envs["GLOG_logtostderr"] = "1"

        local_losses\
            = self._run_local(model_file, required_envs,
W
Wu Yi 已提交
700
                                check_error_log)
W
Wu Yi 已提交
701
        if self._nccl2_mode:
W
Wu Yi 已提交
702 703 704 705 706 707
            if self._nccl2_reduce_layer:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
                    model_file, required_envs, True, check_error_log)
            else:
                tr0_losses, tr1_losses = self._run_cluster_nccl2(
                    model_file, required_envs, False, check_error_log)
W
Wu Yi 已提交
708 709 710
        else:
            tr0_losses, tr1_losses = self._run_cluster(
                model_file, required_envs, check_error_log)
711 712

        for step_id in range(RUN_STEP):
W
Wu Yi 已提交
713 714 715
            local_loss = local_losses[step_id]
            tr0_loss = tr0_losses[step_id]
            tr1_loss = tr1_losses[step_id]
Y
Yan Xu 已提交
716
            dist_loss = (np.array([tr0_loss]) + np.array([tr1_loss])) / 2
W
Wu Yi 已提交
717 718
            print("=======", local_loss, ":", dist_loss[0], "=======")
            self.assertAlmostEqual(local_loss, dist_loss[0], delta=delta)