distribute_transpiler.py 74.6 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13
# 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
16 17 18 19 20
"""
Steps to transpile trainer:
1. split variable to multiple blocks, aligned by product(dim[1:]) (width).
2. rename splited grad variables to add trainer_id suffix ".trainer_%d".
3. modify trainer program add split_op to each grad variable.
Q
Qiyang Min 已提交
21
4. append send_op to send splited variables to server and
22 23
5. add recv_op to fetch params(splited blocks or origin param) from server.
6. append concat_op to merge splited blocks to update local weights.
24 25 26 27 28 29 30 31

Steps to transpile pserver:
1. create new program for parameter server.
2. create params and grad variables that assigned to current server instance.
3. create a sub-block in the server side program
4. append ops that should run on current server instance.
5. add listen_and_serv op
"""
D
dzhwinter 已提交
32

T
typhoonzero 已提交
33
import math
W
Wu Yi 已提交
34
import sys
35
import numpy as np
36
import collections
37
import six
38

39
from .ps_dispatcher import RoundRobin, HashName, PSDispatcher
Y
Yancey 已提交
40
from .. import core, framework
T
typhoonzero 已提交
41
from ..framework import Program, default_main_program, \
T
tangwei12 已提交
42 43
    default_startup_program, Block, \
    Parameter, grad_var_name
44 45
from .details import *
from functools import reduce
46 47 48

LOOKUP_TABLE_TYPE = "lookup_table"
LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
49
OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
Y
Yancey1989 已提交
50 51
RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName(
)
X
fix  
Xin Pan 已提交
52
OPT_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Optimize
Y
Yancey1989 已提交
53
RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
54 55 56 57 58 59 60 61 62
DIST_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Dist
LR_SCHED_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.LRSched

PRINT_LOG = False


def log(*args):
    if PRINT_LOG:
        print(args)
T
done  
typhoonzero 已提交
63 64


T
typhoonzero 已提交
65 66 67 68 69 70
class VarBlock:
    def __init__(self, varname, offset, size):
        self.varname = varname
        # NOTE: real offset is offset * size
        self.offset = offset
        self.size = size
T
done  
typhoonzero 已提交
71

T
typhoonzero 已提交
72 73
    def __str__(self):
        return "%s:%d:%d" % (self.varname, self.offset, self.size)
T
done  
typhoonzero 已提交
74 75


76 77 78 79
def same_or_split_var(p_name, var_name):
    return p_name == var_name or p_name.startswith(var_name + ".block")


G
gongweibao 已提交
80
def slice_variable(var_list, slice_count, min_block_size):
T
typhoonzero 已提交
81
    """
82 83 84 85 86 87
    We may need to split dense tensor to one or more blocks and put
    them equally onto parameter server. One block is a sub-tensor
    aligned by dim[0] of the tensor.

    We need to have a minimal block size so that the calculations in
    the parameter server side can gain better performance. By default
88
    minimum block size 8K elements (maybe 16bit or 32bit or 64bit).
89 90 91

    Args:
        var_list (list): List of variables.
92 93
        slice_count (int): Numel of count that variables will be sliced, which
            could be the pserver services' count.
94 95
        min_block_size (int): Minimum splitted block size.
    Returns:
96
        blocks (list[(varname, block_id, current_block_size)]): A list
97
            of VarBlocks. Each VarBlock specifies a shard of the var.
T
typhoonzero 已提交
98 99 100
    """
    blocks = []
    for var in var_list:
101
        split_count = slice_count
T
typhoonzero 已提交
102 103 104 105
        var_numel = reduce(lambda x, y: x * y, var.shape)
        max_pserver_count = int(math.floor(var_numel / float(min_block_size)))
        if max_pserver_count == 0:
            max_pserver_count = 1
106
        if max_pserver_count < slice_count:
T
typhoonzero 已提交
107 108 109 110 111 112 113 114 115
            split_count = max_pserver_count
        block_size = int(math.ceil(var_numel / float(split_count)))

        if len(var.shape) >= 2:
            # align by dim1(width)
            dim1 = reduce(lambda x, y: x * y, var.shape[1:])
            remains = block_size % dim1
            if remains != 0:
                block_size += dim1 - remains
116
        # update split_count after aligning
T
typhoonzero 已提交
117
        split_count = int(math.ceil(var_numel / float(block_size)))
118
        for block_id in range(split_count):
T
typhoonzero 已提交
119 120 121 122 123 124 125
            curr_block_size = min(block_size, var_numel - (
                (block_id) * block_size))
            block = VarBlock(var.name, block_id, curr_block_size)
            blocks.append(str(block))
    return blocks


G
gongweibao 已提交
126 127 128 129 130 131 132
class DistributeTranspilerConfig(object):
    """
    slice_var_up (bool): Do Tensor slice for pservers, default is True.
    split_method (PSDispatcher): RoundRobin or HashName can be used
        try to choose the best method to balance loads for pservers.
    min_block_size (int): Minimum splitted element number in block.
        According:https://github.com/PaddlePaddle/Paddle/issues/8638#issuecomment-369912156
133
        We can use bandwidth effiently when data size is larger than 2MB.If you
G
gongweibao 已提交
134 135 136 137 138 139
        want to change it, please be sure you see the slice_variable function.
    """

    slice_var_up = True
    split_method = None
    min_block_size = 8192
W
Wu Yi 已提交
140 141
    # supported modes: pserver, nccl2
    mode = "pserver"
142
    print_log = False
G
gongweibao 已提交
143 144


Y
gen rst  
yi.wu 已提交
145
class DistributeTranspiler(object):
Y
yi.wu 已提交
146 147 148 149
    """
    **DistributeTranspiler**

    Convert the fluid program to distributed data-parallelism programs.
W
Wu Yi 已提交
150
    Supports two modes: pserver mode and nccl2 mode.
Y
yi.wu 已提交
151

W
Wu Yi 已提交
152 153 154 155 156 157 158 159 160
    In pserver mode, the main_program will be transformed to use a remote
    parameter server to do parameter optimization. And the optimization
    graph will be put into a parameter server program.

    In nccl2 mode, the transpiler will append a NCCL_ID broadcasting
    op in startup_program to share the NCCL_ID across the job nodes.
    After transpile_nccl2 called, you ***must*** pass trainer_id and
    num_trainers argument to ParallelExecutor to enable NCCL2 distributed
    mode.
Y
yi.wu 已提交
161 162 163 164

    Examples:
        .. code-block:: python

W
Wu Yi 已提交
165 166 167 168 169 170
           # for pserver mode
           pserver_endpoints = "192.168.0.1:6174,192.168.0.2:6174"
           trainer_endpoints = "192.168.0.1:6174,192.168.0.2:6174"
           current_endpoint = "192.168.0.1:6174"
           trainer_id = 0
           trainers = 4
Y
yi.wu 已提交
171 172
           role = os.getenv("PADDLE_TRAINING_ROLE")

W
Wu Yi 已提交
173
           t = fluid.DistributeTranspiler()
Y
yi.wu 已提交
174 175 176 177 178 179 180 181
           t.transpile(
                trainer_id, pservers=pserver_endpoints, trainers=trainers)
           if role == "PSERVER":
                pserver_program = t.get_pserver_program(current_endpoint)
                pserver_startup_program = t.get_startup_program(current_endpoint,
                                                                pserver_program)
           elif role == "TRAINER":
                trainer_program = t.get_trainer_program()
T
tangwei12 已提交
182

W
Wu Yi 已提交
183 184 185 186 187 188 189 190 191 192 193
           # for nccl2 mode
           config = fluid.DistributeTranspilerConfig()
           config.mode = "nccl2"
           t = fluid.DistributeTranspiler(config=config)
           t.transpile(trainer_id, workers=workers, current_endpoint=curr_ep)
           exe = fluid.ParallelExecutor(
               use_cuda,
               loss_name=loss_var.name,
               num_trainers=len(trainers.split(",)),
               trainer_id=trainer_id
           )
Y
yi.wu 已提交
194
    """
Y
Yancey1989 已提交
195

G
gongweibao 已提交
196 197 198 199 200 201 202 203 204
    def __init__(self, config=None):
        if config is not None:
            self.config = config
        else:
            self.config = DistributeTranspilerConfig()

        if self.config.split_method is None:
            self.config.split_method = RoundRobin

205 206 207
        global PRINT_LOG
        if self.config.print_log:
            PRINT_LOG = True
G
gongweibao 已提交
208 209 210
        assert (self.config.min_block_size >= 8192)
        assert (self.config.split_method.__bases__[0] == PSDispatcher)

W
Wu Yi 已提交
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
    def _transpile_nccl2(self,
                         trainer_id,
                         trainers,
                         current_endpoint,
                         startup_program=None):
        if not startup_program:
            startup_program = default_startup_program()
        if trainer_id >= 0:
            worker_endpoints = trainers.split(",")
            # send NCCL_ID to others or recv from trainer 0
            worker_endpoints.remove(current_endpoint)

            nccl_id_var = startup_program.global_block().create_var(
                name="NCCLID", persistable=True, type=core.VarDesc.VarType.RAW)
            startup_program.global_block().append_op(
                type="gen_nccl_id",
                inputs={},
                outputs={"NCCLID": nccl_id_var},
                attrs={
                    "endpoint": current_endpoint,
                    "endpoint_list": worker_endpoints,
                    "trainer_id": trainer_id
                })
            return nccl_id_var
        else:
            raise ValueError("must set trainer_id > 0")

238 239 240 241 242
    def transpile(self,
                  trainer_id,
                  program=None,
                  pservers="127.0.0.1:6174",
                  trainers=1,
W
Wu Yi 已提交
243
                  sync_mode=True,
W
Wu Yi 已提交
244 245
                  startup_program=None,
                  current_endpoint="127.0.0.1:6174"):
246
        """
Y
yi.wu 已提交
247 248 249 250 251 252 253 254 255
        Run the transpiler.

        Args:
            trainer_id (int): id for current trainer worker, if you have
                n workers, the id may range from 0 ~ n-1
            program (Program|None): program to transpile,
                default is fluid.default_main_program().
            pservers (str): comma separated ip:port string for the pserver
                list.
W
Wu Yi 已提交
256 257 258
            trainers (int|str): in pserver mode this is the number of
                trainers, in nccl2 mode this is a string of trainer
                endpoints.
Y
yi.wu 已提交
259
            sync_mode (bool): Do sync training or not, default is True.
W
Wu Yi 已提交
260 261
            startup_program (Program|None): startup_program to transpile,
                default is fluid.default_main_program().
W
Wu Yi 已提交
262 263 264
            current_endpoint (str): need pass current endpoint when
                transpile as nccl2 distributed mode. In pserver mode
                this argument is not used.
265 266 267
        """
        if program is None:
            program = default_main_program()
W
Wu Yi 已提交
268 269
        if startup_program is None:
            startup_program = default_startup_program()
270
        self.origin_program = program
W
Wu Yi 已提交
271 272
        self.startup_program = startup_program
        self.origin_startup_program = self.startup_program.clone()
G
gongweibao 已提交
273

W
Wu Yi 已提交
274 275 276 277 278 279 280 281 282
        if self.config.mode == "nccl2":
            assert (isinstance(trainers, str))
            self._transpile_nccl2(
                trainer_id,
                trainers,
                current_endpoint,
                startup_program=startup_program)
            return

283 284 285 286 287 288 289
        self.trainer_num = trainers
        self.sync_mode = sync_mode
        self.trainer_id = trainer_id
        pserver_endpoints = pservers.split(",")
        self.pserver_endpoints = pserver_endpoints
        self.optimize_ops, self.params_grads = self._get_optimize_pass()

G
gongweibao 已提交
290
        ps_dispatcher = self.config.split_method(self.pserver_endpoints)
291
        self.has_distributed_lookup_table = self._has_distributed_lookup_table()
292
        self.param_name_to_grad_name = dict()
W
Wu Yi 已提交
293
        self.grad_name_to_param_name = dict()
294 295
        for param_var, grad_var in self.params_grads:
            self.param_name_to_grad_name[param_var.name] = grad_var.name
W
Wu Yi 已提交
296
            self.grad_name_to_param_name[grad_var.name] = param_var.name
297

T
tangwei12 已提交
298 299 300 301 302 303
        # add distributed attrs to program
        self.origin_program._is_distributed = True
        self.origin_program._endpoints = self.pserver_endpoints
        self.origin_program._is_chief = self.trainer_id == 0
        self.origin_program._distributed_lookup_table = self.table_name if self.table_name else None

304
        # split and create vars, then put splited vars in dicts for later use.
G
gongweibao 已提交
305
        # step 1: split and create vars, then put splited vars in dicts for later use.
G
gongweibao 已提交
306
        self._init_splited_vars()
307

G
gongweibao 已提交
308
        # step 2: insert send op to send gradient vars to parameter servers
Y
Yancey1989 已提交
309
        ps_dispatcher.reset()
Y
update  
Yancey1989 已提交
310
        send_vars = []
311 312 313 314 315 316

        # in general cases, the number of pservers is times of 2, and this
        # will lead to uneven distribution among weights and bias:
        #       fc_w@GRAD_trainer_0, fc_w@GRAD_trainer_1 --> pserver1
        #       fc_b@GRAD_trainer_0, fc_b@GRAD_trainer_1 --> pserver2
        # shuffle the map will avoid the uneven distribution above
M
minqiyang 已提交
317
        grad_var_mapping_items = list(six.iteritems(self.grad_var_mapping))
318

G
gongweibao 已提交
319
        if not self.config.slice_var_up:
320 321
            np.random.seed(self.origin_program.random_seed)
            np.random.shuffle(grad_var_mapping_items)
322

323
        self.grad_name_to_send_dummy_out = dict()
324
        for grad_varname, splited_vars in grad_var_mapping_items:
Y
update  
Yancey1989 已提交
325
            eplist = ps_dispatcher.dispatch(splited_vars)
326

G
gongweibao 已提交
327
            if not self.config.slice_var_up:
328 329
                assert (len(splited_vars) == 1)

330
            splited_grad_varname = grad_varname
Y
Yancey1989 已提交
331
            if len(splited_vars) == 1:
332
                splited_grad_varname = splited_vars[0].name
333 334
                index = find_op_by_output_arg(
                    program.global_block(), splited_grad_varname, reverse=True)
Y
Yancey1989 已提交
335
            elif len(splited_vars) > 1:
336
                orig_var = program.global_block().vars[splited_grad_varname]
337 338
                index = find_op_by_output_arg(
                    program.global_block(), splited_grad_varname, reverse=True)
Y
Yancey1989 已提交
339
                self._insert_split_op(program, orig_var, index, splited_vars)
Y
update  
Yancey1989 已提交
340
                index += 1
Y
Yancey1989 已提交
341 342
            else:
                AssertionError("Can not insert the send op by original "
343
                               "variable name :", splited_grad_varname)
Y
Yancey1989 已提交
344

W
Wu Yi 已提交
345 346
            dummy_output = program.global_block().create_var(
                name=framework.generate_control_dev_var_name())
347
            self.grad_name_to_send_dummy_out[grad_varname] = dummy_output
W
Wu Yi 已提交
348

W
Wu Yi 已提交
349 350 351 352
            # get send op_role_var, if not splited, the grad should have .trainer suffix
            # if splited, grad should be the original grad var name (split_by_ref and send
            # will be on the same place). ParallelExecutor
            # will use op_role_var to get expected device place to run this op.
W
Wu Yi 已提交
353
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
354
                index=index + 1,
355
                type="send",
Y
update  
Yancey1989 已提交
356
                inputs={"X": splited_vars},
357
                outputs={"Out": dummy_output},
Y
Yancey1989 已提交
358 359
                attrs={
                    "epmap": eplist,
360
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
W
Wu Yi 已提交
361 362 363 364
                    OP_ROLE_VAR_ATTR_NAME: [
                        self.grad_name_to_param_name[grad_varname],
                        splited_grad_varname
                    ],
365
                    "sync_mode": not self.sync_mode,
Y
Yancey1989 已提交
366
                })
Y
update  
Yancey1989 已提交
367 368
            for _, var in enumerate(splited_vars):
                send_vars.append(var)
Y
Yancey1989 已提交
369 370

        if self.sync_mode:
W
Wu Yi 已提交
371 372
            send_barrier_out = program.global_block().create_var(
                name=framework.generate_control_dev_var_name())
373 374 375 376
            if self.has_distributed_lookup_table:
                self.grad_name_to_send_dummy_out[
                    self.table_name] = program.global_block().create_var(
                        name=framework.generate_control_dev_var_name())
377
            input_deps = list(self.grad_name_to_send_dummy_out.values())
378

Y
Yancey1989 已提交
379 380
            program.global_block().append_op(
                type="send_barrier",
M
minqiyang 已提交
381
                inputs={"X": list(input_deps)},
W
Wu Yi 已提交
382
                outputs={"Out": send_barrier_out},
Y
Yancey1989 已提交
383 384
                attrs={
                    "endpoints": pserver_endpoints,
Y
Yancey1989 已提交
385
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Y
Yancey1989 已提交
386
                })
Y
Yancey1989 已提交
387

G
gongweibao 已提交
388
        # step 3: insert recv op to receive parameters from parameter server
Y
Yancey1989 已提交
389
        recv_vars = []
Y
update  
Yancey1989 已提交
390
        for _, var in enumerate(send_vars):
391
            recv_vars.append(self.grad_param_mapping[var])
Y
update  
Yancey1989 已提交
392
        ps_dispatcher.reset()
Y
Yancey1989 已提交
393 394
        eplist = ps_dispatcher.dispatch(recv_vars)

T
typhoonzero 已提交
395
        for i, ep in enumerate(eplist):
Y
Yancey1989 已提交
396 397
            self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
            self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
398

Y
Yancey1989 已提交
399
        # step4: Concat the parameters splits together after recv.
W
Wu Yi 已提交
400
        all_recv_outputs = []
401
        for param_varname, splited_var in six.iteritems(self.param_var_mapping):
Y
Yancey1989 已提交
402 403 404 405
            eps = []
            for var in splited_var:
                index = [v.name for v in recv_vars].index(var.name)
                eps.append(eplist[index])
W
Wu Yi 已提交
406 407 408 409
            if self.sync_mode:
                recv_dep_in = send_barrier_out
            else:
                # connect deps to send op in async mode
410
                recv_dep_in = self.grad_name_to_send_dummy_out[
W
Wu Yi 已提交
411 412
                    self.param_name_to_grad_name[param_varname]]
            all_recv_outputs.extend(splited_var)
W
Wu Yi 已提交
413 414 415 416 417 418 419 420 421
            # get recv op_role_var, if not splited, the grad should have .trainer suffix
            # if splited, grad should be the original grad var name. ParallelExecutor
            # will use op_role_var to get expected device place to run this op.
            orig_grad_name = self.param_name_to_grad_name[param_varname]
            recv_op_role_var_name = orig_grad_name
            splited_trainer_grad = self.grad_var_mapping[orig_grad_name]
            if len(splited_trainer_grad) == 1:
                recv_op_role_var_name = splited_trainer_grad[0].name

Y
Yancey1989 已提交
422 423
            program.global_block().append_op(
                type="recv",
W
Wu Yi 已提交
424
                inputs={"X": [recv_dep_in]},
Y
Yancey1989 已提交
425 426 427
                outputs={"Out": splited_var},
                attrs={
                    "epmap": eps,
428
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
W
Wu Yi 已提交
429 430
                    OP_ROLE_VAR_ATTR_NAME:
                    [param_varname, recv_op_role_var_name],
431
                    "sync_mode": not self.sync_mode
Y
Yancey1989 已提交
432
                })
T
typhoonzero 已提交
433

Q
qiaolongfei 已提交
434
        if self.sync_mode:
W
Wu Yi 已提交
435
            # form a WAW dependency
Q
qiaolongfei 已提交
436 437 438
            program.global_block().append_op(
                type="fetch_barrier",
                inputs={},
W
Wu Yi 已提交
439
                outputs={"Out": all_recv_outputs},
Q
qiaolongfei 已提交
440 441 442 443
                attrs={
                    "endpoints": pserver_endpoints,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })
Y
Yancey1989 已提交
444

445
        for param_varname, splited_var in six.iteritems(self.param_var_mapping):
T
typhoonzero 已提交
446 447
            if len(splited_var) <= 1:
                continue
448
            orig_param = program.global_block().vars[param_varname]
T
typhoonzero 已提交
449
            program.global_block().append_op(
T
typhoonzero 已提交
450
                type="concat",
T
typhoonzero 已提交
451
                inputs={"X": splited_var},
T
typhoonzero 已提交
452
                outputs={"Out": [orig_param]},
453 454 455 456
                attrs={
                    "axis": 0,
                    RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE
                })
T
typhoonzero 已提交
457

G
gongweibao 已提交
458 459
        self._get_trainer_startup_program(recv_vars=recv_vars, eplist=eplist)

460
        if self.has_distributed_lookup_table:
Q
update  
qiaolongfei 已提交
461 462
            self._replace_lookup_table_op_with_prefetch(program,
                                                        pserver_endpoints)
Y
Yancey1989 已提交
463
            self._split_table_grad_and_add_send_vars(program, pserver_endpoints)
464

W
Wu Yi 已提交
465
    def get_trainer_program(self, wait_port=True):
Y
yi.wu 已提交
466 467 468 469 470 471
        """
        Get transpiled trainer side program.

        Returns:
            Program: trainer side program.
        """
T
typhoonzero 已提交
472
        # remove optimize ops and add a send op to main_program
X
Xin Pan 已提交
473
        # FIXME(typhoonzero): Also ops like clip_gradient, lrn_decay?
T
typhoonzero 已提交
474
        lr_ops = self._get_lr_ops()
475
        delete_ops(self.origin_program.global_block(), self.optimize_ops)
T
typhoonzero 已提交
476 477
        delete_ops(self.origin_program.global_block(), lr_ops)

478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497
        # delete table init op
        if self.has_distributed_lookup_table:
            table_var = self.startup_program.global_block().vars[
                self.table_name]
            table_param_init_op = []
            for op in self.startup_program.global_block().ops:
                if self.table_name in op.output_arg_names:
                    table_param_init_op.append(op)
            init_op_num = len(table_param_init_op)
            if init_op_num != 1:
                raise ValueError("table init op num should be 1, now is " + str(
                    init_op_num))
            table_init_op = table_param_init_op[0]
            self.startup_program.global_block().append_op(
                type="fake_init",
                inputs={},
                outputs={"Out": table_var},
                attrs={"shape": table_init_op.attr('shape')})
            delete_ops(self.startup_program.global_block(), table_param_init_op)

498
        self.origin_program.__str__()
G
gongweibao 已提交
499

W
Wu Yi 已提交
500 501 502
        if wait_port:
            wait_server_ready(self.pserver_endpoints)

503
        return self.origin_program
T
typhoonzero 已提交
504

W
Wu Yi 已提交
505
    def _get_trainer_startup_program(self, recv_vars, eplist):
G
gongweibao 已提交
506 507 508 509
        """
        Get transpiled trainer side startup program.

        Args:
W
Wu Yi 已提交
510
            recv_vars (list): Variable list to recv for current trainer_id
M
minqiyang 已提交
511
            eplist (list): A list of strings indicating
G
gongweibao 已提交
512 513 514 515

        Returns:
            Program: trainer side startup program.
        """
W
Wu Yi 已提交
516
        startup_program = self.startup_program
G
gongweibao 已提交
517 518 519 520

        # FIXME(gongwb): delete not need ops.
        # note that: some parameter is not trainable and those ops can't be deleted.

M
minqiyang 已提交
521
        for varname, splited_var in six.iteritems(self.param_var_mapping):
G
gongweibao 已提交
522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
            # Get the eplist of recv vars
            eps = []
            for var in splited_var:
                index = [v.name for v in recv_vars].index(var.name)
                eps.append(eplist[index])

            for var in splited_var:
                if startup_program.global_block().has_var(var.name):
                    continue

                startup_program.global_block().create_var(
                    name=var.name,
                    persistable=False,
                    type=var.type,
                    dtype=var.dtype,
                    shape=var.shape,
                    lod_level=var.lod_level)

            op = startup_program.global_block().append_op(
                type="recv",
542
                inputs={"X": []},
G
gongweibao 已提交
543 544 545 546 547 548
                outputs={"Out": splited_var},
                attrs={
                    "epmap": eps,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })

W
Wu Yi 已提交
549 550
        fetch_barrier_out = startup_program.global_block().create_var(
            name=framework.generate_control_dev_var_name())
G
gongweibao 已提交
551 552 553
        startup_program.global_block().append_op(
            type="fetch_barrier",
            inputs={},
W
Wu Yi 已提交
554
            outputs={"Out": fetch_barrier_out},
G
gongweibao 已提交
555 556 557 558 559
            attrs={
                "endpoints": self.pserver_endpoints,
                RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
            })

M
minqiyang 已提交
560
        for varname, splited_var in six.iteritems(self.param_var_mapping):
T
tangwei12 已提交
561
            # add concat ops to merge splited parameters received from parameter servers.
G
gongweibao 已提交
562 563
            if len(splited_var) <= 1:
                continue
W
Wu Yi 已提交
564
            # NOTE: if enable memory optimization, origin vars maybe removed.
M
minqiyang 已提交
565
            if varname in startup_program.global_block().vars:
W
Wu Yi 已提交
566 567 568 569 570 571 572 573 574 575
                orig_param = startup_program.global_block().vars[varname]
            else:
                origin_param_var = self.origin_program.global_block().vars[
                    varname]
                orig_param = startup_program.global_block().create_var(
                    name=varname,
                    persistable=origin_param_var.persistable,
                    type=origin_param_var.type,
                    dtype=origin_param_var.dtype,
                    shape=origin_param_var.shape)
G
gongweibao 已提交
576 577 578 579 580 581 582 583
            startup_program.global_block().append_op(
                type="concat",
                inputs={"X": splited_var},
                outputs={"Out": [orig_param]},
                attrs={"axis": 0})

        return startup_program

T
typhoonzero 已提交
584 585
    def get_pserver_program(self, endpoint):
        """
Y
yi.wu 已提交
586
        Get parameter server side program.
587

Y
yi.wu 已提交
588 589
        Args:
            endpoint (str): current parameter server endpoint.
590

Y
yi.wu 已提交
591 592
        Returns:
            Program: the program for current parameter server to run.
T
typhoonzero 已提交
593
        """
Y
yi.wu 已提交
594 595 596 597
        # TODO(panyx0718): Revisit this assumption. what if #blocks > #pservers.
        # NOTE: assume blocks of the same variable is not distributed
        # on the same pserver, only change param/grad varnames for
        # trainers to fetch.
598 599 600
        sys.stderr.write("get_pserver_program() is deprecated, call \
get_pserver_programs() to get pserver main and startup \
in a single call.")
T
typhoonzero 已提交
601 602
        # step1
        pserver_program = Program()
X
Xin Pan 已提交
603
        pserver_program.random_seed = self.origin_program.random_seed
604
        # step2: Create vars to receive vars at parameter servers.
T
typhoonzero 已提交
605 606 607 608 609 610 611 612
        recv_inputs = []
        for v in self.param_grad_ep_mapping[endpoint]["params"]:
            self._clone_var(pserver_program.global_block(), v)
        for v in self.param_grad_ep_mapping[endpoint]["grads"]:
            # create vars for each trainer in global scope, so
            # we don't need to create them when grad arrives.
            # change client side var name to origin name by
            # removing ".trainer_%d" suffix
T
update  
typhoonzero 已提交
613 614 615 616 617
            suff_idx = v.name.find(".trainer_")
            if suff_idx >= 0:
                orig_var_name = v.name[:suff_idx]
            else:
                orig_var_name = v.name
T
typhoonzero 已提交
618 619 620 621 622 623 624 625 626
            # NOTE: single_trainer_var must be created for multi-trainer
            # case to merge grads from multiple trainers
            single_trainer_var = \
                pserver_program.global_block().create_var(
                    name=orig_var_name,
                    persistable=True,
                    type=v.type,
                    dtype=v.dtype,
                    shape=v.shape)
627
            if self.sync_mode and self.trainer_num > 1:
628
                for trainer_id in range(self.trainer_num):
T
typhoonzero 已提交
629 630 631 632 633 634 635 636 637
                    var = pserver_program.global_block().create_var(
                        name="%s.trainer_%d" % (orig_var_name, trainer_id),
                        persistable=False,
                        type=v.type,
                        dtype=v.dtype,
                        shape=v.shape)
                    recv_inputs.append(var)
            else:
                recv_inputs.append(single_trainer_var)
638

Q
qiaolongfei 已提交
639
        # step 3
640
        # Create a union-find data structure from optimize ops,
T
typhoonzero 已提交
641 642 643
        # If two ops are connected, we could add these two ops
        # into one set.
        ufind = self._create_ufind(self.optimize_ops)
Q
qiaolongfei 已提交
644
        # step 3.2
T
typhoonzero 已提交
645 646 647 648
        # Iterate through the ops and append optimize op which
        # located on current pserver
        opt_op_on_pserver = []
        for _, op in enumerate(self.optimize_ops):
649 650
            if self._is_optimizer_op(op) and self._is_opt_op_on_pserver(
                    endpoint, op):
T
typhoonzero 已提交
651
                opt_op_on_pserver.append(op)
Q
qiaolongfei 已提交
652
        # step 3.3
T
typhoonzero 已提交
653
        # Iterate through the ops, and if an op and the optimize ops
654
        # which located on current pserver are in one set, then
T
typhoonzero 已提交
655
        # append it into the sub program.
T
typhoonzero 已提交
656 657 658

        global_ops = []

Y
wip  
yi.wu 已提交
659 660
        def __append_optimize_op__(op, block, grad_to_block_id, merged_var,
                                   lr_ops):
661
            if self._is_optimizer_op(op):
Q
qiaolongfei 已提交
662
                self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
663
                                         self.origin_program, merged_var)
Y
wip  
yi.wu 已提交
664
            elif op not in lr_ops:
Q
Qiyang Min 已提交
665
                self._append_pserver_non_opt_ops(block, op)
666 667 668 669 670 671

        def __op_have_grad_input__(op):
            for varname in op.input_arg_names:
                if varname.find("@GRAD") >= 0:
                    return varname
            return ""
T
typhoonzero 已提交
672

Y
Yancey1989 已提交
673
        def __clone_lr_op_sub_block__(op, program, lr_block):
Q
Qiyang Min 已提交
674 675 676 677 678 679 680 681
            if not op.has_attr('sub_block'):
                return

            origin_block_desc = op.attr('sub_block')
            origin_block = self.origin_program.block(origin_block_desc.id)
            assert isinstance(origin_block, Block)
            # we put the new sub block to new block to follow the block
            # hierarchy of the original blocks
W
Wu Yi 已提交
682
            new_sub_block = program._create_block(lr_block.idx)
Q
Qiyang Min 已提交
683 684 685

            # clone vars
            for var in origin_block.vars:
W
Wu Yi 已提交
686
                new_sub_block._clone_variable(var)
Q
Qiyang Min 已提交
687 688

            # clone ops
Y
Yancey1989 已提交
689 690
            for origin_op in origin_block.ops:
                cloned_op = self._clone_lr_op(program, new_sub_block, origin_op)
Q
Qiyang Min 已提交
691
                # clone sub_block of op
Y
Yancey1989 已提交
692
                __clone_lr_op_sub_block__(cloned_op, program, new_sub_block)
Q
Qiyang Min 已提交
693 694

            # reset the block of op
W
Wu Yi 已提交
695
            op._set_attr('sub_block', new_sub_block)
Q
Qiyang Min 已提交
696

697
        # append lr decay ops to the child block if exists
698
        lr_ops = self._get_lr_ops()
699 700
        # record optimize blocks and we can run them on pserver parallel
        optimize_blocks = []
701
        if len(lr_ops) > 0:
W
Wu Yi 已提交
702
            lr_decay_block = pserver_program._create_block(
Q
qiaolongfei 已提交
703
                pserver_program.num_blocks - 1)
704
            optimize_blocks.append(lr_decay_block)
705
            for _, op in enumerate(lr_ops):
Y
Yancey1989 已提交
706
                cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op)
Q
Qiyang Min 已提交
707
                # append sub blocks to pserver_program in lr_decay_op
Y
Yancey1989 已提交
708 709
                __clone_lr_op_sub_block__(cloned_op, pserver_program,
                                          lr_decay_block)
710

T
typhoonzero 已提交
711
        # append op to the current block
Q
qiaolongfei 已提交
712
        grad_to_block_id = []
Q
qiaolongfei 已提交
713
        pre_block_idx = pserver_program.num_blocks - 1
T
typhoonzero 已提交
714
        for idx, opt_op in enumerate(opt_op_on_pserver):
W
Wu Yi 已提交
715
            per_opt_block = pserver_program._create_block(pre_block_idx)
716
            optimize_blocks.append(per_opt_block)
717
            optimize_target_param_name = opt_op.attr(OP_ROLE_VAR_ATTR_NAME)[0]
718
            # append grad merging ops before clip and weight decay
719 720
            # e.g. merge grad -> L2Decay op -> clip op -> optimize
            merged_var = None
721
            for _, op in enumerate(self.optimize_ops):
722 723 724 725 726
                # find the origin grad var before clipping/L2Decay,
                # merged_var should be the input var name of L2Decaybuil
                grad_varname_for_block = op.attr(OP_ROLE_VAR_ATTR_NAME)[1]
                if op.attr(OP_ROLE_VAR_ATTR_NAME)[
                        0] == optimize_target_param_name:
727 728 729
                    merged_var = self._append_pserver_grad_merge_ops(
                        per_opt_block, grad_varname_for_block, endpoint,
                        grad_to_block_id, self.origin_program)
730 731 732 733 734 735
                    if merged_var:
                        break  # append optimize op once then append other ops.
            if merged_var:
                for _, op in enumerate(self.optimize_ops):
                    # optimizer is connected to itself
                    if op.attr(OP_ROLE_VAR_ATTR_NAME)[0] == optimize_target_param_name and \
736
                            op not in global_ops:
737 738 739 740 741
                        log("append opt op: ", op.type, op.input_arg_names,
                            merged_var)
                        __append_optimize_op__(op, per_opt_block,
                                               grad_to_block_id, merged_var,
                                               lr_ops)
T
typhoonzero 已提交
742

W
Wu Yi 已提交
743 744
        # dedup grad to ids list
        grad_to_block_id = list(set(grad_to_block_id))
T
typhoonzero 已提交
745
        # append global ops
746
        if global_ops:
W
Wu Yi 已提交
747
            opt_state_block = pserver_program._create_block(
Q
qiaolongfei 已提交
748
                pserver_program.num_blocks - 1)
749
            optimize_blocks.append(opt_state_block)
Q
qiaolongfei 已提交
750
            for glb_op in global_ops:
X
Xi Chen 已提交
751
                __append_optimize_op__(glb_op, opt_state_block,
Y
wip  
yi.wu 已提交
752
                                       grad_to_block_id, None, lr_ops)
T
typhoonzero 已提交
753

754
        # process distributed lookup_table
Q
qiaolongfei 已提交
755
        prefetch_var_name_to_block_id = []
756 757
        if self.has_distributed_lookup_table:
            pserver_index = self.pserver_endpoints.index(endpoint)
758
            table_opt_block = self._create_table_optimize_block(
759
                pserver_index, pserver_program, pre_block_idx, grad_to_block_id)
760
            optimize_blocks.append(table_opt_block)
T
tangwei12 已提交
761
            lookup_table_var_name_to_block_id = self._create_prefetch_block(
762
                pserver_index, pserver_program, table_opt_block)
T
tangwei12 已提交
763 764
            checkpoint_block_id = self._create_checkpoint_save_block(
                pserver_program, table_opt_block.idx)
765

T
tangwei12 已提交
766
            pserver_program._distributed_lookup_table = self.table_name
T
tangwei12 已提交
767 768
            prefetch_var_name_to_block_id.extend(
                lookup_table_var_name_to_block_id)
769

770
        attrs = {
771
            "optimize_blocks": optimize_blocks,
772 773 774
            "endpoint": endpoint,
            "Fanin": self.trainer_num,
            "sync_mode": self.sync_mode,
Y
Yancey1989 已提交
775
            "grad_to_block_id": grad_to_block_id,
776
        }
T
tangwei12 已提交
777 778

        if self.has_distributed_lookup_table:
T
tangwei12 已提交
779
            attrs['checkpint_block_id'] = checkpoint_block_id
780

T
tangwei12 已提交
781 782 783 784
        if len(prefetch_var_name_to_block_id) > 0:
            attrs[
                'prefetch_var_name_to_block_id'] = prefetch_var_name_to_block_id

T
typhoonzero 已提交
785 786 787 788 789
        # step5 append the listen_and_serv op
        pserver_program.global_block().append_op(
            type="listen_and_serv",
            inputs={'X': recv_inputs},
            outputs={},
790
            attrs=attrs)
791

T
tangwei12 已提交
792
        # add distributed attrs
T
tangwei12 已提交
793
        pserver_program._slice_vars_and_attrs = self._get_slice_vars_and_attrs(
T
tangwei12 已提交
794
            endpoint)
795

W
Wu Yi 已提交
796
        pserver_program._sync_with_cpp()
W
Wu Yi 已提交
797 798
        # save pserver program to generate pserver side startup relatively.
        self.pserver_program = pserver_program
T
typhoonzero 已提交
799 800
        return pserver_program

W
Wu Yi 已提交
801 802 803 804 805 806
    def get_pserver_programs(self, endpoint):
        """
        Get pserver side main program and startup program for distributed training.

        Args:
            endpoint (str): current pserver endpoint.
M
minqiyang 已提交
807

W
Wu Yi 已提交
808 809 810 811
        Returns:
            tuple: (main_program, startup_program), of type "Program"
        """
        pserver_prog = self.get_pserver_program(endpoint)
W
Wu Yi 已提交
812 813
        pserver_startup = self.get_startup_program(
            endpoint, pserver_program=pserver_prog)
W
Wu Yi 已提交
814 815
        return pserver_prog, pserver_startup

816 817
    def get_startup_program(self,
                            endpoint,
W
Wu Yi 已提交
818
                            pserver_program=None,
819
                            startup_program=None):
T
typhoonzero 已提交
820
        """
W
Wu Yi 已提交
821 822
        **Deprecated**

T
typhoonzero 已提交
823 824 825
        Get startup program for current parameter server.
        Modify operator input variables if there are variables that
        were split to several blocks.
Y
yi.wu 已提交
826 827 828

        Args:
            endpoint (str): current pserver endpoint.
W
Wu Yi 已提交
829 830
            pserver_program (Program): deprecated, call get_pserver_program first.
            startup_program (Program): deprecated, should pass startup_program
M
minqiyang 已提交
831
                when initalizing
832

Y
yi.wu 已提交
833 834
        Returns:
            Program: parameter server side startup program.
T
typhoonzero 已提交
835
        """
836 837 838
        sys.stderr.write("get_startup_program() is deprecated, call \
get_pserver_programs() to get pserver main and startup \
in a single call.")
W
Wu Yi 已提交
839
        if pserver_program != None:
840 841 842
            sys.stderr.write("passing pserver_program to get_startup_program() \
is deprecated, you can use new API get_pserver_programs() to \
get both pserver main program and startup program.")
W
Wu Yi 已提交
843
        if startup_program != None:
844 845 846
            sys.stderr.write("passing startup_program to get_startup_program() \
is deprecated, use fluid.program_guard() or pass this argument \
to transpile() call.")
W
Wu Yi 已提交
847

T
typhoonzero 已提交
848
        s_prog = Program()
W
Wu Yi 已提交
849
        orig_s_prog = self.startup_program
X
Xin Pan 已提交
850
        s_prog.random_seed = orig_s_prog.random_seed
T
typhoonzero 已提交
851 852 853 854 855 856 857 858 859 860 861
        params = self.param_grad_ep_mapping[endpoint]["params"]

        def _get_splited_name_and_shape(varname):
            for idx, splited_param in enumerate(params):
                pname = splited_param.name
                if same_or_split_var(pname, varname) and varname != pname:
                    return pname, splited_param.shape
            return "", []

        # 1. create vars in pserver program to startup program
        pserver_vars = pserver_program.global_block().vars
862
        created_var_map = collections.OrderedDict()
M
minqiyang 已提交
863
        for _, var in six.iteritems(pserver_vars):
W
Wu Yi 已提交
864
            tmpvar = s_prog.global_block()._clone_variable(var)
T
typhoonzero 已提交
865 866 867 868
            created_var_map[var.name] = tmpvar

        # 2. rename op outputs
        for op in orig_s_prog.global_block().ops:
869
            new_outputs = collections.OrderedDict()
T
typhoonzero 已提交
870 871
            # do not append startup op if var is not on this pserver
            op_on_pserver = False
G
gongweibao 已提交
872 873 874 875 876 877 878 879 880 881
            # TODO(gongwb): remove this line.
            if op.type not in ["recv", "fetch_barrier", "concat"]:
                for key in op.output_names:
                    newname, _ = _get_splited_name_and_shape(op.output(key)[0])
                    if newname:
                        op_on_pserver = True
                        new_outputs[key] = created_var_map[newname]
                    elif op.output(key)[0] in pserver_vars:
                        op_on_pserver = True
                        new_outputs[key] = pserver_vars[op.output(key)[0]]
T
typhoonzero 已提交
882 883

            if op_on_pserver:
884 885 886
                # most startup program ops have no inputs
                new_inputs = self._get_input_map_from_op(pserver_vars, op)

T
typhoonzero 已提交
887 888 889
                if op.type in [
                        "gaussian_random", "fill_constant", "uniform_random"
                ]:
W
Wu Yi 已提交
890
                    op._set_attr("shape", list(new_outputs["Out"].shape))
T
typhoonzero 已提交
891 892 893 894
                s_prog.global_block().append_op(
                    type=op.type,
                    inputs=new_inputs,
                    outputs=new_outputs,
G
gongweibao 已提交
895
                    attrs=op.all_attrs())
896 897

        # add slice vars
T
tangwei12 已提交
898
        s_prog._slice_vars_and_attrs = self._get_slice_vars_and_attrs(endpoint)
899

T
typhoonzero 已提交
900 901
        return s_prog

T
tangwei12 已提交
902 903 904
    def _get_slice_vars_and_attrs(self, endpoint):
        slice_vars_and_attrs = []
        block_suffix = "block"
905
        for param in self.param_grad_ep_mapping[endpoint]["params"]:
T
tangwei12 已提交
906
            orig_var_name, block_name, _ = self._get_varname_parts(param.name)
T
tangwei12 已提交
907
            if not block_name:
908 909
                continue

T
tangwei12 已提交
910
            block_idx = int(block_name.split(block_suffix)[1])
911 912 913 914 915 916
            orig_var = self.origin_program.global_block().vars[orig_var_name]

            skip_numel = 0
            slice_vars = self.param_var_mapping[orig_var_name]
            for slice_var in slice_vars[:block_idx]:
                skip_numel += reduce(lambda x, y: x * y, slice_var.shape)
T
tangwei12 已提交
917
            slice_vars_and_attrs.append([orig_var, skip_numel, param])
918

T
tangwei12 已提交
919
        return slice_vars_and_attrs
920

921 922
    # ====================== private transpiler functions =====================

Y
yi.wu 已提交
923 924 925 926 927 928 929 930 931
    def _has_distributed_lookup_table(self):
        # process lookup_table_op
        # 1. check all lookup_table_op is distributed
        # 2. check all lookup_table_op share the same table.
        distributed_lookup_table_ops = []
        # support only one distributed_lookup_table now
        self.table_name = None
        for op in self.origin_program.global_block().ops:
            if op.type == LOOKUP_TABLE_TYPE:
G
gongweibao 已提交
932
                if op.attr('is_distributed') is True:
Y
yi.wu 已提交
933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983
                    if self.table_name is None:
                        self.table_name = op.input("W")[0]
                    if self.table_name != op.input("W")[0]:
                        raise RuntimeError("all distributed lookup_table_ops"
                                           " should have only one table")
                    distributed_lookup_table_ops.append(op)
                else:
                    if self.table_name is not None:
                        assert op.input("W")[0] != self.table_name

        return len(distributed_lookup_table_ops) > 0

    def _update_dist_lookup_table_vars(self, param_list, grad_list,
                                       params_grads):
        # TODO(wuyi): put find a way to put dist lookup table stuff all together.
        # update self.table_param_grad and self.trainer_side_table_grad_list
        program = self.origin_program
        if self.has_distributed_lookup_table:
            param_list = [
                param for param in param_list if param.name != self.table_name
            ]
            grad_list = [
                grad for grad in grad_list
                if grad.name != grad_var_name(self.table_name)
            ]
            self.table_param_grad = [
                param_grad for param_grad in params_grads
                if param_grad[0].name == self.table_name
            ][0]
            table_grad_var = self.table_param_grad[1]
            if self.sync_mode:
                self.trainer_side_table_grad_list = [
                    program.global_block().create_var(
                        name="%s.trainer_%d.pserver_%d" %
                        (table_grad_var.name, self.trainer_id, index),
                        type=table_grad_var.type,
                        shape=table_grad_var.shape,
                        dtype=table_grad_var.dtype)
                    for index in range(len(self.pserver_endpoints))
                ]
            else:
                self.trainer_side_table_grad_list = [
                    program.global_block().create_var(
                        name="%s.pserver_%d" % (table_grad_var.name, index),
                        type=table_grad_var.type,
                        shape=table_grad_var.shape,
                        dtype=table_grad_var.dtype)
                    for index in range(len(self.pserver_endpoints))
                ]
        return param_list, grad_list

G
gongweibao 已提交
984
    def _init_splited_vars(self):
Y
yi.wu 已提交
985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
        # update these mappings for further transpile:
        # 1. param_var_mapping: param var name -> [splited params vars]
        # 2. grad_var_mapping: grad var name -> [splited grads vars]
        # 3. grad_param_mapping: grad.blockx -> param.blockx
        # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []}

        param_list = []
        grad_list = []
        param_grad_set = set()
        for p, g in self.params_grads:
            # skip parameter marked not trainable
            if type(p) == Parameter and p.trainable == False:
                continue
            if p.name not in param_grad_set:
                param_list.append(p)
                param_grad_set.add(p.name)
            if g.name not in param_grad_set:
                grad_list.append(g)
                param_grad_set.add(g.name)

        param_list, grad_list = self._update_dist_lookup_table_vars(
            param_list, grad_list, self.params_grads)

G
gongweibao 已提交
1008
        if self.config.slice_var_up:
Y
yi.wu 已提交
1009 1010
            # when we slice var up into blocks, we will slice the var according to
            # pserver services' count. A pserver may have two or more listening ports.
G
gongweibao 已提交
1011 1012 1013
            grad_blocks = slice_variable(grad_list,
                                         len(self.pserver_endpoints),
                                         self.config.min_block_size)
Y
yi.wu 已提交
1014
            param_blocks = slice_variable(param_list,
G
gongweibao 已提交
1015 1016
                                          len(self.pserver_endpoints),
                                          self.config.min_block_size)
Y
yi.wu 已提交
1017 1018 1019
        else:
            # when we do NOT slice var up into blocks, we will always slice params
            # grads into one block.
G
gongweibao 已提交
1020 1021 1022 1023
            grad_blocks = slice_variable(grad_list, 1,
                                         self.config.min_block_size)
            param_blocks = slice_variable(param_list, 1,
                                          self.config.min_block_size)
Y
yi.wu 已提交
1024 1025
        assert (len(grad_blocks) == len(param_blocks))

1026
        # origin_param_name -> [splited_param_vars]
Y
yi.wu 已提交
1027 1028
        self.param_var_mapping = self._create_vars_from_blocklist(
            self.origin_program, param_blocks)
1029
        # origin_grad_name -> [splited_grad_vars]
Y
yi.wu 已提交
1030 1031 1032 1033
        self.grad_var_mapping = self._create_vars_from_blocklist(
            self.origin_program,
            grad_blocks,
            add_trainer_suffix=self.trainer_num > 1)
1034
        # dict(grad_splited_var -> param_splited_var)
1035
        self.grad_param_mapping = collections.OrderedDict()
Y
yi.wu 已提交
1036 1037 1038
        for g, p in zip(grad_blocks, param_blocks):
            g_name, g_bid, _ = g.split(":")
            p_name, p_bid, _ = p.split(":")
T
tangwei12 已提交
1039
            self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \
1040
                self.param_var_mapping[p_name][int(p_bid)]
Y
yi.wu 已提交
1041 1042

        # create mapping of endpoint -> split var to create pserver side program
1043
        self.param_grad_ep_mapping = collections.OrderedDict()
Y
yi.wu 已提交
1044 1045 1046 1047 1048 1049 1050 1051 1052
        [
            self.param_grad_ep_mapping.update({
                ep: {
                    "params": [],
                    "grads": []
                }
            }) for ep in self.pserver_endpoints
        ]

1053
    # transpiler function for dis lookup_table
Q
update  
qiaolongfei 已提交
1054 1055
    def _replace_lookup_table_op_with_prefetch(self, program,
                                               pserver_endpoints):
1056
        # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
1057
        self.all_in_ids_vars = []
Q
qiaolongfei 已提交
1058 1059
        self.all_prefetch_input_vars = []
        self.all_prefetch_output_vars = []
1060 1061
        self.all_out_emb_vars = []
        lookup_table_op_index = -1
1062 1063 1064 1065 1066 1067 1068 1069 1070

        continue_search_lookup_table_op = True
        while continue_search_lookup_table_op:
            continue_search_lookup_table_op = False
            all_ops = program.global_block().ops
            for op in all_ops:
                if op.type == LOOKUP_TABLE_TYPE:
                    continue_search_lookup_table_op = True

1071 1072
                    lookup_table_op_index = lookup_table_op_index if lookup_table_op_index != -1 else list(
                        all_ops).index(op)
1073 1074 1075
                    ids_name = op.input("Ids")
                    out_name = op.output("Out")

Q
qiaolongfei 已提交
1076
                    ids_var = program.global_block().vars[ids_name[0]]
1077
                    self.all_in_ids_vars.append(ids_var)
Q
qiaolongfei 已提交
1078 1079

                    out_var = program.global_block().vars[out_name[0]]
1080
                    self.all_out_emb_vars.append(out_var)
1081 1082

                    # delete lookup_table_op
1083
                    delete_ops(program.global_block(), [op])
1084 1085 1086
                    # break for loop
                    break

1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
        for index in range(len(self.pserver_endpoints)):
            in_var = program.global_block().create_var(
                name=str("prefetch_compress_in_tmp_" + str(index)),
                type=self.all_in_ids_vars[0].type,
                shape=self.all_in_ids_vars[0].shape,
                dtype=self.all_in_ids_vars[0].dtype)
            self.all_prefetch_input_vars.append(in_var)

            out_var = program.global_block().create_var(
                name=str("prefetch_compress_out_tmp_" + str(index)),
                type=self.all_out_emb_vars[0].type,
                shape=self.all_out_emb_vars[0].shape,
                dtype=self.all_out_emb_vars[0].dtype)
            self.all_prefetch_output_vars.append(out_var)

        # insert split_ids_op
        program.global_block()._insert_op(
            index=lookup_table_op_index,
            type="split_ids",
            inputs={'Ids': self.all_in_ids_vars},
            outputs={"Out": self.all_prefetch_input_vars})

        # insert prefetch_op
        program.global_block()._insert_op(
            index=lookup_table_op_index + 1,
            type="prefetch",
            inputs={'X': self.all_prefetch_input_vars},
            outputs={"Out": self.all_prefetch_output_vars},
            attrs={
                "epmap": pserver_endpoints,
                # FIXME(qiao) temporarily disable this config because prefetch
                # is not act as other rpc op, it's more like a forward op
                # RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
            })

        # insert concat_op
        program.global_block()._insert_op(
            index=lookup_table_op_index + 2,
            type="merge_ids",
            inputs={
                'Ids': self.all_in_ids_vars,
                'Rows': self.all_prefetch_input_vars,
                'X': self.all_prefetch_output_vars
            },
            outputs={"Out": self.all_out_emb_vars})

Y
Yancey1989 已提交
1133
    def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
1134
        # 2. add split_ids_op and send_op to send gradient to pservers
1135

1136 1137
        # there should only be one table_name
        all_ops = program.global_block().ops
T
typhoonzero 已提交
1138
        table_grad_name = grad_var_name(self.table_name)
1139 1140 1141 1142
        for op in all_ops:
            if table_grad_name in op.output_arg_names:
                op_index = list(all_ops).index(op)
                # insert split_ids_op
W
Wu Yi 已提交
1143
                program.global_block()._insert_op(
1144 1145 1146 1147 1148
                    index=op_index + 1,
                    type="split_ids",
                    inputs={
                        'Ids': [program.global_block().vars[table_grad_name]]
                    },
1149 1150
                    outputs={"Out": self.trainer_side_table_grad_list},
                    attrs={RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE})
W
Wu Yi 已提交
1151
                program.global_block()._insert_op(
1152
                    index=op_index + 2,
1153
                    type="send",
1154
                    inputs={'X': self.trainer_side_table_grad_list},
1155 1156 1157 1158 1159
                    outputs={
                        'Out':
                        [self.grad_name_to_send_dummy_out[self.table_name]]
                        if self.sync_mode else []
                    },
Y
Yancey1989 已提交
1160
                    attrs={
1161
                        "sync_mode": not self.sync_mode,
Y
Yancey1989 已提交
1162
                        "epmap": pserver_endpoints,
W
Wu Yi 已提交
1163 1164 1165 1166 1167
                        RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
                        OP_ROLE_VAR_ATTR_NAME: [
                            self.grad_name_to_param_name[table_grad_name],
                            table_grad_name
                        ]
Y
Yancey1989 已提交
1168
                    })
1169 1170 1171 1172 1173 1174
                break

    def _create_prefetch_block(self, pserver_index, pserver_program,
                               optimize_block):
        # STEP: create prefetch block
        table_var = pserver_program.global_block().vars[self.table_name]
Q
qiaolongfei 已提交
1175
        prefetch_var_name_to_block_id = []
1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200
        prefetch_block = pserver_program._create_block(optimize_block.idx)
        trainer_ids = self.all_prefetch_input_vars[pserver_index]
        pserver_ids = pserver_program.global_block().create_var(
            name=trainer_ids.name,
            type=trainer_ids.type,
            shape=trainer_ids.shape,
            dtype=trainer_ids.dtype)
        trainer_out = self.all_prefetch_output_vars[pserver_index]
        pserver_out = pserver_program.global_block().create_var(
            name=trainer_out.name,
            type=trainer_out.type,
            shape=trainer_out.shape,
            dtype=trainer_out.dtype)
        prefetch_block.append_op(
            type="lookup_sparse_table",
            inputs={'Ids': pserver_ids,
                    "W": table_var},
            outputs={"Out": pserver_out},
            attrs={
                "is_sparse": True,  # has no effect on lookup_table op
                "is_distributed": True,
                "padding_idx": -1
            })
        prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str(
            prefetch_block.idx))
Q
qiaolongfei 已提交
1201
        return prefetch_var_name_to_block_id
1202 1203

    def _create_table_optimize_block(self, pserver_index, pserver_program,
1204
                                     pre_block_idx, grad_to_block_id):
1205
        # STEP: create table optimize block
1206
        table_opt_block = pserver_program._create_block(pre_block_idx)
1207
        # create table param and grad var in pserver program
1208 1209 1210 1211 1212 1213 1214
        # create table optimize block in pserver program
        table_opt_op = [
            op for op in self.optimize_ops
            if 'Param' in op.input_names and op.input("Param")[0] ==
            self.table_name
        ][0]

Y
Yancey1989 已提交
1215 1216
        origin_param_var = self.origin_program.global_block().vars[
            self.table_name]
T
tangwei12 已提交
1217

T
tangwei12 已提交
1218
        zero_dim = int(
T
bug fix  
tangwei12 已提交
1219 1220
            math.ceil(origin_param_var.shape[0] / float(
                len(self.pserver_endpoints))))
T
tangwei12 已提交
1221 1222 1223
        table_shape = list(origin_param_var.shape)
        table_shape[0] = zero_dim

Y
Yancey1989 已提交
1224 1225
        param_var = pserver_program.global_block().create_var(
            name=origin_param_var.name,
T
tangwei12 已提交
1226
            shape=table_shape,
Y
Yancey1989 已提交
1227 1228 1229
            dtype=origin_param_var.dtype,
            type=core.VarDesc.VarType.SELECTED_ROWS,
            persistable=True)
1230

1231 1232
        # parameter must be selected rows
        param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS)
W
Wu Yi 已提交
1233
        grad_var = pserver_program.global_block()._clone_variable(
T
typhoonzero 已提交
1234
            self.origin_program.global_block().vars[grad_var_name(
1235
                self.table_name)])
1236

1237 1238 1239
        lr_var = pserver_program.global_block()._clone_variable(
            self.origin_program.global_block().vars[table_opt_op.input(
                "LearningRate")[0]])
1240

1241 1242 1243
        if self.sync_mode:
            # create grad vars in pserver program
            table_grad_var = self.table_param_grad[1]
1244
            pserver_side_table_grad_list = [
1245 1246 1247 1248 1249 1250 1251 1252 1253
                pserver_program.global_block().create_var(
                    name="%s.trainer_%d.pserver_%d" %
                    (table_grad_var.name, index, pserver_index),
                    type=table_grad_var.type,
                    shape=table_grad_var.shape,
                    dtype=table_grad_var.dtype)
                for index in range(self.trainer_num)
            ]

1254
            # append sum op for pserver_side_table_grad_list
1255 1256
            table_opt_block.append_op(
                type="sum",
1257
                inputs={"X": pserver_side_table_grad_list},
1258 1259
                outputs={"Out": [grad_var]},
                attrs={"use_mkldnn": False})
1260 1261
        else:
            # in async_mode, for table gradient, it also need to be splited to each parameter server
1262
            origin_grad_name = grad_var.name
1263 1264
            splited_grad_name = self.trainer_side_table_grad_list[
                pserver_index].name
1265 1266
            if not splited_grad_name.startswith(origin_grad_name):
                raise ValueError("origin_grad_var: " + splited_grad_name +
1267
                                 " grad_var:" + grad_var.name)
W
Wu Yi 已提交
1268
            grad_var = pserver_program.global_block()._rename_var(
1269
                origin_grad_name, splited_grad_name)
1270 1271 1272 1273 1274 1275 1276

        inputs = {
            "Param": [param_var],
            "Grad": [grad_var],
            "LearningRate": [lr_var]
        }
        outputs = {"ParamOut": [param_var]}
1277
        # only support sgd now
1278 1279 1280 1281
        import logging
        logging.warn(
            "distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of "
            + table_opt_op.type)
1282
        table_opt_block.append_op(type="sgd", inputs=inputs, outputs=outputs)
1283

1284 1285 1286
        # add table parameter gradient and it's block id to grad_to_block_id
        grad_to_block_id.append(grad_var.name + ":" + str(table_opt_block.idx))

1287 1288
        return table_opt_block

T
tangwei12 已提交
1289 1290 1291 1292 1293 1294
    def _create_checkpoint_save_block(self, pserver_program, pre_block_idx):
        """
        create a new block to handle save checkpoint.
        """
        import os

T
tangwei12 已提交
1295
        pserver_program.global_block().create_var(
T
tangwei12 已提交
1296
            name="kLookupTablePath",
T
tangwei12 已提交
1297 1298
            persistable=True,
            type=core.VarDesc.VarType.RAW)
T
tangwei12 已提交
1299

W
Wu Yi 已提交
1300
        checkpoint_save_block = pserver_program._create_block(pre_block_idx)
T
tangwei12 已提交
1301
        # this 'file_path' do not be used in save lookup table variable
T
tangwei12 已提交
1302 1303 1304 1305
        checkpoint_save_block.append_op(
            type='save',
            inputs={'X': [self.table_name]},
            outputs={},
T
tangwei12 已提交
1306
            attrs={'file_path': "none"})
T
tangwei12 已提交
1307 1308 1309

        return checkpoint_save_block.idx

T
typhoonzero 已提交
1310 1311 1312 1313 1314
    def _create_vars_from_blocklist(self,
                                    program,
                                    block_list,
                                    add_trainer_suffix=False):
        """
1315
        Create vars for each split.
T
typhoonzero 已提交
1316 1317
        NOTE: only grads need to be named for different trainers, use
              add_trainer_suffix to rename the grad vars.
1318 1319 1320 1321
        Args:
            program (ProgramDesc): ProgramDesc which gradients blong.
            block_list (list[(varname, block_id, block_size)]): List of gradient blocks.
            add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True.
1322
        Returns:
1323
            var_mapping (collections.OrderedDict(varname->[new_varname_variable])):A dict mapping
1324
                from original var name to each var split.
T
typhoonzero 已提交
1325
        """
1326 1327

        # varname->[(block_id, current_block_size)]
1328
        block_map = collections.OrderedDict()
1329

1330
        var_mapping = collections.OrderedDict()
T
typhoonzero 已提交
1331 1332
        for block_str in block_list:
            varname, offset, size = block_str.split(":")
1333
            if varname not in block_map:
T
typhoonzero 已提交
1334
                block_map[varname] = []
1335
            block_map[varname].append((int(offset), int(size)))
Y
yi.wu 已提交
1336

M
minqiyang 已提交
1337
        for varname, splited in six.iteritems(block_map):
T
typhoonzero 已提交
1338
            orig_var = program.global_block().var(varname)
T
typhoonzero 已提交
1339
            if len(splited) == 1:
1340
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1341
                    new_var_name = "%s.trainer_%d" % \
T
tangwei12 已提交
1342
                                   (orig_var.name, self.trainer_id)
W
Wu Yi 已提交
1343
                    program.global_block()._rename_var(varname, new_var_name)
T
typhoonzero 已提交
1344 1345 1346 1347 1348
                    var_mapping[varname] = \
                        [program.global_block().var(new_var_name)]
                else:
                    var_mapping[varname] = \
                        [program.global_block().var(orig_var.name)]
T
typhoonzero 已提交
1349
                continue
T
typhoonzero 已提交
1350
            var_mapping[varname] = []
T
typhoonzero 已提交
1351 1352 1353 1354
            orig_shape = orig_var.shape
            orig_dim1_flatten = 1
            if len(orig_shape) >= 2:
                orig_dim1_flatten = reduce(lambda x, y: x * y, orig_shape[1:])
T
typhoonzero 已提交
1355

T
typhoonzero 已提交
1356
            for i, block in enumerate(splited):
T
typhoonzero 已提交
1357
                size = block[1]
M
minqiyang 已提交
1358
                rows = size // orig_dim1_flatten
T
typhoonzero 已提交
1359 1360 1361
                splited_shape = [rows]
                if len(orig_shape) >= 2:
                    splited_shape.extend(orig_shape[1:])
T
typhoonzero 已提交
1362
                new_var_name = ""
1363
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1364
                    new_var_name = "%s.block%d.trainer_%d" % \
T
tangwei12 已提交
1365
                                   (varname, i, self.trainer_id)
T
typhoonzero 已提交
1366 1367
                else:
                    new_var_name = "%s.block%d" % \
T
tangwei12 已提交
1368
                                   (varname, i)
T
typhoonzero 已提交
1369
                var = program.global_block().create_var(
T
typhoonzero 已提交
1370 1371
                    name=new_var_name,
                    persistable=False,
T
typhoonzero 已提交
1372
                    dtype=orig_var.dtype,
1373
                    type=orig_var.type,
T
typhoonzero 已提交
1374
                    shape=splited_shape)  # flattend splited var
T
typhoonzero 已提交
1375
                var_mapping[varname].append(var)
W
Wu Yi 已提交
1376
            program.global_block()._sync_with_cpp()
T
typhoonzero 已提交
1377
        return var_mapping
T
done  
typhoonzero 已提交
1378

1379
    def _clone_var(self, block, var, persistable=True):
T
done  
typhoonzero 已提交
1380 1381 1382 1383 1384 1385
        return block.create_var(
            name=var.name,
            shape=var.shape,
            dtype=var.dtype,
            type=var.type,
            lod_level=var.lod_level,
1386
            persistable=persistable)
T
done  
typhoonzero 已提交
1387

Y
Yancey1989 已提交
1388
    def _insert_split_op(self, program, orig_var, index, splited_vars):
Y
update  
Yancey1989 已提交
1389 1390 1391 1392
        if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS:
            height_sections = []
            for v in splited_vars:
                height_sections.append(v.shape[0])
W
Wu Yi 已提交
1393
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1394 1395 1396 1397
                index=index + 1,
                type="split_selected_rows",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
1398 1399 1400 1401
                attrs={
                    "height_sections": height_sections,
                    RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE
                })
Y
update  
Yancey1989 已提交
1402 1403 1404 1405
        elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR:
            sections = []
            for v in splited_vars:
                sections.append(v.shape[0])
W
Wu Yi 已提交
1406
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1407 1408 1409 1410
                index=index + 1,
                type="split_byref",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
1411 1412 1413 1414
                attrs={
                    "sections": sections,
                    RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE
                })
Y
update  
Yancey1989 已提交
1415 1416 1417
        else:
            AssertionError("Variable type should be in set "
                           "[LOD_TENSOR, SELECTED_ROWS]")
T
done  
typhoonzero 已提交
1418

T
typhoonzero 已提交
1419 1420 1421 1422
    def _get_optimizer_input_shape(self, op_type, varkey, orig_shape,
                                   param_shape):
        """
        Returns the shape for optimizer inputs that need to be reshaped when
1423
        Param and Grad is split to multiple servers.
T
typhoonzero 已提交
1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435
        """
        # HACK(typhoonzero): Should use functions of corresponding optimizer in
        # optimizer.py to get the shape, do not  bind this in the transpiler.
        if op_type == "adam":
            if varkey in ["Moment1", "Moment2"]:
                return param_shape
        elif op_type == "adagrad":
            if varkey == "Moment":
                return param_shape
        elif op_type == "adamax":
            if varkey in ["Moment", "InfNorm"]:
                return param_shape
1436
        elif op_type in ["momentum", "lars_momentum"]:
T
typhoonzero 已提交
1437 1438
            if varkey == "Velocity":
                return param_shape
W
Wu Yi 已提交
1439 1440
        elif op_type == "rmsprop":
            if varkey in ["Moment", "MeanSquare"]:
T
typhoonzero 已提交
1441
                return param_shape
1442 1443 1444
        elif op_type == "decayed_adagrad":
            if varkey == "Moment":
                return param_shape
T
typhoonzero 已提交
1445 1446
        elif op_type == "sgd":
            pass
1447 1448 1449 1450
        else:
            raise ValueError(
                "Not supported optimizer for distributed training: %s" %
                op_type)
T
typhoonzero 已提交
1451 1452
        return orig_shape

1453 1454
    def _get_varname_parts(self, varname):
        # returns origin, blockid, trainerid
T
typhoonzero 已提交
1455
        orig_var_name = ""
1456 1457 1458 1459 1460 1461 1462 1463 1464 1465
        trainer_part = ""
        block_part = ""
        trainer_idx = varname.find(".trainer_")
        if trainer_idx >= 0:
            trainer_part = varname[trainer_idx + 1:]
        else:
            trainer_idx = len(varname)
        block_index = varname.find(".block")
        if block_index >= 0:
            block_part = varname[block_index + 1:trainer_idx]
T
typhoonzero 已提交
1466
        else:
1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488
            block_index = len(varname)
        orig_var_name = varname[0:min(block_index, trainer_idx)]
        return orig_var_name, block_part, trainer_part

    def _orig_varname(self, varname):
        orig, _, _ = self._get_varname_parts(varname)
        return orig

    def _append_pserver_grad_merge_ops(self, optimize_block,
                                       grad_varname_for_block, endpoint,
                                       grad_to_block_id, origin_program):
        program = optimize_block.program
        pserver_block = program.global_block()
        grad_block = None
        for g in self.param_grad_ep_mapping[endpoint]["grads"]:
            if self._orig_varname(g.name) == \
                    self._orig_varname(grad_varname_for_block):
                grad_block = g
                break
        if not grad_block:
            # do not append this op if current endpoint
            # is not dealing with this grad block
1489
            return None
1490 1491 1492 1493
        orig_varname, block_name, trainer_name = self._get_varname_parts(
            grad_block.name)
        if block_name:
            merged_var_name = '.'.join([orig_varname, block_name])
T
typhoonzero 已提交
1494
        else:
1495
            merged_var_name = orig_varname
1496 1497

        merged_var = pserver_block.vars[merged_var_name]
1498 1499 1500
        grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx))
        if self.sync_mode and self.trainer_num > 1:
            vars2merge = []
1501
            for i in range(self.trainer_num):
1502
                per_trainer_name = "%s.trainer_%d" % \
T
tangwei12 已提交
1503
                                   (merged_var_name, i)
1504 1505 1506 1507
                vars2merge.append(pserver_block.vars[per_trainer_name])
            optimize_block.append_op(
                type="sum",
                inputs={"X": vars2merge},
1508 1509
                outputs={"Out": merged_var},
                attrs={"use_mkldnn": False})
Q
qiaolongfei 已提交
1510 1511 1512 1513 1514
            optimize_block.append_op(
                type="scale",
                inputs={"X": merged_var},
                outputs={"Out": merged_var},
                attrs={"scale": 1.0 / float(self.trainer_num)})
1515
        return merged_var
T
typhoonzero 已提交
1516

1517
    def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
1518
                            grad_to_block_id, origin_program, merged_var):
1519
        program = optimize_block.program
T
typhoonzero 已提交
1520
        pserver_block = program.global_block()
1521
        new_inputs = collections.OrderedDict()
W
Wu Yi 已提交
1522 1523 1524 1525 1526 1527 1528 1529 1530 1531

        def _get_param_block(opt_op):
            # param is already created on global program
            param_block = None
            for p in self.param_grad_ep_mapping[endpoint]["params"]:
                if same_or_split_var(p.name, opt_op.input("Param")[0]):
                    param_block = p
                    break
            return param_block

T
typhoonzero 已提交
1532
        for key in opt_op.input_names:
T
typhoonzero 已提交
1533 1534 1535
            if key == "Grad":
                new_inputs[key] = merged_var
            elif key == "Param":
W
Wu Yi 已提交
1536
                param_block = _get_param_block(opt_op)
T
typhoonzero 已提交
1537 1538
                if not param_block:
                    return
T
typhoonzero 已提交
1539
                tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1540
                    name=param_block.name,
T
typhoonzero 已提交
1541
                    persistable=True,
T
typhoonzero 已提交
1542 1543 1544
                    dtype=param_block.dtype,
                    shape=param_block.shape)
                new_inputs[key] = tmpvar
1545
            elif key == "LearningRate":
1546
                # learning rate variable has already be created by non-optimize op,
1547
                # don't create it once again.
1548
                lr_varname = opt_op.input(key)[0]
1549
                if lr_varname in pserver_block.vars:
1550 1551 1552 1553 1554 1555 1556 1557 1558
                    new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]]
                else:
                    origin_var = origin_program.global_block().vars[lr_varname]
                    tmpvar = pserver_block.create_var(
                        name=origin_var.name,
                        persistable=origin_var.persistable,
                        dtype=origin_var.dtype,
                        shape=origin_var.shape)
                    new_inputs[key] = tmpvar
T
typhoonzero 已提交
1559

T
typhoonzero 已提交
1560
        for key in opt_op.input_names:
1561
            new_shape = None
W
Wu Yi 已提交
1562
            if key in ["Param", "Grad", "LearningRate"]:
T
typhoonzero 已提交
1563
                continue
1564
            var = self.origin_program.global_block().vars[opt_op.input(key)[0]]
T
typhoonzero 已提交
1565 1566 1567 1568
            # update accumulator variable shape
            param_shape = new_inputs["Param"].shape
            new_shape = self._get_optimizer_input_shape(opt_op.type, key,
                                                        var.shape, param_shape)
T
typhoonzero 已提交
1569
            tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1570 1571 1572 1573 1574
                name=var.name,
                persistable=var.persistable,
                dtype=var.dtype,
                shape=new_shape)
            new_inputs[key] = tmpvar
T
typhoonzero 已提交
1575

1576
        # change output's ParamOut variable
1577 1578
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1579
        outputs["ParamOut"] = new_inputs["Param"]
1580
        optimize_block.append_op(
T
typhoonzero 已提交
1581 1582
            type=opt_op.type,
            inputs=new_inputs,
T
typhoonzero 已提交
1583
            outputs=outputs,
G
gongweibao 已提交
1584
            attrs=opt_op.all_attrs())
T
typhoonzero 已提交
1585

1586 1587
    def _is_splited_grad_var(self, var, var_dict):
        grad_block = None
M
minqiyang 已提交
1588
        for _, g in six.iteritems(var_dict):
1589 1590 1591 1592 1593 1594
            if self._orig_varname(g.name) == self._orig_varname(var.name):
                if g.name.find(".trainer_") == -1:
                    grad_block = g
                    break
        return grad_block

Q
Qiyang Min 已提交
1595 1596 1597
    def _clone_lr_op(self, program, block, op):
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, op)
M
minqiyang 已提交
1598
        for key, varlist in six.iteritems(inputs):
Q
Qiyang Min 已提交
1599 1600 1601 1602
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
                if var not in program.global_block().vars:
W
Wu Yi 已提交
1603
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1604 1605 1606

        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, op)
M
minqiyang 已提交
1607
        for key, varlist in six.iteritems(outputs):
Q
Qiyang Min 已提交
1608 1609 1610 1611
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
                if var not in program.global_block().vars:
W
Wu Yi 已提交
1612
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1613

Y
Yancey1989 已提交
1614
        return block.append_op(
G
gongweibao 已提交
1615
            type=op.type, inputs=inputs, outputs=outputs, attrs=op.all_attrs())
Q
Qiyang Min 已提交
1616 1617

    def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
1618
        program = optimize_block.program
1619
        # Append the ops for parameters that do not need to be optimized/updated
1620 1621
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, opt_op)
M
minqiyang 已提交
1622
        for key, varlist in six.iteritems(inputs):
1623 1624
            if not isinstance(varlist, list):
                varlist = [varlist]
T
typhoonzero 已提交
1625
            for var in varlist:
1626 1627 1628 1629 1630 1631
                # for ops like clipping and weight decay, get the splited var
                # for inputs/outputs
                grad_block = self._is_splited_grad_var(
                    var, program.global_block().vars)
                if grad_block:
                    inputs[key] = grad_block
1632
                elif var.name not in program.global_block().vars:
1633
                    program.global_block().create_var(
T
typhoonzero 已提交
1634 1635 1636 1637 1638
                        name=var.name,
                        persistable=var.persistable,
                        dtype=var.dtype,
                        shape=var.shape)

1639 1640
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
M
minqiyang 已提交
1641
        for key, varlist in six.iteritems(outputs):
1642 1643 1644
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
1645 1646 1647 1648
                grad_block = self._is_splited_grad_var(
                    var, program.global_block().vars)
                if grad_block:
                    outputs[key] = grad_block
1649
                elif var.name not in program.global_block().vars:
W
Wu Yi 已提交
1650
                    program.global_block()._clone_variable(var)
1651

Y
Yancey1989 已提交
1652
        return optimize_block.append_op(
T
typhoonzero 已提交
1653
            type=opt_op.type,
T
typhoonzero 已提交
1654 1655
            inputs=inputs,
            outputs=outputs,
G
gongweibao 已提交
1656
            attrs=opt_op.all_attrs())
T
typhoonzero 已提交
1657

1658 1659 1660 1661
    def _is_op_connected(self, op1, op2):
        # If one op's input is another op's output or
        # one op's output is another op's input, we say
        # the two operator is connected.
Q
qiaolongfei 已提交
1662
        if set(op1.desc.output_arg_names()) & set(op2.desc.input_arg_names()) or \
T
tangwei12 已提交
1663
                set(op1.desc.input_arg_names()) & set(op2.desc.output_arg_names()):
1664 1665 1666 1667 1668 1669
            return True
        return False

    def _create_ufind(self, optimize_ops):
        # Create a unit find data struct by optimize ops
        ufind = UnionFind(optimize_ops)
1670 1671
        for i in range(len(optimize_ops)):
            for j in range(i, len(optimize_ops)):
1672 1673 1674 1675 1676 1677
                op1 = optimize_ops[i]
                op2 = optimize_ops[j]
                if self._is_op_connected(op1, op2):
                    ufind.union(op1, op2)
        return ufind

1678
    def _is_optimizer_op(self, op):
T
typhoonzero 已提交
1679
        if "Param" in op.input_names and \
T
tangwei12 已提交
1680
                "LearningRate" in op.input_names:
1681 1682 1683 1684 1685 1686 1687
            return True
        return False

    def _is_opt_op_on_pserver(self, endpoint, op):
        param_names = [
            p.name for p in self.param_grad_ep_mapping[endpoint]["params"]
        ]
T
typhoonzero 已提交
1688
        if op.input("Param")[0] in param_names:
1689 1690 1691
            return True
        else:
            for n in param_names:
T
typhoonzero 已提交
1692
                param = op.input("Param")[0]
T
typhoonzero 已提交
1693
                if same_or_split_var(n, param) and n != param:
1694 1695 1696
                    return True
            return False

T
typhoonzero 已提交
1697
    def _get_input_map_from_op(self, varmap, op):
1698
        """Returns a dict from op input name to the vars in varmap."""
1699
        iomap = collections.OrderedDict()
T
typhoonzero 已提交
1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710
        for key in op.input_names:
            vars = []
            for varname in op.input(key):
                vars.append(varmap[varname])
            if len(vars) == 1:
                iomap[key] = vars[0]
            else:
                iomap[key] = vars
        return iomap

    def _get_output_map_from_op(self, varmap, op):
1711
        """Returns a dict from op output name to the vars in varmap."""
1712
        iomap = collections.OrderedDict()
T
typhoonzero 已提交
1713 1714 1715 1716 1717 1718 1719 1720 1721
        for key in op.output_names:
            vars = []
            for varname in op.output(key):
                vars.append(varmap[varname])
            if len(vars) == 1:
                iomap[key] = vars[0]
            else:
                iomap[key] = vars
        return iomap
1722 1723

    def _get_lr_ops(self):
1724 1725 1726
        lr_ops = []
        block = self.origin_program.global_block()
        for op in block.ops:
X
fix  
Xin Pan 已提交
1727 1728 1729 1730
            role_id = int(op.attr(RPC_OP_ROLE_ATTR_NAME))
            if role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) or \
                role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) | \
                    int(OPT_OP_ROLE_ATTR_VALUE):
1731 1732 1733 1734 1735
                lr_ops.append(op)
                log("append lr op: ", op.type)
        return lr_ops

    def _get_lr_ops_deprecated(self):
1736 1737 1738 1739
        lr_ops = []
        # find learning rate variables by optimize op
        lr_vars = set()
        for op in self.optimize_ops:
1740
            if self._is_optimizer_op(op):
1741 1742 1743 1744
                lr_vars.add(op.input("LearningRate")[0])

        find_ops = []
        # find ops which output is lr var
1745
        block = self.origin_program.global_block()
1746 1747 1748 1749 1750
        for op in block.ops:
            if set(op.output_arg_names) & lr_vars:
                find_ops.append(op)
        # make a union find struct by the ops in default_main_program
        ufind = UnionFind(block.ops)
1751

1752 1753 1754 1755 1756
        for op1 in block.ops:
            for op2 in block.ops:
                # NOTE: we need to skip all optimize ops, since it is connected
                # with forward/backward ops and lr ops, we only need the lr ops.
                if op1 != op2 and self._is_op_connected(op1, op2) and \
T
tangwei12 已提交
1757
                        not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2):
1758 1759 1760 1761 1762 1763
                    ufind.union(op1, op2)
        # find all ops which is related with lr var
        for op1 in block.ops:
            for op2 in find_ops:
                if ufind.is_connected(op1, op2):
                    lr_ops.append(op1)
1764 1765
                    # we only need to append op for once
                    break
1766
        return lr_ops
Y
Yancey1989 已提交
1767

W
Wu Yi 已提交
1768 1769 1770 1771 1772
    def _is_opt_role_op(self, op):
        # NOTE: depend on oprole to find out whether this op is for
        # optimize
        op_maker = core.op_proto_and_checker_maker
        optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize
G
gongweibao 已提交
1773 1774
        if op_maker.kOpRoleAttrName() in op.attr_names and \
                int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(optimize_role):
W
Wu Yi 已提交
1775 1776 1777
            return True
        return False

Y
Yancey1989 已提交
1778
    def _get_optimize_pass(self):
1779
        """
1780
        Get optimizer operators, parameters and gradients from origin_program
1781 1782 1783 1784
        Returns:
            opt_ops (list): optimize operators.
            params_grads (dict): paramter->gradient.
        """
Y
Yancey1989 已提交
1785 1786 1787
        block = self.origin_program.global_block()
        opt_ops = []
        params_grads = []
1788 1789
        # tmp set to dedup
        optimize_params = set()
1790
        origin_var_dict = self.origin_program.global_block().vars
Y
Yancey1989 已提交
1791
        for op in block.ops:
W
Wu Yi 已提交
1792
            if self._is_opt_role_op(op):
Y
Yancey1989 已提交
1793
                opt_ops.append(op)
1794 1795 1796 1797 1798 1799
                if op.attr(OP_ROLE_VAR_ATTR_NAME):
                    param_name = op.attr(OP_ROLE_VAR_ATTR_NAME)[0]
                    grad_name = op.attr(OP_ROLE_VAR_ATTR_NAME)[1]
                    if not param_name in optimize_params:
                        optimize_params.add(param_name)
                        log("adding param_grad pair: ", param_name, grad_name)
1800 1801
                        params_grads.append([
                            origin_var_dict[param_name],
1802
                            origin_var_dict[grad_name]
1803
                        ])
Y
Yancey1989 已提交
1804 1805 1806
            else:
                pass
        return opt_ops, params_grads