distribute_transpiler.py 69.5 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
Y
fix ut  
yi.wu 已提交
37
import random
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, \
Q
Qiyang Min 已提交
42
                        default_startup_program, Block, \
W
Wu Yi 已提交
43
                        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 52
RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName(
)
RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
T
done  
typhoonzero 已提交
53 54


T
typhoonzero 已提交
55 56 57 58 59 60
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
61

T
typhoonzero 已提交
62 63
    def __str__(self):
        return "%s:%d:%d" % (self.varname, self.offset, self.size)
T
done  
typhoonzero 已提交
64 65


66 67 68 69
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
70
def slice_variable(var_list, slice_count, min_block_size):
T
typhoonzero 已提交
71
    """
72 73 74 75 76 77
    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
78
    minimum block size 8K elements (maybe 16bit or 32bit or 64bit).
79 80 81

    Args:
        var_list (list): List of variables.
82 83
        slice_count (int): Numel of count that variables will be sliced, which
            could be the pserver services' count.
84 85
        min_block_size (int): Minimum splitted block size.
    Returns:
86
        blocks (list[(varname, block_id, current_block_size)]): A list
87
            of VarBlocks. Each VarBlock specifies a shard of the var.
T
typhoonzero 已提交
88 89 90
    """
    blocks = []
    for var in var_list:
91
        split_count = slice_count
T
typhoonzero 已提交
92 93 94 95
        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
96
        if max_pserver_count < slice_count:
T
typhoonzero 已提交
97 98 99 100 101 102 103 104 105
            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
106
        # update split_count after aligning
T
typhoonzero 已提交
107
        split_count = int(math.ceil(var_numel / float(block_size)))
108
        for block_id in range(split_count):
T
typhoonzero 已提交
109 110 111 112 113 114 115
            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
116 117 118 119 120 121 122
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
123
        We can use bandwidth effiently when data size is larger than 2MB.If you
G
gongweibao 已提交
124 125 126 127 128 129 130 131
        want to change it, please be sure you see the slice_variable function.
    """

    slice_var_up = True
    split_method = None
    min_block_size = 8192


Y
gen rst  
yi.wu 已提交
132
class DistributeTranspiler(object):
Y
yi.wu 已提交
133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166
    """
    **DistributeTranspiler**

    Convert the fluid program to distributed data-parallelism programs.

    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.

    Examples:
        .. code-block:: python

           # Define your model before these codes.
           port = os.getenv("PADDLE_PSERVER_PORT", "6174")
           pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "")
           eplist = []
           for ip in pserver_ips.split(","):
                eplist.append(':'.join([ip, port]))
           pserver_endpoints = ",".join(eplist)
           trainers = int(os.getenv("PADDLE_TRAINERS"))
           current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port
           trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
           role = os.getenv("PADDLE_TRAINING_ROLE")

           t = distribute_transpiler.DistributeTranspiler()
           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()
    """
Y
Yancey1989 已提交
167

G
gongweibao 已提交
168 169 170 171 172 173 174 175 176 177 178 179
    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

        assert (self.config.min_block_size >= 8192)
        assert (self.config.split_method.__bases__[0] == PSDispatcher)

180 181 182 183 184
    def transpile(self,
                  trainer_id,
                  program=None,
                  pservers="127.0.0.1:6174",
                  trainers=1,
W
Wu Yi 已提交
185 186
                  sync_mode=True,
                  startup_program=None):
187
        """
Y
yi.wu 已提交
188 189 190 191 192 193 194 195 196 197 198
        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.
            trainers (int): number of trainers in the distributed job.
            sync_mode (bool): Do sync training or not, default is True.
W
Wu Yi 已提交
199 200
            startup_program (Program|None): startup_program to transpile,
                default is fluid.default_main_program().
201 202 203
        """
        if program is None:
            program = default_main_program()
W
Wu Yi 已提交
204 205
        if startup_program is None:
            startup_program = default_startup_program()
206
        self.origin_program = program
W
Wu Yi 已提交
207 208
        self.startup_program = startup_program
        self.origin_startup_program = self.startup_program.clone()
G
gongweibao 已提交
209

210 211 212 213 214 215 216
        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
217
        ps_dispatcher = self.config.split_method(self.pserver_endpoints)
218
        self.has_distributed_lookup_table = self._has_distributed_lookup_table()
219
        self.param_name_to_grad_name = dict()
W
Wu Yi 已提交
220
        self.grad_name_to_param_name = dict()
221 222
        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
223
            self.grad_name_to_param_name[grad_var.name] = param_var.name
224

G
gongweibao 已提交
225
        # step 1: split and create vars, then put splited vars in dicts for later use.
G
gongweibao 已提交
226
        self._init_splited_vars()
227

G
gongweibao 已提交
228
        # step 2: insert send op to send gradient vars to parameter servers
Y
Yancey1989 已提交
229
        ps_dispatcher.reset()
Y
update  
Yancey1989 已提交
230
        send_vars = []
231 232 233 234 235 236

        # 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
237
        grad_var_mapping_items = list(six.iteritems(self.grad_var_mapping))
238

G
gongweibao 已提交
239
        if not self.config.slice_var_up:
240
            random.seed(self.origin_program.random_seed)
S
seiriosPlus 已提交
241
            random.shuffle(grad_var_mapping_items)
242

243 244
        grad_name_to_send_dummy_out = dict()
        for grad_varname, splited_vars in grad_var_mapping_items:
Y
update  
Yancey1989 已提交
245
            eplist = ps_dispatcher.dispatch(splited_vars)
246

G
gongweibao 已提交
247
            if not self.config.slice_var_up:
248 249
                assert (len(splited_vars) == 1)

250
            splited_grad_varname = grad_varname
Y
Yancey1989 已提交
251
            if len(splited_vars) == 1:
252
                splited_grad_varname = splited_vars[0].name
Y
Yancey1989 已提交
253
                index = find_op_by_output_arg(program.global_block(),
254
                                              splited_grad_varname)
Y
Yancey1989 已提交
255
            elif len(splited_vars) > 1:
256
                orig_var = program.global_block().vars[splited_grad_varname]
Y
Yancey1989 已提交
257
                index = find_op_by_output_arg(program.global_block(),
258
                                              splited_grad_varname)
Y
Yancey1989 已提交
259
                self._insert_split_op(program, orig_var, index, splited_vars)
Y
update  
Yancey1989 已提交
260
                index += 1
Y
Yancey1989 已提交
261 262
            else:
                AssertionError("Can not insert the send op by original "
263
                               "variable name :", splited_grad_varname)
Y
Yancey1989 已提交
264

W
Wu Yi 已提交
265 266
            dummy_output = program.global_block().create_var(
                name=framework.generate_control_dev_var_name())
267
            grad_name_to_send_dummy_out[grad_varname] = dummy_output
W
Wu Yi 已提交
268

W
Wu Yi 已提交
269 270 271 272
            # 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
273
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
274
                index=index + 1,
275
                type="send",
Y
update  
Yancey1989 已提交
276
                inputs={"X": splited_vars},
277
                outputs={"Out": dummy_output},
Y
Yancey1989 已提交
278 279
                attrs={
                    "epmap": eplist,
280
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
W
Wu Yi 已提交
281 282 283 284
                    OP_ROLE_VAR_ATTR_NAME: [
                        self.grad_name_to_param_name[grad_varname],
                        splited_grad_varname
                    ],
285
                    "sync_mode": not self.sync_mode,
Y
Yancey1989 已提交
286
                })
Y
update  
Yancey1989 已提交
287 288
            for _, var in enumerate(splited_vars):
                send_vars.append(var)
Y
Yancey1989 已提交
289 290

        if self.sync_mode:
W
Wu Yi 已提交
291 292 293
            send_barrier_out = program.global_block().create_var(
                name=framework.generate_control_dev_var_name())
            input_deps = grad_name_to_send_dummy_out.values()
Y
Yancey1989 已提交
294 295
            program.global_block().append_op(
                type="send_barrier",
W
Wu Yi 已提交
296 297
                inputs={"X": input_deps},
                outputs={"Out": send_barrier_out},
Y
Yancey1989 已提交
298 299
                attrs={
                    "endpoints": pserver_endpoints,
Y
Yancey1989 已提交
300
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Y
Yancey1989 已提交
301
                })
Y
Yancey1989 已提交
302

G
gongweibao 已提交
303
        # step 3: insert recv op to receive parameters from parameter server
Y
Yancey1989 已提交
304
        recv_vars = []
Y
update  
Yancey1989 已提交
305
        for _, var in enumerate(send_vars):
306
            recv_vars.append(self.grad_param_mapping[var])
Y
update  
Yancey1989 已提交
307
        ps_dispatcher.reset()
Y
Yancey1989 已提交
308 309
        eplist = ps_dispatcher.dispatch(recv_vars)

T
typhoonzero 已提交
310
        for i, ep in enumerate(eplist):
Y
Yancey1989 已提交
311 312
            self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
            self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
313

Y
Yancey1989 已提交
314
        # step4: Concat the parameters splits together after recv.
W
Wu Yi 已提交
315
        all_recv_outputs = []
316
        for param_varname, splited_var in six.iteritems(self.param_var_mapping):
Y
Yancey1989 已提交
317 318 319 320
            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
321 322 323 324 325 326 327
            if self.sync_mode:
                recv_dep_in = send_barrier_out
            else:
                # connect deps to send op in async mode
                recv_dep_in = grad_name_to_send_dummy_out[
                    self.param_name_to_grad_name[param_varname]]
            all_recv_outputs.extend(splited_var)
W
Wu Yi 已提交
328 329 330 331 332 333 334 335 336
            # 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
337 338
            program.global_block().append_op(
                type="recv",
W
Wu Yi 已提交
339
                inputs={"X": [recv_dep_in]},
Y
Yancey1989 已提交
340 341 342
                outputs={"Out": splited_var},
                attrs={
                    "epmap": eps,
343
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
W
Wu Yi 已提交
344 345
                    OP_ROLE_VAR_ATTR_NAME:
                    [param_varname, recv_op_role_var_name],
346
                    "sync_mode": not self.sync_mode
Y
Yancey1989 已提交
347
                })
T
typhoonzero 已提交
348

Q
qiaolongfei 已提交
349
        if self.sync_mode:
W
Wu Yi 已提交
350
            # form a WAW dependency
Q
qiaolongfei 已提交
351 352 353
            program.global_block().append_op(
                type="fetch_barrier",
                inputs={},
W
Wu Yi 已提交
354
                outputs={"Out": all_recv_outputs},
Q
qiaolongfei 已提交
355 356 357 358
                attrs={
                    "endpoints": pserver_endpoints,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })
Y
Yancey1989 已提交
359

360
        for param_varname, splited_var in six.iteritems(self.param_var_mapping):
T
typhoonzero 已提交
361 362
            if len(splited_var) <= 1:
                continue
363
            orig_param = program.global_block().vars[param_varname]
T
typhoonzero 已提交
364
            program.global_block().append_op(
T
typhoonzero 已提交
365
                type="concat",
T
typhoonzero 已提交
366
                inputs={"X": splited_var},
T
typhoonzero 已提交
367
                outputs={"Out": [orig_param]},
T
typhoonzero 已提交
368
                attrs={"axis": 0})
T
typhoonzero 已提交
369

G
gongweibao 已提交
370 371
        self._get_trainer_startup_program(recv_vars=recv_vars, eplist=eplist)

372
        if self.has_distributed_lookup_table:
Q
update  
qiaolongfei 已提交
373 374
            self._replace_lookup_table_op_with_prefetch(program,
                                                        pserver_endpoints)
Y
Yancey1989 已提交
375
            self._split_table_grad_and_add_send_vars(program, pserver_endpoints)
376

T
typhoonzero 已提交
377
    def get_trainer_program(self):
Y
yi.wu 已提交
378 379 380 381 382 383
        """
        Get transpiled trainer side program.

        Returns:
            Program: trainer side program.
        """
T
typhoonzero 已提交
384
        # remove optimize ops and add a send op to main_program
X
Xin Pan 已提交
385
        # FIXME(typhoonzero): Also ops like clip_gradient, lrn_decay?
386
        delete_ops(self.origin_program.global_block(), self.optimize_ops)
387
        self.origin_program.__str__()
G
gongweibao 已提交
388

389
        return self.origin_program
T
typhoonzero 已提交
390

W
Wu Yi 已提交
391
    def _get_trainer_startup_program(self, recv_vars, eplist):
G
gongweibao 已提交
392 393 394 395
        """
        Get transpiled trainer side startup program.

        Args:
W
Wu Yi 已提交
396 397
            recv_vars (list): Variable list to recv for current trainer_id
            eplist (list): A list of strings indicating 
G
gongweibao 已提交
398 399 400 401

        Returns:
            Program: trainer side startup program.
        """
W
Wu Yi 已提交
402
        startup_program = self.startup_program
G
gongweibao 已提交
403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427

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

        for varname, splited_var in self.param_var_mapping.iteritems():
            # 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",
428
                inputs={"X": []},
G
gongweibao 已提交
429 430 431 432 433 434
                outputs={"Out": splited_var},
                attrs={
                    "epmap": eps,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })

W
Wu Yi 已提交
435 436
        fetch_barrier_out = startup_program.global_block().create_var(
            name=framework.generate_control_dev_var_name())
G
gongweibao 已提交
437 438 439
        startup_program.global_block().append_op(
            type="fetch_barrier",
            inputs={},
W
Wu Yi 已提交
440
            outputs={"Out": fetch_barrier_out},
G
gongweibao 已提交
441 442 443 444 445 446 447 448 449
            attrs={
                "endpoints": self.pserver_endpoints,
                RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
            })

        for varname, splited_var in self.param_var_mapping.iteritems():
            #add concat ops to merge splited parameters received from parameter servers.
            if len(splited_var) <= 1:
                continue
W
Wu Yi 已提交
450 451 452 453 454 455 456 457 458 459 460 461
            # NOTE: if enable memory optimization, origin vars maybe removed.
            if startup_program.global_block().vars.has_key(varname):
                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
462 463 464 465 466 467 468 469
            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
470 471
    def get_pserver_program(self, endpoint):
        """
Y
yi.wu 已提交
472
        Get parameter server side program.
473

Y
yi.wu 已提交
474 475
        Args:
            endpoint (str): current parameter server endpoint.
476

Y
yi.wu 已提交
477 478
        Returns:
            Program: the program for current parameter server to run.
T
typhoonzero 已提交
479
        """
Y
yi.wu 已提交
480 481 482 483
        # 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.
W
Wu Yi 已提交
484 485 486
        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
487 488
        # step1
        pserver_program = Program()
X
Xin Pan 已提交
489
        pserver_program.random_seed = self.origin_program.random_seed
490
        # step2: Create vars to receive vars at parameter servers.
T
typhoonzero 已提交
491 492 493 494 495 496 497 498
        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
499 500 501 502 503
            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
504 505 506 507 508 509 510 511 512
            # 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)
513
            if self.sync_mode and self.trainer_num > 1:
514
                for trainer_id in range(self.trainer_num):
T
typhoonzero 已提交
515 516 517 518 519 520 521 522 523
                    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)
524

Q
qiaolongfei 已提交
525
        # step 3
526
        # Create a union-find data structure from optimize ops,
T
typhoonzero 已提交
527 528 529
        # 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
530
        # step 3.2
T
typhoonzero 已提交
531 532 533 534
        # Iterate through the ops and append optimize op which
        # located on current pserver
        opt_op_on_pserver = []
        for _, op in enumerate(self.optimize_ops):
535 536
            if self._is_optimizer_op(op) and self._is_opt_op_on_pserver(
                    endpoint, op):
T
typhoonzero 已提交
537
                opt_op_on_pserver.append(op)
Q
qiaolongfei 已提交
538
        # step 3.3
T
typhoonzero 已提交
539
        # Iterate through the ops, and if an op and the optimize ops
540
        # which located on current pserver are in one set, then
T
typhoonzero 已提交
541
        # append it into the sub program.
T
typhoonzero 已提交
542 543 544

        global_ops = []

Y
wip  
yi.wu 已提交
545 546
        def __append_optimize_op__(op, block, grad_to_block_id, merged_var,
                                   lr_ops):
547
            if self._is_optimizer_op(op):
Q
qiaolongfei 已提交
548
                self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
549
                                         self.origin_program, merged_var)
Y
wip  
yi.wu 已提交
550
            elif op not in lr_ops:
Q
Qiyang Min 已提交
551
                self._append_pserver_non_opt_ops(block, op)
552 553 554 555 556 557

        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
558

Y
Yancey1989 已提交
559
        def __clone_lr_op_sub_block__(op, program, lr_block):
Q
Qiyang Min 已提交
560 561 562 563 564 565 566 567
            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
Y
Yancey1989 已提交
568
            new_sub_block = program.create_block(lr_block.idx)
Q
Qiyang Min 已提交
569 570 571

            # clone vars
            for var in origin_block.vars:
W
Wu Yi 已提交
572
                new_sub_block._clone_variable(var)
Q
Qiyang Min 已提交
573 574

            # clone ops
Y
Yancey1989 已提交
575 576
            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
577
                # clone sub_block of op
Y
Yancey1989 已提交
578
                __clone_lr_op_sub_block__(cloned_op, program, new_sub_block)
Q
Qiyang Min 已提交
579 580 581 582

            # reset the block of op
            op.set_attr('sub_block', new_sub_block)

583
        # append lr decay ops to the child block if exists
584
        lr_ops = self._get_lr_ops()
585 586
        # record optimize blocks and we can run them on pserver parallel
        optimize_blocks = []
587
        if len(lr_ops) > 0:
Q
qiaolongfei 已提交
588 589
            lr_decay_block = pserver_program.create_block(
                pserver_program.num_blocks - 1)
590
            optimize_blocks.append(lr_decay_block)
591
            for _, op in enumerate(lr_ops):
Y
Yancey1989 已提交
592
                cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op)
Q
Qiyang Min 已提交
593
                # append sub blocks to pserver_program in lr_decay_op
Y
Yancey1989 已提交
594 595
                __clone_lr_op_sub_block__(cloned_op, pserver_program,
                                          lr_decay_block)
596

T
typhoonzero 已提交
597
        # append op to the current block
Q
qiaolongfei 已提交
598
        grad_to_block_id = []
Q
qiaolongfei 已提交
599
        pre_block_idx = pserver_program.num_blocks - 1
T
typhoonzero 已提交
600
        for idx, opt_op in enumerate(opt_op_on_pserver):
601
            per_opt_block = pserver_program.create_block(pre_block_idx)
602
            optimize_blocks.append(per_opt_block)
603
            # append grad merging ops before clip and weight decay
604
            # cases may like:
T
typhoonzero 已提交
605
            # L2Decay op -> clip op -> optimize
606 607 608 609 610 611 612
            for _, op in enumerate(self.optimize_ops):
                # find the origin @GRAD var before clipping
                grad_varname_for_block = __op_have_grad_input__(op)
                if ufind.is_connected(op, opt_op) and grad_varname_for_block:
                    merged_var = self._append_pserver_grad_merge_ops(
                        per_opt_block, grad_varname_for_block, endpoint,
                        grad_to_block_id, self.origin_program)
T
typhoonzero 已提交
613
                    break  # append optimize op once then append other ops.
T
typhoonzero 已提交
614 615
            for _, op in enumerate(self.optimize_ops):
                # optimizer is connected to itself
616
                if ufind.is_connected(op, opt_op) and op not in global_ops:
617
                    __append_optimize_op__(op, per_opt_block, grad_to_block_id,
Y
wip  
yi.wu 已提交
618
                                           merged_var, lr_ops)
T
typhoonzero 已提交
619

W
Wu Yi 已提交
620 621
        # dedup grad to ids list
        grad_to_block_id = list(set(grad_to_block_id))
T
typhoonzero 已提交
622
        # append global ops
623
        if global_ops:
Q
qiaolongfei 已提交
624 625
            opt_state_block = pserver_program.create_block(
                pserver_program.num_blocks - 1)
626
            optimize_blocks.append(opt_state_block)
Q
qiaolongfei 已提交
627
            for glb_op in global_ops:
X
Xi Chen 已提交
628
                __append_optimize_op__(glb_op, opt_state_block,
Y
wip  
yi.wu 已提交
629
                                       grad_to_block_id, None, lr_ops)
T
typhoonzero 已提交
630

631
        # process distributed lookup_table
Q
qiaolongfei 已提交
632
        prefetch_var_name_to_block_id = []
633 634
        if self.has_distributed_lookup_table:
            pserver_index = self.pserver_endpoints.index(endpoint)
635
            table_opt_block = self._create_table_optimize_block(
636
                pserver_index, pserver_program, pre_block_idx, grad_to_block_id)
637
            optimize_blocks.append(table_opt_block)
Q
qiaolongfei 已提交
638
            prefetch_var_name_to_block_id = self._create_prefetch_block(
639
                pserver_index, pserver_program, table_opt_block)
T
tangwei12 已提交
640 641
            checkpoint_block_id = self._create_checkpoint_save_block(
                pserver_program, table_opt_block.idx)
642 643 644 645

        # NOTE: if has_distributed_lookup_table is False, then prefetch_block will
        # not be executed, so it's safe to use optimize_block to hold the place
        if self.has_distributed_lookup_table:
Q
qiaolongfei 已提交
646
            assert len(prefetch_var_name_to_block_id) > 0
647
        else:
Q
qiaolongfei 已提交
648
            assert len(prefetch_var_name_to_block_id) == 0
649

650
        attrs = {
651
            "optimize_blocks": optimize_blocks,
652 653 654
            "endpoint": endpoint,
            "Fanin": self.trainer_num,
            "sync_mode": self.sync_mode,
Y
Yancey1989 已提交
655
            "grad_to_block_id": grad_to_block_id,
656 657 658 659
        }
        if len(prefetch_var_name_to_block_id) > 0:
            attrs['prefetch_var_name_to_block_id'] \
                = prefetch_var_name_to_block_id
T
tangwei12 已提交
660
            attrs['checkpint_block_id'] = checkpoint_block_id
661

T
typhoonzero 已提交
662 663 664 665 666
        # step5 append the listen_and_serv op
        pserver_program.global_block().append_op(
            type="listen_and_serv",
            inputs={'X': recv_inputs},
            outputs={},
667
            attrs=attrs)
668

W
Wu Yi 已提交
669
        pserver_program._sync_with_cpp()
W
Wu Yi 已提交
670 671
        # save pserver program to generate pserver side startup relatively.
        self.pserver_program = pserver_program
T
typhoonzero 已提交
672 673
        return pserver_program

W
Wu Yi 已提交
674 675 676 677 678 679 680 681 682 683 684 685 686 687
    def get_pserver_programs(self, endpoint):
        """
        Get pserver side main program and startup program for distributed training.

        Args:
            endpoint (str): current pserver endpoint.
        
        Returns:
            tuple: (main_program, startup_program), of type "Program"
        """
        pserver_prog = self.get_pserver_program(endpoint)
        pserver_startup = self.get_startup_program(endpoint)
        return pserver_prog, pserver_startup

688 689
    def get_startup_program(self,
                            endpoint,
W
Wu Yi 已提交
690
                            pserver_program=None,
691
                            startup_program=None):
T
typhoonzero 已提交
692
        """
W
Wu Yi 已提交
693 694
        **Deprecated**

T
typhoonzero 已提交
695 696 697
        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
698 699 700

        Args:
            endpoint (str): current pserver endpoint.
W
Wu Yi 已提交
701 702 703
            pserver_program (Program): deprecated, call get_pserver_program first.
            startup_program (Program): deprecated, should pass startup_program
                when initalizing 
704

Y
yi.wu 已提交
705 706
        Returns:
            Program: parameter server side startup program.
T
typhoonzero 已提交
707
        """
W
Wu Yi 已提交
708 709 710 711 712 713 714 715 716 717 718 719
        sys.stderr.write("get_startup_program() is deprecated, call\
            get_pserver_programs() to get pserver main and startup\
            in a single call.")
        if pserver_program != None:
            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.")
        if startup_program != None:
            sys.stderr.write("passing startup_program to get_startup_program()\
                is deprecated, use fluid.program_guard() or pass this argument\
                to transpile() call.")

T
typhoonzero 已提交
720
        s_prog = Program()
W
Wu Yi 已提交
721
        orig_s_prog = self.startup_program
X
Xin Pan 已提交
722
        s_prog.random_seed = orig_s_prog.random_seed
T
typhoonzero 已提交
723 724 725 726 727 728 729 730 731 732 733
        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
734
        created_var_map = collections.OrderedDict()
M
minqiyang 已提交
735
        for _, var in six.iteritems(pserver_vars):
W
Wu Yi 已提交
736
            tmpvar = s_prog.global_block()._clone_variable(var)
T
typhoonzero 已提交
737 738 739 740
            created_var_map[var.name] = tmpvar

        # 2. rename op outputs
        for op in orig_s_prog.global_block().ops:
741
            new_outputs = collections.OrderedDict()
T
typhoonzero 已提交
742 743
            # do not append startup op if var is not on this pserver
            op_on_pserver = False
G
gongweibao 已提交
744 745 746 747 748 749 750 751 752 753
            # 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
754 755

            if op_on_pserver:
756 757 758
                # most startup program ops have no inputs
                new_inputs = self._get_input_map_from_op(pserver_vars, op)

T
typhoonzero 已提交
759 760 761
                if op.type in [
                        "gaussian_random", "fill_constant", "uniform_random"
                ]:
G
gongweibao 已提交
762
                    op.set_attr("shape", list(new_outputs["Out"].shape))
T
typhoonzero 已提交
763 764 765 766
                s_prog.global_block().append_op(
                    type=op.type,
                    inputs=new_inputs,
                    outputs=new_outputs,
G
gongweibao 已提交
767
                    attrs=op.all_attrs())
T
typhoonzero 已提交
768 769
        return s_prog

770 771
    # ====================== private transpiler functions =====================

Y
yi.wu 已提交
772 773 774 775 776 777 778 779 780
    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
781
                if op.attr('is_distributed') is True:
Y
yi.wu 已提交
782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832
                    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
833
    def _init_splited_vars(self):
Y
yi.wu 已提交
834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856
        # 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
857
        if self.config.slice_var_up:
Y
yi.wu 已提交
858 859
            # 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
860 861 862
            grad_blocks = slice_variable(grad_list,
                                         len(self.pserver_endpoints),
                                         self.config.min_block_size)
Y
yi.wu 已提交
863
            param_blocks = slice_variable(param_list,
G
gongweibao 已提交
864 865
                                          len(self.pserver_endpoints),
                                          self.config.min_block_size)
Y
yi.wu 已提交
866 867 868
        else:
            # when we do NOT slice var up into blocks, we will always slice params
            # grads into one block.
G
gongweibao 已提交
869 870 871 872
            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
873 874
        assert (len(grad_blocks) == len(param_blocks))

875
        # origin_param_name -> [splited_param_vars]
Y
yi.wu 已提交
876 877
        self.param_var_mapping = self._create_vars_from_blocklist(
            self.origin_program, param_blocks)
878
        # origin_grad_name -> [splited_grad_vars]
Y
yi.wu 已提交
879 880 881 882
        self.grad_var_mapping = self._create_vars_from_blocklist(
            self.origin_program,
            grad_blocks,
            add_trainer_suffix=self.trainer_num > 1)
883
        # dict(grad_splited_var -> param_splited_var)
884
        self.grad_param_mapping = collections.OrderedDict()
Y
yi.wu 已提交
885 886 887 888
        for g, p in zip(grad_blocks, param_blocks):
            g_name, g_bid, _ = g.split(":")
            p_name, p_bid, _ = p.split(":")
            self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] =  \
889
                self.param_var_mapping[p_name][int(p_bid)]
Y
yi.wu 已提交
890 891

        # create mapping of endpoint -> split var to create pserver side program
892
        self.param_grad_ep_mapping = collections.OrderedDict()
Y
yi.wu 已提交
893 894 895 896 897 898 899 900 901
        [
            self.param_grad_ep_mapping.update({
                ep: {
                    "params": [],
                    "grads": []
                }
            }) for ep in self.pserver_endpoints
        ]

902
    # transpiler function for dis lookup_table
Q
update  
qiaolongfei 已提交
903 904
    def _replace_lookup_table_op_with_prefetch(self, program,
                                               pserver_endpoints):
905
        # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
Q
qiaolongfei 已提交
906 907 908 909 910 911 912 913 914
        # self.all_prefetch_input_vars =
        #       [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
        #        [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
        self.all_prefetch_input_vars = []

        # self.all_prefetch_input_vars =
        #       [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
        #        [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
        self.all_prefetch_output_vars = []
915 916 917 918 919 920 921 922 923

        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

924
                    lookup_table_op_index = list(all_ops).index(op)
925 926 927
                    ids_name = op.input("Ids")
                    out_name = op.output("Out")

Q
qiaolongfei 已提交
928
                    ids_var = program.global_block().vars[ids_name[0]]
W
Wu Yi 已提交
929
                    prefetch_input_vars = self._create_splited_vars(
Q
qiaolongfei 已提交
930 931 932 933 934 935
                        source_var=ids_var,
                        block=program.global_block(),
                        tag="_prefetch_in_")
                    self.all_prefetch_input_vars.append(prefetch_input_vars)

                    out_var = program.global_block().vars[out_name[0]]
W
Wu Yi 已提交
936
                    prefetch_output_vars = self._create_splited_vars(
Q
qiaolongfei 已提交
937 938 939 940
                        source_var=out_var,
                        block=program.global_block(),
                        tag="_prefetch_out_")
                    self.all_prefetch_output_vars.append(prefetch_output_vars)
941 942

                    # insert split_ids_op
W
Wu Yi 已提交
943
                    program.global_block()._insert_op(
944
                        index=lookup_table_op_index,
945 946 947 948 949 950 951
                        type="split_ids",
                        inputs={
                            'Ids': [
                                program.global_block().vars[varname]
                                for varname in ids_name
                            ]
                        },
Q
qiaolongfei 已提交
952
                        outputs={"Out": prefetch_input_vars})
953 954

                    # insert prefetch_op
W
Wu Yi 已提交
955
                    program.global_block()._insert_op(
956
                        index=lookup_table_op_index + 1,
957
                        type="prefetch",
Q
qiaolongfei 已提交
958 959
                        inputs={'X': prefetch_input_vars},
                        outputs={"Out": prefetch_output_vars},
Y
Yancey1989 已提交
960
                        attrs={
961
                            "epmap": pserver_endpoints,
962 963 964
                            # 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
Y
Yancey1989 已提交
965
                        })
966 967

                    # insert concat_op
W
Wu Yi 已提交
968
                    program.global_block()._insert_op(
969 970 971 972 973 974 975
                        index=lookup_table_op_index + 2,
                        type="merge_ids",
                        inputs={
                            'Ids': [
                                program.global_block().vars[varname]
                                for varname in ids_name
                            ],
976
                            'X': prefetch_output_vars
977
                        },
978 979 980 981 982
                        outputs={
                            "Out": [
                                program.global_block().vars[varname]
                                for varname in out_name
                            ]
983
                        })
984 985

                    # delete lookup_table_op
986
                    delete_ops(program.global_block(), [op])
987 988 989
                    # break for loop
                    break

Y
Yancey1989 已提交
990
    def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
991
        # 2. add split_ids_op and send_op to send gradient to pservers
992 993
        # there should only be one table_name
        all_ops = program.global_block().ops
T
typhoonzero 已提交
994
        table_grad_name = grad_var_name(self.table_name)
995 996 997 998
        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
999
                program.global_block()._insert_op(
1000 1001 1002 1003 1004
                    index=op_index + 1,
                    type="split_ids",
                    inputs={
                        'Ids': [program.global_block().vars[table_grad_name]]
                    },
1005
                    outputs={"Out": self.trainer_side_table_grad_list})
W
Wu Yi 已提交
1006
                program.global_block()._insert_op(
1007
                    index=op_index + 2,
1008
                    type="send",
1009
                    inputs={'X': self.trainer_side_table_grad_list},
1010
                    outputs={'Out': []},
Y
Yancey1989 已提交
1011
                    attrs={
1012
                        "sync_mode": True,
Y
Yancey1989 已提交
1013
                        "epmap": pserver_endpoints,
W
Wu Yi 已提交
1014 1015 1016 1017 1018
                        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
1019
                    })
1020 1021 1022 1023 1024 1025
                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
1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053
        prefetch_var_name_to_block_id = []
        for index in range(len(self.all_prefetch_input_vars)):
            prefetch_block = pserver_program.create_block(optimize_block.idx)
            trainer_ids = self.all_prefetch_input_vars[index][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[index][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))
        return prefetch_var_name_to_block_id
1054 1055

    def _create_table_optimize_block(self, pserver_index, pserver_program,
1056
                                     pre_block_idx, grad_to_block_id):
1057 1058
        # STEP: create table optimize block
        # create table param and grad var in pserver program
Y
Yancey1989 已提交
1059 1060
        origin_param_var = self.origin_program.global_block().vars[
            self.table_name]
T
tangwei12 已提交
1061

T
tangwei12 已提交
1062
        zero_dim = int(
T
tangwei12 已提交
1063 1064 1065 1066
            math.ceil(origin_param_var.shape[0] / len(self.pserver_endpoints)))
        table_shape = list(origin_param_var.shape)
        table_shape[0] = zero_dim

Y
Yancey1989 已提交
1067 1068
        param_var = pserver_program.global_block().create_var(
            name=origin_param_var.name,
T
tangwei12 已提交
1069
            shape=table_shape,
Y
Yancey1989 已提交
1070 1071 1072
            dtype=origin_param_var.dtype,
            type=core.VarDesc.VarType.SELECTED_ROWS,
            persistable=True)
1073 1074
        # parameter must be selected rows
        param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS)
W
Wu Yi 已提交
1075
        grad_var = pserver_program.global_block()._clone_variable(
T
typhoonzero 已提交
1076
            self.origin_program.global_block().vars[grad_var_name(
1077
                self.table_name)])
1078 1079 1080 1081

        # create table optimize block in pserver program
        table_opt_op = [
            op for op in self.optimize_ops
1082 1083
            if 'Param' in op.input_names and op.input("Param")[0] ==
            self.table_name
1084
        ][0]
Q
qiaolongfei 已提交
1085
        table_opt_block = pserver_program.create_block(pre_block_idx)
1086

1087 1088 1089
        if self.sync_mode:
            # create grad vars in pserver program
            table_grad_var = self.table_param_grad[1]
1090
            pserver_side_table_grad_list = [
1091 1092 1093 1094 1095 1096 1097 1098 1099
                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)
            ]

1100
            # append sum op for pserver_side_table_grad_list
1101 1102
            table_opt_block.append_op(
                type="sum",
1103
                inputs={"X": pserver_side_table_grad_list},
1104 1105
                outputs={"Out": [grad_var]},
                attrs={"use_mkldnn": False})
1106 1107
        else:
            # in async_mode, for table gradient, it also need to be splited to each parameter server
1108
            origin_grad_name = grad_var.name
1109 1110
            splited_grad_name = self.trainer_side_table_grad_list[
                pserver_index].name
1111 1112
            if not splited_grad_name.startswith(origin_grad_name):
                raise ValueError("origin_grad_var: " + splited_grad_name +
1113
                                 " grad_var:" + grad_var.name)
W
Wu Yi 已提交
1114
            grad_var = pserver_program.global_block()._rename_var(
1115
                origin_grad_name, splited_grad_name)
1116 1117 1118 1119 1120 1121 1122 1123 1124

        lr_var = pserver_program.global_block().vars[table_opt_op.input(
            "LearningRate")[0]]
        inputs = {
            "Param": [param_var],
            "Grad": [grad_var],
            "LearningRate": [lr_var]
        }
        outputs = {"ParamOut": [param_var]}
1125
        # only support sgd now
1126 1127 1128 1129
        import logging
        logging.warn(
            "distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of "
            + table_opt_op.type)
1130
        table_opt_block.append_op(type="sgd", inputs=inputs, outputs=outputs)
1131

1132 1133 1134
        # 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))

1135 1136
        return table_opt_block

T
tangwei12 已提交
1137 1138 1139 1140 1141 1142
    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
1143
        pserver_program.global_block().create_var(
T
tangwei12 已提交
1144
            name="kLookupTablePath",
T
tangwei12 已提交
1145 1146
            persistable=True,
            type=core.VarDesc.VarType.RAW)
T
tangwei12 已提交
1147

T
tangwei12 已提交
1148
        checkpoint_save_block = pserver_program.create_block(pre_block_idx)
T
tangwei12 已提交
1149
        # this 'file_path' do not be used in save lookup table variable
T
tangwei12 已提交
1150 1151 1152 1153
        checkpoint_save_block.append_op(
            type='save',
            inputs={'X': [self.table_name]},
            outputs={},
T
tangwei12 已提交
1154
            attrs={'file_path': "none"})
T
tangwei12 已提交
1155 1156 1157

        return checkpoint_save_block.idx

T
typhoonzero 已提交
1158 1159 1160 1161 1162
    def _create_vars_from_blocklist(self,
                                    program,
                                    block_list,
                                    add_trainer_suffix=False):
        """
1163
        Create vars for each split.
T
typhoonzero 已提交
1164 1165
        NOTE: only grads need to be named for different trainers, use
              add_trainer_suffix to rename the grad vars.
1166 1167 1168 1169
        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.
1170
        Returns:
1171
            var_mapping (collections.OrderedDict(varname->[new_varname_variable])):A dict mapping
1172
                from original var name to each var split.
T
typhoonzero 已提交
1173
        """
1174 1175

        # varname->[(block_id, current_block_size)]
1176
        block_map = collections.OrderedDict()
1177

1178
        var_mapping = collections.OrderedDict()
T
typhoonzero 已提交
1179 1180
        for block_str in block_list:
            varname, offset, size = block_str.split(":")
1181
            if varname not in block_map:
T
typhoonzero 已提交
1182
                block_map[varname] = []
1183
            block_map[varname].append((int(offset), int(size)))
Y
yi.wu 已提交
1184

M
minqiyang 已提交
1185
        for varname, splited in six.iteritems(block_map):
T
typhoonzero 已提交
1186
            orig_var = program.global_block().var(varname)
T
typhoonzero 已提交
1187
            if len(splited) == 1:
1188
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1189 1190
                    new_var_name = "%s.trainer_%d" % \
                        (orig_var.name, self.trainer_id)
W
Wu Yi 已提交
1191
                    program.global_block()._rename_var(varname, new_var_name)
T
typhoonzero 已提交
1192 1193 1194 1195 1196
                    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
1197
                continue
T
typhoonzero 已提交
1198
            var_mapping[varname] = []
T
typhoonzero 已提交
1199 1200 1201 1202
            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
1203

T
typhoonzero 已提交
1204
            for i, block in enumerate(splited):
T
typhoonzero 已提交
1205
                size = block[1]
M
minqiyang 已提交
1206
                rows = size // orig_dim1_flatten
T
typhoonzero 已提交
1207 1208 1209
                splited_shape = [rows]
                if len(orig_shape) >= 2:
                    splited_shape.extend(orig_shape[1:])
T
typhoonzero 已提交
1210
                new_var_name = ""
1211
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1212 1213 1214 1215 1216
                    new_var_name = "%s.block%d.trainer_%d" % \
                        (varname, i, self.trainer_id)
                else:
                    new_var_name = "%s.block%d" % \
                        (varname, i)
T
typhoonzero 已提交
1217
                var = program.global_block().create_var(
T
typhoonzero 已提交
1218 1219
                    name=new_var_name,
                    persistable=False,
T
typhoonzero 已提交
1220
                    dtype=orig_var.dtype,
1221
                    type=orig_var.type,
T
typhoonzero 已提交
1222
                    shape=splited_shape)  # flattend splited var
T
typhoonzero 已提交
1223
                var_mapping[varname].append(var)
W
Wu Yi 已提交
1224
            program.global_block()._sync_with_cpp()
T
typhoonzero 已提交
1225
        return var_mapping
T
done  
typhoonzero 已提交
1226

W
Wu Yi 已提交
1227
    def _create_splited_vars(self, source_var, block, tag):
1228 1229 1230 1231 1232 1233 1234 1235 1236 1237
        return [
            block.create_var(
                name=str(source_var.name + tag + str(index)),
                type=source_var.type,
                shape=source_var.shape,
                dtype=source_var.dtype)
            for index in range(len(self.pserver_endpoints))
        ]

    def _clone_var(self, block, var, persistable=True):
T
done  
typhoonzero 已提交
1238 1239 1240 1241 1242 1243
        return block.create_var(
            name=var.name,
            shape=var.shape,
            dtype=var.dtype,
            type=var.type,
            lod_level=var.lod_level,
1244
            persistable=persistable)
T
done  
typhoonzero 已提交
1245

Y
Yancey1989 已提交
1246
    def _insert_split_op(self, program, orig_var, index, splited_vars):
Y
update  
Yancey1989 已提交
1247 1248 1249 1250
        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
1251
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1252 1253 1254 1255 1256 1257 1258 1259 1260
                index=index + 1,
                type="split_selected_rows",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
                attrs={"height_sections": height_sections})
        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
1261
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1262 1263 1264 1265 1266 1267 1268 1269 1270
                index=index + 1,
                type="split_byref",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
                attrs={"sections": sections}  # assume split evenly
            )
        else:
            AssertionError("Variable type should be in set "
                           "[LOD_TENSOR, SELECTED_ROWS]")
T
done  
typhoonzero 已提交
1271

T
typhoonzero 已提交
1272 1273 1274 1275
    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
1276
        Param and Grad is split to multiple servers.
T
typhoonzero 已提交
1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
        """
        # 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
        elif op_type == "momentum":
            if varkey == "Velocity":
                return param_shape
        elif op_type == "":
            if varkey == "Moment":
                return param_shape
        elif op_type == "sgd":
            pass
        return orig_shape

1299 1300
    def _get_varname_parts(self, varname):
        # returns origin, blockid, trainerid
T
typhoonzero 已提交
1301
        orig_var_name = ""
1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
        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
1312
        else:
1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339
            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
            return
        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
1340
        else:
1341 1342 1343 1344 1345 1346
            merged_var_name = orig_varname
        merged_var = \
            pserver_block.vars[merged_var_name]
        grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx))
        if self.sync_mode and self.trainer_num > 1:
            vars2merge = []
1347
            for i in range(self.trainer_num):
1348 1349 1350 1351 1352 1353 1354
                per_trainer_name = "%s.trainer_%d" % \
                (merged_var_name, i)
                vars2merge.append(pserver_block.vars[per_trainer_name])

            optimize_block.append_op(
                type="sum",
                inputs={"X": vars2merge},
1355 1356
                outputs={"Out": merged_var},
                attrs={"use_mkldnn": False})
Q
qiaolongfei 已提交
1357 1358 1359 1360 1361
            optimize_block.append_op(
                type="scale",
                inputs={"X": merged_var},
                outputs={"Out": merged_var},
                attrs={"scale": 1.0 / float(self.trainer_num)})
1362
        return merged_var
T
typhoonzero 已提交
1363

1364
    def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
1365
                            grad_to_block_id, origin_program, merged_var):
1366
        program = optimize_block.program
T
typhoonzero 已提交
1367
        pserver_block = program.global_block()
1368
        new_inputs = collections.OrderedDict()
W
Wu Yi 已提交
1369

T
typhoonzero 已提交
1370 1371
        # update param/grad shape first, then other inputs like
        # moment can use the updated shape
W
Wu Yi 已提交
1372 1373 1374 1375 1376 1377 1378 1379 1380
        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
1381
        for key in opt_op.input_names:
T
typhoonzero 已提交
1382 1383
            if key == "Grad":
                new_inputs[key] = merged_var
W
Wu Yi 已提交
1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399
            # For RMSProp optimizer
            elif key == "Moment" or key == "MeanSquare":
                param_block = _get_param_block(opt_op)
                if not param_block:
                    return
                moment_var = origin_program.global_block().vars[opt_op.input(
                    key)[0]]
                tmpvar = pserver_block.create_var(
                    name=moment_var.name,
                    persistable=moment_var.persistable,
                    dtype=moment_var.dtype,
                    # change to use same shape as param
                    # TODO(typhoonzero): didn't append .block in the var name,
                    # may affect checkpoint saving? Need to verify.
                    shape=param_block.shape)
                new_inputs[key] = tmpvar
T
typhoonzero 已提交
1400
            elif key == "Param":
W
Wu Yi 已提交
1401
                param_block = _get_param_block(opt_op)
T
typhoonzero 已提交
1402 1403
                if not param_block:
                    return
T
typhoonzero 已提交
1404
                tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1405
                    name=param_block.name,
T
typhoonzero 已提交
1406
                    persistable=True,
T
typhoonzero 已提交
1407 1408 1409
                    dtype=param_block.dtype,
                    shape=param_block.shape)
                new_inputs[key] = tmpvar
1410
            elif key == "LearningRate":
1411
                # learning rate variable has already be created by non-optimize op,
1412
                # don't create it once again.
1413
                lr_varname = opt_op.input(key)[0]
1414
                if lr_varname in pserver_block.vars:
1415 1416 1417 1418 1419 1420 1421 1422 1423
                    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
1424

T
typhoonzero 已提交
1425
        for key in opt_op.input_names:
1426
            new_shape = None
W
Wu Yi 已提交
1427
            if key in ["Param", "Grad", "LearningRate", "Moment", "MeanSquare"]:
T
typhoonzero 已提交
1428
                continue
1429
            var = self.origin_program.global_block().vars[opt_op.input(key)[0]]
T
typhoonzero 已提交
1430 1431 1432 1433
            # 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
1434
            tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1435 1436 1437 1438 1439
                name=var.name,
                persistable=var.persistable,
                dtype=var.dtype,
                shape=new_shape)
            new_inputs[key] = tmpvar
T
typhoonzero 已提交
1440

1441
        # change output's ParamOut variable
1442 1443
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1444
        outputs["ParamOut"] = new_inputs["Param"]
T
typhoonzero 已提交
1445

1446
        optimize_block.append_op(
T
typhoonzero 已提交
1447 1448
            type=opt_op.type,
            inputs=new_inputs,
T
typhoonzero 已提交
1449
            outputs=outputs,
G
gongweibao 已提交
1450
            attrs=opt_op.all_attrs())
T
typhoonzero 已提交
1451

1452 1453
    def _is_splited_grad_var(self, var, var_dict):
        grad_block = None
M
minqiyang 已提交
1454
        for _, g in six.iteritems(var_dict):
1455 1456 1457 1458 1459 1460
            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
1461 1462 1463
    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
1464
        for key, varlist in six.iteritems(inputs):
Q
Qiyang Min 已提交
1465 1466 1467 1468
            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
1469
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1470 1471 1472

        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, op)
M
minqiyang 已提交
1473
        for key, varlist in six.iteritems(outputs):
Q
Qiyang Min 已提交
1474 1475 1476 1477
            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
1478
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1479

Y
Yancey1989 已提交
1480
        return block.append_op(
G
gongweibao 已提交
1481
            type=op.type, inputs=inputs, outputs=outputs, attrs=op.all_attrs())
Q
Qiyang Min 已提交
1482 1483

    def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
1484
        program = optimize_block.program
1485
        # Append the ops for parameters that do not need to be optimized/updated
1486 1487
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, opt_op)
M
minqiyang 已提交
1488
        for key, varlist in six.iteritems(inputs):
1489 1490
            if not isinstance(varlist, list):
                varlist = [varlist]
T
typhoonzero 已提交
1491
            for var in varlist:
1492 1493 1494 1495 1496 1497
                # 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
1498
                elif var.name not in program.global_block().vars:
1499
                    program.global_block().create_var(
T
typhoonzero 已提交
1500 1501 1502 1503 1504
                        name=var.name,
                        persistable=var.persistable,
                        dtype=var.dtype,
                        shape=var.shape)

1505 1506
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
M
minqiyang 已提交
1507
        for key, varlist in six.iteritems(outputs):
1508 1509 1510
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
1511 1512 1513 1514
                grad_block = self._is_splited_grad_var(
                    var, program.global_block().vars)
                if grad_block:
                    outputs[key] = grad_block
1515
                elif var.name not in program.global_block().vars:
W
Wu Yi 已提交
1516
                    program.global_block()._clone_variable(var)
1517

Y
Yancey1989 已提交
1518
        return optimize_block.append_op(
T
typhoonzero 已提交
1519
            type=opt_op.type,
T
typhoonzero 已提交
1520 1521
            inputs=inputs,
            outputs=outputs,
G
gongweibao 已提交
1522
            attrs=opt_op.all_attrs())
T
typhoonzero 已提交
1523

1524 1525 1526 1527
    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
1528 1529
        if set(op1.desc.output_arg_names()) & set(op2.desc.input_arg_names()) or \
           set(op1.desc.input_arg_names()) & set(op2.desc.output_arg_names()):
1530 1531 1532 1533 1534 1535
            return True
        return False

    def _create_ufind(self, optimize_ops):
        # Create a unit find data struct by optimize ops
        ufind = UnionFind(optimize_ops)
1536 1537
        for i in range(len(optimize_ops)):
            for j in range(i, len(optimize_ops)):
1538 1539 1540 1541 1542 1543
                op1 = optimize_ops[i]
                op2 = optimize_ops[j]
                if self._is_op_connected(op1, op2):
                    ufind.union(op1, op2)
        return ufind

1544
    def _is_optimizer_op(self, op):
T
typhoonzero 已提交
1545 1546
        if "Param" in op.input_names and \
            "LearningRate" in op.input_names:
1547 1548 1549 1550 1551 1552 1553
            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
1554
        if op.input("Param")[0] in param_names:
1555 1556 1557
            return True
        else:
            for n in param_names:
T
typhoonzero 已提交
1558
                param = op.input("Param")[0]
T
typhoonzero 已提交
1559
                if same_or_split_var(n, param) and n != param:
1560 1561 1562
                    return True
            return False

T
typhoonzero 已提交
1563
    def _get_input_map_from_op(self, varmap, op):
1564
        """Returns a dict from op input name to the vars in varmap."""
1565
        iomap = collections.OrderedDict()
T
typhoonzero 已提交
1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576
        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):
1577
        """Returns a dict from op output name to the vars in varmap."""
1578
        iomap = collections.OrderedDict()
T
typhoonzero 已提交
1579 1580 1581 1582 1583 1584 1585 1586 1587
        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
1588 1589 1590 1591 1592 1593

    def _get_lr_ops(self):
        lr_ops = []
        # find learning rate variables by optimize op
        lr_vars = set()
        for op in self.optimize_ops:
1594
            if self._is_optimizer_op(op):
1595 1596 1597 1598
                lr_vars.add(op.input("LearningRate")[0])

        find_ops = []
        # find ops which output is lr var
1599
        block = self.origin_program.global_block()
1600 1601 1602 1603 1604
        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)
1605

1606 1607 1608 1609 1610
        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 \
1611
                    not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2):
1612 1613 1614 1615 1616 1617
                    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)
1618 1619
                    # we only need to append op for once
                    break
1620
        return lr_ops
Y
Yancey1989 已提交
1621

W
Wu Yi 已提交
1622 1623 1624 1625 1626
    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
1627 1628
        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
1629 1630 1631
            return True
        return False

Y
Yancey1989 已提交
1632
    def _get_optimize_pass(self):
1633
        """
1634
        Get optimizer operators, parameters and gradients from origin_program
1635 1636 1637 1638
        Returns:
            opt_ops (list): optimize operators.
            params_grads (dict): paramter->gradient.
        """
Y
Yancey1989 已提交
1639 1640 1641
        block = self.origin_program.global_block()
        opt_ops = []
        params_grads = []
1642
        origin_var_dict = self.origin_program.global_block().vars
Y
Yancey1989 已提交
1643
        for op in block.ops:
W
Wu Yi 已提交
1644
            if self._is_opt_role_op(op):
Y
Yancey1989 已提交
1645
                opt_ops.append(op)
1646 1647 1648 1649 1650
                # HACK(wuyi): if we find grad vars from input of optimize
                # ops, we may get the output of clip op. Use syntax "@GRAD"
                # and op_role_var to get the pair.
                for input_name in op.input_arg_names:
                    if input_name.find("@GRAD") != -1 and \
G
gongweibao 已提交
1651 1652
                        op.attr(RPC_OP_ROLE_ATTR_NAME):
                        param_name = op.attr(OP_ROLE_VAR_ATTR_NAME)[0]
1653 1654 1655 1656
                        params_grads.append([
                            origin_var_dict[param_name],
                            origin_var_dict[input_name]
                        ])
Y
Yancey1989 已提交
1657 1658 1659
            else:
                pass
        return opt_ops, params_grads