distribute_transpiler.py 81.3 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
34
import numpy as np
35
import collections
Q
Qiao Longfei 已提交
36
import logging
37

38
from .ps_dispatcher import RoundRobin, PSDispatcher
W
Wu Yi 已提交
39
from .. import core, framework, unique_name
T
typhoonzero 已提交
40
from ..framework import Program, default_main_program, \
T
tangwei12 已提交
41 42
    default_startup_program, Block, \
    Parameter, grad_var_name
43
from .details import *
Q
Qiao Longfei 已提交
44
from ..distribute_lookup_table import find_distributed_lookup_table
45
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
    enable_dc_asgd = False
W
Wu Yi 已提交
141 142
    # supported modes: pserver, nccl2
    mode = "pserver"
143
    print_log = False
G
gongweibao 已提交
144 145


Y
gen rst  
yi.wu 已提交
146
class DistributeTranspiler(object):
Y
yi.wu 已提交
147 148 149 150
    """
    **DistributeTranspiler**

    Convert the fluid program to distributed data-parallelism programs.
W
Wu Yi 已提交
151
    Supports two modes: pserver mode and nccl2 mode.
Y
yi.wu 已提交
152

W
Wu Yi 已提交
153 154 155 156 157 158 159 160 161
    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
162 163 164 165

    Examples:
        .. code-block:: python

W
Wu Yi 已提交
166 167 168 169 170 171
           # 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
172 173
           role = os.getenv("PADDLE_TRAINING_ROLE")

W
Wu Yi 已提交
174
           t = fluid.DistributeTranspiler()
Y
yi.wu 已提交
175 176 177 178 179 180 181 182
           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
183

W
Wu Yi 已提交
184 185 186 187 188 189 190 191 192 193 194
           # 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
195
    """
Y
Yancey1989 已提交
196

G
gongweibao 已提交
197 198 199 200 201 202 203 204 205
    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

206 207 208
        global PRINT_LOG
        if self.config.print_log:
            PRINT_LOG = True
G
gongweibao 已提交
209 210 211
        assert (self.config.min_block_size >= 8192)
        assert (self.config.split_method.__bases__[0] == PSDispatcher)

W
Wu Yi 已提交
212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238
    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")

Q
Qiao Longfei 已提交
239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
    def _get_all_sparse_update_op(self, main_program):
        sparse_update_ops = []
        sparse_update_op_types = ["lookup_table"]
        for op in main_program.global_block().ops:
            if op.type in sparse_update_op_types and op.attr(
                    'is_sparse') is True and not op.attr('is_distributed'):
                sparse_update_ops.append(op)
        return sparse_update_ops

    def _update_sparse_update_op(self, param_varname, height_sections,
                                 endpint_map):
        for op in self.sparse_update_ops:
            if param_varname in op.input_arg_names:
                op._set_attr('epmap', endpint_map)
                op._set_attr('height_sections', height_sections)

255 256 257 258 259
    def transpile(self,
                  trainer_id,
                  program=None,
                  pservers="127.0.0.1:6174",
                  trainers=1,
W
Wu Yi 已提交
260
                  sync_mode=True,
W
Wu Yi 已提交
261 262
                  startup_program=None,
                  current_endpoint="127.0.0.1:6174"):
263
        """
Y
yi.wu 已提交
264 265 266 267 268 269 270
        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().
W
Wu Yi 已提交
271 272
            startup_program (Program|None): startup_program to transpile,
                default is fluid.default_startup_program().
Y
yi.wu 已提交
273 274
            pservers (str): comma separated ip:port string for the pserver
                list.
W
Wu Yi 已提交
275 276 277
            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
278
            sync_mode (bool): Do sync training or not, default is True.
W
Wu Yi 已提交
279 280
            startup_program (Program|None): startup_program to transpile,
                default is fluid.default_main_program().
W
Wu Yi 已提交
281 282 283
            current_endpoint (str): need pass current endpoint when
                transpile as nccl2 distributed mode. In pserver mode
                this argument is not used.
284 285 286
        """
        if program is None:
            program = default_main_program()
W
Wu Yi 已提交
287 288
        if startup_program is None:
            startup_program = default_startup_program()
289
        self.origin_program = program
W
Wu Yi 已提交
290 291
        self.startup_program = startup_program
        self.origin_startup_program = self.startup_program.clone()
G
gongweibao 已提交
292

W
Wu Yi 已提交
293 294 295 296 297 298 299 300 301
        if self.config.mode == "nccl2":
            assert (isinstance(trainers, str))
            self._transpile_nccl2(
                trainer_id,
                trainers,
                current_endpoint,
                startup_program=startup_program)
            return

302 303 304 305 306 307 308
        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
309
        ps_dispatcher = self.config.split_method(self.pserver_endpoints)
310 311
        self.table_name = find_distributed_lookup_table(self.origin_program)
        self.has_distributed_lookup_table = self.table_name != None
312
        self.param_name_to_grad_name = dict()
W
Wu Yi 已提交
313
        self.grad_name_to_param_name = dict()
314 315
        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
316
            self.grad_name_to_param_name[grad_var.name] = param_var.name
317

Q
Qiao Longfei 已提交
318 319 320 321 322
        # get all sparse update ops
        self.sparse_update_ops = self._get_all_sparse_update_op(
            self.origin_program)
        self.sparse_param_to_height_sections = dict()

T
tangwei12 已提交
323 324 325 326 327 328
        # 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

329
        # split and create vars, then put splited vars in dicts for later use.
G
gongweibao 已提交
330
        # step 1: split and create vars, then put splited vars in dicts for later use.
G
gongweibao 已提交
331
        self._init_splited_vars()
332

G
gongweibao 已提交
333
        # step 2: insert send op to send gradient vars to parameter servers
Y
Yancey1989 已提交
334
        ps_dispatcher.reset()
Y
update  
Yancey1989 已提交
335
        send_vars = []
336 337 338 339 340 341

        # 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
342
        grad_var_mapping_items = list(six.iteritems(self.grad_var_mapping))
343

G
gongweibao 已提交
344
        if not self.config.slice_var_up:
345 346
            np.random.seed(self.origin_program.random_seed)
            np.random.shuffle(grad_var_mapping_items)
347

348
        self.grad_name_to_send_dummy_out = dict()
349
        for grad_varname, splited_vars in grad_var_mapping_items:
Y
update  
Yancey1989 已提交
350
            eplist = ps_dispatcher.dispatch(splited_vars)
351

G
gongweibao 已提交
352
            if not self.config.slice_var_up:
353 354
                assert (len(splited_vars) == 1)

355
            splited_grad_varname = grad_varname
Y
Yancey1989 已提交
356
            if len(splited_vars) == 1:
357
                splited_grad_varname = splited_vars[0].name
358 359
                index = find_op_by_output_arg(
                    program.global_block(), splited_grad_varname, reverse=True)
Y
Yancey1989 已提交
360
            elif len(splited_vars) > 1:
361
                orig_var = program.global_block().vars[splited_grad_varname]
362 363
                index = find_op_by_output_arg(
                    program.global_block(), splited_grad_varname, reverse=True)
Y
Yancey1989 已提交
364
                self._insert_split_op(program, orig_var, index, splited_vars)
Y
update  
Yancey1989 已提交
365
                index += 1
Y
Yancey1989 已提交
366 367
            else:
                AssertionError("Can not insert the send op by original "
368
                               "variable name :", splited_grad_varname)
Y
Yancey1989 已提交
369

W
Wu Yi 已提交
370 371
            dummy_output = program.global_block().create_var(
                name=framework.generate_control_dev_var_name())
372
            self.grad_name_to_send_dummy_out[grad_varname] = dummy_output
W
Wu Yi 已提交
373

W
Wu Yi 已提交
374 375 376 377
            # 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
378
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
379
                index=index + 1,
380
                type="send",
Y
update  
Yancey1989 已提交
381
                inputs={"X": splited_vars},
382
                outputs={"Out": dummy_output},
Y
Yancey1989 已提交
383 384
                attrs={
                    "epmap": eplist,
385
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
W
Wu Yi 已提交
386 387 388 389
                    OP_ROLE_VAR_ATTR_NAME: [
                        self.grad_name_to_param_name[grad_varname],
                        splited_grad_varname
                    ],
390
                    "sync_mode": not self.sync_mode,
Y
Yancey1989 已提交
391
                })
Y
update  
Yancey1989 已提交
392 393
            for _, var in enumerate(splited_vars):
                send_vars.append(var)
Y
Yancey1989 已提交
394 395

        if self.sync_mode:
W
Wu Yi 已提交
396 397
            send_barrier_out = program.global_block().create_var(
                name=framework.generate_control_dev_var_name())
398 399 400 401
            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())
402
            input_deps = list(self.grad_name_to_send_dummy_out.values())
403

Y
Yancey1989 已提交
404 405
            program.global_block().append_op(
                type="send_barrier",
M
minqiyang 已提交
406
                inputs={"X": list(input_deps)},
W
Wu Yi 已提交
407
                outputs={"Out": send_barrier_out},
Y
Yancey1989 已提交
408 409
                attrs={
                    "endpoints": pserver_endpoints,
W
Wu Yi 已提交
410 411
                    "sync_mode": self.sync_mode,
                    "trainer_id": self.trainer_id,
Y
Yancey1989 已提交
412
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Y
Yancey1989 已提交
413
                })
Y
Yancey1989 已提交
414

G
gongweibao 已提交
415
        # step 3: insert recv op to receive parameters from parameter server
Y
Yancey1989 已提交
416
        recv_vars = []
Y
update  
Yancey1989 已提交
417
        for _, var in enumerate(send_vars):
418
            recv_vars.append(self.grad_param_mapping[var])
Y
update  
Yancey1989 已提交
419
        ps_dispatcher.reset()
Y
Yancey1989 已提交
420 421
        eplist = ps_dispatcher.dispatch(recv_vars)

T
typhoonzero 已提交
422
        for i, ep in enumerate(eplist):
Y
Yancey1989 已提交
423 424
            self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
            self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
425

Y
Yancey1989 已提交
426
        # step4: Concat the parameters splits together after recv.
W
Wu Yi 已提交
427
        all_recv_outputs = []
428
        for param_varname, splited_var in six.iteritems(self.param_var_mapping):
Y
Yancey1989 已提交
429 430 431 432
            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
433 434 435 436
            if self.sync_mode:
                recv_dep_in = send_barrier_out
            else:
                # connect deps to send op in async mode
437
                recv_dep_in = self.grad_name_to_send_dummy_out[
W
Wu Yi 已提交
438 439
                    self.param_name_to_grad_name[param_varname]]
            all_recv_outputs.extend(splited_var)
W
Wu Yi 已提交
440 441 442 443 444 445 446 447 448
            # 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

Q
Qiao Longfei 已提交
449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466
            if param_varname in self.sparse_param_to_height_sections:
                height_sections = self.sparse_param_to_height_sections[
                    param_varname]
                self._update_sparse_update_op(param_varname, height_sections,
                                              eps)
            else:
                program.global_block().append_op(
                    type="recv",
                    inputs={"X": [recv_dep_in]},
                    outputs={"Out": splited_var},
                    attrs={
                        "epmap": eps,
                        "trainer_id": self.trainer_id,
                        RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE,
                        OP_ROLE_VAR_ATTR_NAME:
                        [param_varname, recv_op_role_var_name],
                        "sync_mode": not self.sync_mode
                    })
T
typhoonzero 已提交
467

Q
qiaolongfei 已提交
468
        if self.sync_mode:
W
Wu Yi 已提交
469
            # form a WAW dependency
Q
qiaolongfei 已提交
470 471 472
            program.global_block().append_op(
                type="fetch_barrier",
                inputs={},
W
Wu Yi 已提交
473
                outputs={"Out": all_recv_outputs},
Q
qiaolongfei 已提交
474 475
                attrs={
                    "endpoints": pserver_endpoints,
W
Wu Yi 已提交
476
                    "trainer_id": self.trainer_id,
Q
qiaolongfei 已提交
477 478
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })
Y
Yancey1989 已提交
479

480
        for param_varname, splited_var in six.iteritems(self.param_var_mapping):
T
typhoonzero 已提交
481 482
            if len(splited_var) <= 1:
                continue
483
            orig_param = program.global_block().vars[param_varname]
Q
Qiao Longfei 已提交
484 485 486 487 488 489 490 491 492 493 494
            print("sparse_param_to_height_sections: " + str(
                self.sparse_param_to_height_sections))
            if param_varname not in self.sparse_param_to_height_sections:
                program.global_block().append_op(
                    type="concat",
                    inputs={"X": splited_var},
                    outputs={"Out": [orig_param]},
                    attrs={
                        "axis": 0,
                        RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE
                    })
T
typhoonzero 已提交
495

G
gongweibao 已提交
496 497
        self._get_trainer_startup_program(recv_vars=recv_vars, eplist=eplist)

498
        if self.has_distributed_lookup_table:
Q
update  
qiaolongfei 已提交
499 500
            self._replace_lookup_table_op_with_prefetch(program,
                                                        pserver_endpoints)
Y
Yancey1989 已提交
501
            self._split_table_grad_and_add_send_vars(program, pserver_endpoints)
502

W
Wu Yi 已提交
503
    def get_trainer_program(self, wait_port=True):
Y
yi.wu 已提交
504 505 506 507 508 509
        """
        Get transpiled trainer side program.

        Returns:
            Program: trainer side program.
        """
T
typhoonzero 已提交
510
        # remove optimize ops and add a send op to main_program
X
Xin Pan 已提交
511
        # FIXME(typhoonzero): Also ops like clip_gradient, lrn_decay?
T
typhoonzero 已提交
512
        lr_ops = self._get_lr_ops()
513
        delete_ops(self.origin_program.global_block(), self.optimize_ops)
T
typhoonzero 已提交
514 515
        delete_ops(self.origin_program.global_block(), lr_ops)

516 517
        # delete table init op
        if self.has_distributed_lookup_table:
518 519 520
            table_var = self.startup_program.global_block().vars[
                self.table_name]
            table_param_init_op = []
521 522
            for op in self.startup_program.global_block().ops:
                if self.table_name in op.output_arg_names:
523 524 525 526 527
                    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))
Q
Qiao Longfei 已提交
528
            table_init_op = table_param_init_op[0]
529 530 531 532 533 534
            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)
535

536
        self.origin_program.__str__()
G
gongweibao 已提交
537

W
Wu Yi 已提交
538 539 540
        if wait_port:
            wait_server_ready(self.pserver_endpoints)

541
        return self.origin_program
T
typhoonzero 已提交
542

W
Wu Yi 已提交
543
    def _get_trainer_startup_program(self, recv_vars, eplist):
G
gongweibao 已提交
544 545 546 547
        """
        Get transpiled trainer side startup program.

        Args:
W
Wu Yi 已提交
548
            recv_vars (list): Variable list to recv for current trainer_id
M
minqiyang 已提交
549
            eplist (list): A list of strings indicating
G
gongweibao 已提交
550 551 552 553

        Returns:
            Program: trainer side startup program.
        """
W
Wu Yi 已提交
554
        startup_program = self.startup_program
G
gongweibao 已提交
555 556 557 558

        # 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
559
        for varname, splited_var in six.iteritems(self.param_var_mapping):
G
gongweibao 已提交
560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579
            # 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",
580
                inputs={"X": []},
G
gongweibao 已提交
581 582 583 584 585 586
                outputs={"Out": splited_var},
                attrs={
                    "epmap": eps,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })

W
Wu Yi 已提交
587 588
        fetch_barrier_out = startup_program.global_block().create_var(
            name=framework.generate_control_dev_var_name())
G
gongweibao 已提交
589 590 591
        startup_program.global_block().append_op(
            type="fetch_barrier",
            inputs={},
W
Wu Yi 已提交
592
            outputs={"Out": fetch_barrier_out},
G
gongweibao 已提交
593 594 595 596 597
            attrs={
                "endpoints": self.pserver_endpoints,
                RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
            })

M
minqiyang 已提交
598
        for varname, splited_var in six.iteritems(self.param_var_mapping):
T
tangwei12 已提交
599
            # add concat ops to merge splited parameters received from parameter servers.
G
gongweibao 已提交
600 601
            if len(splited_var) <= 1:
                continue
W
Wu Yi 已提交
602
            # NOTE: if enable memory optimization, origin vars maybe removed.
M
minqiyang 已提交
603
            if varname in startup_program.global_block().vars:
W
Wu Yi 已提交
604 605 606 607 608 609 610 611 612 613
                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
614 615 616 617 618 619 620 621
            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
622 623
    def get_pserver_program(self, endpoint):
        """
Y
yi.wu 已提交
624
        Get parameter server side program.
625

Y
yi.wu 已提交
626 627
        Args:
            endpoint (str): current parameter server endpoint.
628

Y
yi.wu 已提交
629 630
        Returns:
            Program: the program for current parameter server to run.
T
typhoonzero 已提交
631
        """
Y
yi.wu 已提交
632 633 634 635
        # 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.
636 637 638
        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
639 640
        # step1
        pserver_program = Program()
X
Xin Pan 已提交
641
        pserver_program.random_seed = self.origin_program.random_seed
642
        # step2: Create vars to receive vars at parameter servers.
T
typhoonzero 已提交
643 644 645 646 647 648 649 650
        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
651 652 653 654 655
            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
656 657 658 659 660 661 662 663 664
            # 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)
665
            if self.sync_mode and self.trainer_num > 1:
666
                for trainer_id in range(self.trainer_num):
T
typhoonzero 已提交
667 668 669 670 671 672 673 674 675
                    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)
676

Q
qiaolongfei 已提交
677
        # step 3
678
        # Create a union-find data structure from optimize ops,
T
typhoonzero 已提交
679 680 681
        # 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
682
        # step 3.2
T
typhoonzero 已提交
683 684 685 686
        # Iterate through the ops and append optimize op which
        # located on current pserver
        opt_op_on_pserver = []
        for _, op in enumerate(self.optimize_ops):
687 688
            if self._is_optimizer_op(op) and self._is_opt_op_on_pserver(
                    endpoint, op):
T
typhoonzero 已提交
689
                opt_op_on_pserver.append(op)
Q
qiaolongfei 已提交
690
        # step 3.3
W
Wu Yi 已提交
691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708
        # prepare if dc asgd is enabled
        if self.config.enable_dc_asgd == True:
            assert (self.sync_mode == False)
            self.param_bak_list = []
            # add param_bak for each trainer
            for p in self.param_grad_ep_mapping[endpoint]["params"]:
                # each parameter should have w_bak for each trainer id
                for i in range(self.trainer_num):
                    param_bak_name = "%s.trainer_%d_bak" % (p.name, i)
                    tmpvar = pserver_program.global_block().create_var(
                        # NOTE: this var name format is used in `request_get_handler`
                        name=param_bak_name,
                        type=p.type,
                        shape=p.shape,
                        dtype=p.dtype)
                    self.param_bak_list.append((p, tmpvar))

        # step 3.4
T
typhoonzero 已提交
709
        # Iterate through the ops, and if an op and the optimize ops
710
        # which located on current pserver are in one set, then
T
typhoonzero 已提交
711
        # append it into the sub program.
T
typhoonzero 已提交
712 713 714

        global_ops = []

Y
wip  
yi.wu 已提交
715 716
        def __append_optimize_op__(op, block, grad_to_block_id, merged_var,
                                   lr_ops):
717
            if self._is_optimizer_op(op):
Q
qiaolongfei 已提交
718
                self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
719
                                         self.origin_program, merged_var)
Y
wip  
yi.wu 已提交
720
            elif op not in lr_ops:
Q
Qiyang Min 已提交
721
                self._append_pserver_non_opt_ops(block, op)
722 723 724 725 726 727

        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
728

Y
Yancey1989 已提交
729
        def __clone_lr_op_sub_block__(op, program, lr_block):
Q
Qiyang Min 已提交
730 731 732 733 734 735 736 737
            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
738
            new_sub_block = program._create_block(lr_block.idx)
Q
Qiyang Min 已提交
739 740 741

            # clone vars
            for var in origin_block.vars:
W
Wu Yi 已提交
742
                new_sub_block._clone_variable(var)
Q
Qiyang Min 已提交
743 744

            # clone ops
Y
Yancey1989 已提交
745 746
            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
747
                # clone sub_block of op
Y
Yancey1989 已提交
748
                __clone_lr_op_sub_block__(cloned_op, program, new_sub_block)
Q
Qiyang Min 已提交
749 750

            # reset the block of op
W
Wu Yi 已提交
751
            op._set_attr('sub_block', new_sub_block)
Q
Qiyang Min 已提交
752

753
        # append lr decay ops to the child block if exists
754
        lr_ops = self._get_lr_ops()
755 756
        # record optimize blocks and we can run them on pserver parallel
        optimize_blocks = []
757
        if len(lr_ops) > 0:
W
Wu Yi 已提交
758
            lr_decay_block = pserver_program._create_block(
Q
qiaolongfei 已提交
759
                pserver_program.num_blocks - 1)
760
            optimize_blocks.append(lr_decay_block)
761
            for _, op in enumerate(lr_ops):
Y
Yancey1989 已提交
762
                cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op)
Q
Qiyang Min 已提交
763
                # append sub blocks to pserver_program in lr_decay_op
Y
Yancey1989 已提交
764 765
                __clone_lr_op_sub_block__(cloned_op, pserver_program,
                                          lr_decay_block)
766

T
typhoonzero 已提交
767
        # append op to the current block
Q
qiaolongfei 已提交
768
        grad_to_block_id = []
Q
qiaolongfei 已提交
769
        pre_block_idx = pserver_program.num_blocks - 1
T
typhoonzero 已提交
770
        for idx, opt_op in enumerate(opt_op_on_pserver):
W
Wu Yi 已提交
771
            per_opt_block = pserver_program._create_block(pre_block_idx)
772
            optimize_blocks.append(per_opt_block)
773
            optimize_target_param_name = opt_op.attr(OP_ROLE_VAR_ATTR_NAME)[0]
774
            # append grad merging ops before clip and weight decay
775 776
            # e.g. merge grad -> L2Decay op -> clip op -> optimize
            merged_var = None
777
            for _, op in enumerate(self.optimize_ops):
778 779 780 781 782
                # 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:
783 784 785
                    merged_var = self._append_pserver_grad_merge_ops(
                        per_opt_block, grad_varname_for_block, endpoint,
                        grad_to_block_id, self.origin_program)
786 787 788 789 790 791
                    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 \
S
seiriosPlus 已提交
792
                            op not in global_ops:
793 794 795 796 797
                        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
798

W
Wu Yi 已提交
799
# dedup grad to ids list
W
Wu Yi 已提交
800
        grad_to_block_id = list(set(grad_to_block_id))
T
typhoonzero 已提交
801
        # append global ops
802
        if global_ops:
W
Wu Yi 已提交
803
            opt_state_block = pserver_program._create_block(
Q
qiaolongfei 已提交
804
                pserver_program.num_blocks - 1)
805
            optimize_blocks.append(opt_state_block)
Q
qiaolongfei 已提交
806
            for glb_op in global_ops:
X
Xi Chen 已提交
807
                __append_optimize_op__(glb_op, opt_state_block,
Y
wip  
yi.wu 已提交
808
                                       grad_to_block_id, None, lr_ops)
T
typhoonzero 已提交
809

810
        # process distributed lookup_table
Q
qiaolongfei 已提交
811
        prefetch_var_name_to_block_id = []
812 813
        if self.has_distributed_lookup_table:
            pserver_index = self.pserver_endpoints.index(endpoint)
814
            table_opt_block = self._create_table_optimize_block(
815
                pserver_index, pserver_program, pre_block_idx, grad_to_block_id)
816
            optimize_blocks.append(table_opt_block)
T
tangwei12 已提交
817
            lookup_table_var_name_to_block_id = self._create_prefetch_block(
818
                pserver_index, pserver_program, table_opt_block)
T
tangwei12 已提交
819 820
            checkpoint_block_id = self._create_checkpoint_save_block(
                pserver_program, table_opt_block.idx)
821

T
tangwei12 已提交
822
            pserver_program._distributed_lookup_table = self.table_name
T
tangwei12 已提交
823 824
            prefetch_var_name_to_block_id.extend(
                lookup_table_var_name_to_block_id)
825

826
        if len(optimize_blocks) == 0:
Q
Qiao Longfei 已提交
827 828
            logging.warn("pserver [" + str(endpoint) +
                         "] has no optimize block!!")
829 830 831 832 833 834
            pre_block_idx = pserver_program.num_blocks - 1
            empty_block = pserver_program._create_block(pre_block_idx)
            optimize_blocks.append(empty_block)

        # In some case, some parameter server will have no parameter to optimize
        # So we give an empty optimize block to parameter server.
835
        attrs = {
836
            "optimize_blocks": optimize_blocks,
837 838 839
            "endpoint": endpoint,
            "Fanin": self.trainer_num,
            "sync_mode": self.sync_mode,
Y
Yancey1989 已提交
840
            "grad_to_block_id": grad_to_block_id,
841
        }
T
tangwei12 已提交
842 843

        if self.has_distributed_lookup_table:
T
tangwei12 已提交
844
            attrs['checkpint_block_id'] = checkpoint_block_id
W
Wu Yi 已提交
845 846
        if self.config.enable_dc_asgd:
            attrs['dc_asgd'] = True
847

T
tangwei12 已提交
848 849 850 851
        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
852 853 854 855 856
        # step5 append the listen_and_serv op
        pserver_program.global_block().append_op(
            type="listen_and_serv",
            inputs={'X': recv_inputs},
            outputs={},
857
            attrs=attrs)
858

T
tangwei12 已提交
859
        # add distributed attrs
T
tangwei12 已提交
860
        pserver_program._slice_vars_and_attrs = self._get_slice_vars_and_attrs(
T
tangwei12 已提交
861
            endpoint)
862

W
Wu Yi 已提交
863
        pserver_program._sync_with_cpp()
W
Wu Yi 已提交
864 865
        # save pserver program to generate pserver side startup relatively.
        self.pserver_program = pserver_program
T
typhoonzero 已提交
866 867
        return pserver_program

W
Wu Yi 已提交
868 869 870 871 872 873
    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
874

W
Wu Yi 已提交
875 876 877 878
        Returns:
            tuple: (main_program, startup_program), of type "Program"
        """
        pserver_prog = self.get_pserver_program(endpoint)
W
Wu Yi 已提交
879 880
        pserver_startup = self.get_startup_program(
            endpoint, pserver_program=pserver_prog)
W
Wu Yi 已提交
881 882
        return pserver_prog, pserver_startup

883 884
    def get_startup_program(self,
                            endpoint,
W
Wu Yi 已提交
885
                            pserver_program=None,
886
                            startup_program=None):
T
typhoonzero 已提交
887
        """
W
Wu Yi 已提交
888 889
        **Deprecated**

T
typhoonzero 已提交
890 891 892
        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
893 894 895

        Args:
            endpoint (str): current pserver endpoint.
W
Wu Yi 已提交
896 897
            pserver_program (Program): deprecated, call get_pserver_program first.
            startup_program (Program): deprecated, should pass startup_program
M
minqiyang 已提交
898
                when initalizing
899

Y
yi.wu 已提交
900 901
        Returns:
            Program: parameter server side startup program.
T
typhoonzero 已提交
902
        """
903 904 905
        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
906
        if pserver_program != None:
907 908 909
            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
910
        if startup_program != None:
911 912 913
            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
914

T
typhoonzero 已提交
915
        s_prog = Program()
W
Wu Yi 已提交
916
        orig_s_prog = self.startup_program
X
Xin Pan 已提交
917
        s_prog.random_seed = orig_s_prog.random_seed
T
typhoonzero 已提交
918 919 920 921 922 923 924 925 926 927 928
        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
929
        created_var_map = collections.OrderedDict()
M
minqiyang 已提交
930
        for _, var in six.iteritems(pserver_vars):
W
Wu Yi 已提交
931
            tmpvar = s_prog.global_block()._clone_variable(var)
T
typhoonzero 已提交
932 933 934 935
            created_var_map[var.name] = tmpvar

        # 2. rename op outputs
        for op in orig_s_prog.global_block().ops:
936
            new_outputs = collections.OrderedDict()
T
typhoonzero 已提交
937 938
            # do not append startup op if var is not on this pserver
            op_on_pserver = False
G
gongweibao 已提交
939 940 941 942 943 944 945 946 947 948
            # 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
949 950

            if op_on_pserver:
951 952 953
                # most startup program ops have no inputs
                new_inputs = self._get_input_map_from_op(pserver_vars, op)

T
typhoonzero 已提交
954 955 956
                if op.type in [
                        "gaussian_random", "fill_constant", "uniform_random"
                ]:
W
Wu Yi 已提交
957
                    op._set_attr("shape", list(new_outputs["Out"].shape))
T
typhoonzero 已提交
958 959 960 961
                s_prog.global_block().append_op(
                    type=op.type,
                    inputs=new_inputs,
                    outputs=new_outputs,
G
gongweibao 已提交
962
                    attrs=op.all_attrs())
W
Wu Yi 已提交
963 964 965 966 967 968 969 970 971
        if self.config.enable_dc_asgd:
            for p, p_bak in self.param_bak_list:
                startup_param_var = s_prog.global_block().vars[p.name]
                startup_tmpvar = s_prog.global_block().vars[p_bak.name]
                # copy init random value to param_bak
                s_prog.global_block().append_op(
                    type="assign",
                    inputs={"X": startup_param_var},
                    outputs={"Out": startup_tmpvar})
972 973

        # add slice vars
T
tangwei12 已提交
974
        s_prog._slice_vars_and_attrs = self._get_slice_vars_and_attrs(endpoint)
975

T
typhoonzero 已提交
976 977
        return s_prog

T
tangwei12 已提交
978 979 980
    def _get_slice_vars_and_attrs(self, endpoint):
        slice_vars_and_attrs = []
        block_suffix = "block"
981
        for param in self.param_grad_ep_mapping[endpoint]["params"]:
T
tangwei12 已提交
982
            orig_var_name, block_name, _ = self._get_varname_parts(param.name)
T
tangwei12 已提交
983
            if not block_name:
984 985
                continue

T
tangwei12 已提交
986
            block_idx = int(block_name.split(block_suffix)[1])
987 988
            orig_var = self.origin_program.global_block().vars[orig_var_name]

T
tangwei12 已提交
989
            skip_dim0 = 0
990 991
            slice_vars = self.param_var_mapping[orig_var_name]
            for slice_var in slice_vars[:block_idx]:
T
tangwei12 已提交
992 993
                skip_dim0 += slice_var.shape[0]
            slice_vars_and_attrs.append([orig_var, skip_dim0, param])
994

T
tangwei12 已提交
995
        return slice_vars_and_attrs
996

997 998
    # ====================== private transpiler functions =====================

Y
yi.wu 已提交
999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037
    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
1038
    def _init_splited_vars(self):
Y
yi.wu 已提交
1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
        # 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
1062
        if self.config.slice_var_up:
Y
yi.wu 已提交
1063 1064
            # 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
1065 1066 1067
            grad_blocks = slice_variable(grad_list,
                                         len(self.pserver_endpoints),
                                         self.config.min_block_size)
Y
yi.wu 已提交
1068
            param_blocks = slice_variable(param_list,
G
gongweibao 已提交
1069 1070
                                          len(self.pserver_endpoints),
                                          self.config.min_block_size)
Y
yi.wu 已提交
1071 1072 1073
        else:
            # when we do NOT slice var up into blocks, we will always slice params
            # grads into one block.
G
gongweibao 已提交
1074 1075 1076 1077
            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
1078 1079
        assert (len(grad_blocks) == len(param_blocks))

1080
        # origin_param_name -> [splited_param_vars]
Y
yi.wu 已提交
1081 1082
        self.param_var_mapping = self._create_vars_from_blocklist(
            self.origin_program, param_blocks)
1083
        # origin_grad_name -> [splited_grad_vars]
Y
yi.wu 已提交
1084 1085 1086 1087
        self.grad_var_mapping = self._create_vars_from_blocklist(
            self.origin_program,
            grad_blocks,
            add_trainer_suffix=self.trainer_num > 1)
1088
        # dict(grad_splited_var -> param_splited_var)
1089
        self.grad_param_mapping = collections.OrderedDict()
Y
yi.wu 已提交
1090 1091 1092
        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
1093
            self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \
1094
                self.param_var_mapping[p_name][int(p_bid)]
Y
yi.wu 已提交
1095 1096

        # create mapping of endpoint -> split var to create pserver side program
1097
        self.param_grad_ep_mapping = collections.OrderedDict()
Y
yi.wu 已提交
1098 1099 1100 1101 1102 1103 1104 1105 1106
        [
            self.param_grad_ep_mapping.update({
                ep: {
                    "params": [],
                    "grads": []
                }
            }) for ep in self.pserver_endpoints
        ]

1107
    # transpiler function for dis lookup_table
Q
update  
qiaolongfei 已提交
1108 1109
    def _replace_lookup_table_op_with_prefetch(self, program,
                                               pserver_endpoints):
1110
        # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
S
seiriosPlus 已提交
1111
        self.all_in_ids_vars = []
Q
qiaolongfei 已提交
1112 1113
        self.all_prefetch_input_vars = []
        self.all_prefetch_output_vars = []
S
seiriosPlus 已提交
1114 1115
        self.all_out_emb_vars = []
        lookup_table_op_index = -1
1116 1117 1118 1119 1120 1121

        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:
Q
Qiao Longfei 已提交
1122 1123
                if op.type == LOOKUP_TABLE_TYPE and self.table_name == op.input(
                        "W")[0]:
1124
                    if not op.attr('is_distributed'):
Q
Qiao Longfei 已提交
1125 1126 1127
                        raise RuntimeError(
                            "lookup_table_op that lookup an distributed embedding table"
                            "should set is_distributed to true")
1128 1129
                    continue_search_lookup_table_op = True

S
seiriosPlus 已提交
1130 1131
                    lookup_table_op_index = lookup_table_op_index if lookup_table_op_index != -1 else list(
                        all_ops).index(op)
1132 1133 1134
                    ids_name = op.input("Ids")
                    out_name = op.output("Out")

Q
qiaolongfei 已提交
1135
                    ids_var = program.global_block().vars[ids_name[0]]
S
seiriosPlus 已提交
1136
                    self.all_in_ids_vars.append(ids_var)
Q
qiaolongfei 已提交
1137 1138

                    out_var = program.global_block().vars[out_name[0]]
S
seiriosPlus 已提交
1139
                    self.all_out_emb_vars.append(out_var)
1140 1141

                    # delete lookup_table_op
1142
                    delete_ops(program.global_block(), [op])
1143 1144 1145
                    # break for loop
                    break

S
seiriosPlus 已提交
1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191
        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
1192
    def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
1193
        # 2. add split_ids_op and send_op to send gradient to pservers
1194

1195 1196
        # there should only be one table_name
        all_ops = program.global_block().ops
T
typhoonzero 已提交
1197
        table_grad_name = grad_var_name(self.table_name)
1198 1199 1200 1201
        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
1202
                program.global_block()._insert_op(
1203 1204 1205 1206 1207
                    index=op_index + 1,
                    type="split_ids",
                    inputs={
                        'Ids': [program.global_block().vars[table_grad_name]]
                    },
T
tangwei12 已提交
1208 1209
                    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
1210
                program.global_block()._insert_op(
1211
                    index=op_index + 2,
1212
                    type="send",
1213
                    inputs={'X': self.trainer_side_table_grad_list},
1214 1215 1216 1217 1218
                    outputs={
                        'Out':
                        [self.grad_name_to_send_dummy_out[self.table_name]]
                        if self.sync_mode else []
                    },
Y
Yancey1989 已提交
1219
                    attrs={
1220
                        "sync_mode": not self.sync_mode,
Y
Yancey1989 已提交
1221
                        "epmap": pserver_endpoints,
W
Wu Yi 已提交
1222
                        "trainer_id": self.trainer_id,
W
Wu Yi 已提交
1223 1224 1225 1226 1227
                        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
1228
                    })
1229 1230 1231 1232 1233 1234
                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
1235
        prefetch_var_name_to_block_id = []
S
seiriosPlus 已提交
1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260
        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
1261
        return prefetch_var_name_to_block_id
1262 1263

    def _create_table_optimize_block(self, pserver_index, pserver_program,
1264
                                     pre_block_idx, grad_to_block_id):
1265
        # STEP: create table optimize block
1266
        table_opt_block = pserver_program._create_block(pre_block_idx)
1267
        # create table param and grad var in pserver program
1268 1269
        # create table optimize block in pserver program
        table_opt_op = [
Q
Qiao Longfei 已提交
1270 1271
            op for op in self.optimize_ops if 'Param' in op.input_names and
            op.input("Param")[0] == self.table_name
1272 1273
        ][0]

Y
Yancey1989 已提交
1274 1275
        origin_param_var = self.origin_program.global_block().vars[
            self.table_name]
T
tangwei12 已提交
1276

T
tangwei12 已提交
1277
        zero_dim = int(
T
bug fix  
tangwei12 已提交
1278 1279
            math.ceil(origin_param_var.shape[0] / float(
                len(self.pserver_endpoints))))
T
tangwei12 已提交
1280 1281 1282
        table_shape = list(origin_param_var.shape)
        table_shape[0] = zero_dim

Y
Yancey1989 已提交
1283 1284
        param_var = pserver_program.global_block().create_var(
            name=origin_param_var.name,
T
tangwei12 已提交
1285
            shape=table_shape,
Y
Yancey1989 已提交
1286 1287 1288
            dtype=origin_param_var.dtype,
            type=core.VarDesc.VarType.SELECTED_ROWS,
            persistable=True)
1289

1290 1291
        # parameter must be selected rows
        param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS)
W
Wu Yi 已提交
1292
        grad_var = pserver_program.global_block()._clone_variable(
T
typhoonzero 已提交
1293
            self.origin_program.global_block().vars[grad_var_name(
1294
                self.table_name)])
1295

1296 1297 1298
        lr_var = pserver_program.global_block()._clone_variable(
            self.origin_program.global_block().vars[table_opt_op.input(
                "LearningRate")[0]])
1299

1300 1301 1302
        if self.sync_mode:
            # create grad vars in pserver program
            table_grad_var = self.table_param_grad[1]
1303
            pserver_side_table_grad_list = [
1304 1305 1306 1307 1308 1309 1310 1311 1312
                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)
            ]

1313
            # append sum op for pserver_side_table_grad_list
1314 1315
            table_opt_block.append_op(
                type="sum",
1316
                inputs={"X": pserver_side_table_grad_list},
1317 1318
                outputs={"Out": [grad_var]},
                attrs={"use_mkldnn": False})
1319 1320
        else:
            # in async_mode, for table gradient, it also need to be splited to each parameter server
1321
            origin_grad_name = grad_var.name
1322 1323
            splited_grad_name = self.trainer_side_table_grad_list[
                pserver_index].name
1324 1325
            if not splited_grad_name.startswith(origin_grad_name):
                raise ValueError("origin_grad_var: " + splited_grad_name +
1326
                                 " grad_var:" + grad_var.name)
W
Wu Yi 已提交
1327
            grad_var = pserver_program.global_block()._rename_var(
1328
                origin_grad_name, splited_grad_name)
1329 1330 1331 1332 1333 1334 1335

        inputs = {
            "Param": [param_var],
            "Grad": [grad_var],
            "LearningRate": [lr_var]
        }
        outputs = {"ParamOut": [param_var]}
1336
        # only support sgd now
1337 1338 1339
        logging.warn(
            "distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of "
            + table_opt_op.type)
1340
        table_opt_block.append_op(type="sgd", inputs=inputs, outputs=outputs)
1341

1342 1343 1344
        # 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))

1345 1346
        return table_opt_block

T
tangwei12 已提交
1347 1348 1349 1350 1351
    def _create_checkpoint_save_block(self, pserver_program, pre_block_idx):
        """
        create a new block to handle save checkpoint.
        """

T
tangwei12 已提交
1352
        pserver_program.global_block().create_var(
T
tangwei12 已提交
1353
            name="kLookupTablePath",
T
tangwei12 已提交
1354 1355
            persistable=True,
            type=core.VarDesc.VarType.RAW)
T
tangwei12 已提交
1356

W
Wu Yi 已提交
1357
        checkpoint_save_block = pserver_program._create_block(pre_block_idx)
T
tangwei12 已提交
1358
        # this 'file_path' do not be used in save lookup table variable
T
tangwei12 已提交
1359 1360 1361 1362
        checkpoint_save_block.append_op(
            type='save',
            inputs={'X': [self.table_name]},
            outputs={},
T
tangwei12 已提交
1363
            attrs={'file_path': "none"})
T
tangwei12 已提交
1364 1365 1366

        return checkpoint_save_block.idx

T
typhoonzero 已提交
1367 1368 1369 1370 1371
    def _create_vars_from_blocklist(self,
                                    program,
                                    block_list,
                                    add_trainer_suffix=False):
        """
1372
        Create vars for each split.
T
typhoonzero 已提交
1373 1374
        NOTE: only grads need to be named for different trainers, use
              add_trainer_suffix to rename the grad vars.
1375 1376 1377 1378
        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.
1379
        Returns:
1380
            var_mapping (collections.OrderedDict(varname->[new_varname_variable])):A dict mapping
1381
                from original var name to each var split.
T
typhoonzero 已提交
1382
        """
1383 1384

        # varname->[(block_id, current_block_size)]
1385
        block_map = collections.OrderedDict()
1386

1387
        var_mapping = collections.OrderedDict()
T
typhoonzero 已提交
1388 1389
        for block_str in block_list:
            varname, offset, size = block_str.split(":")
1390
            if varname not in block_map:
T
typhoonzero 已提交
1391
                block_map[varname] = []
1392
            block_map[varname].append((int(offset), int(size)))
Y
yi.wu 已提交
1393

M
minqiyang 已提交
1394
        for varname, splited in six.iteritems(block_map):
T
typhoonzero 已提交
1395
            orig_var = program.global_block().var(varname)
T
typhoonzero 已提交
1396
            if len(splited) == 1:
1397
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1398
                    new_var_name = "%s.trainer_%d" % \
T
tangwei12 已提交
1399
                                   (orig_var.name, self.trainer_id)
W
Wu Yi 已提交
1400
                    program.global_block()._rename_var(varname, new_var_name)
T
typhoonzero 已提交
1401 1402 1403 1404 1405
                    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
1406
                continue
T
typhoonzero 已提交
1407
            var_mapping[varname] = []
T
typhoonzero 已提交
1408 1409 1410 1411
            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
1412

T
typhoonzero 已提交
1413
            for i, block in enumerate(splited):
T
typhoonzero 已提交
1414
                size = block[1]
M
minqiyang 已提交
1415
                rows = size // orig_dim1_flatten
T
typhoonzero 已提交
1416 1417 1418
                splited_shape = [rows]
                if len(orig_shape) >= 2:
                    splited_shape.extend(orig_shape[1:])
T
typhoonzero 已提交
1419
                new_var_name = ""
1420
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1421
                    new_var_name = "%s.block%d.trainer_%d" % \
T
tangwei12 已提交
1422
                                   (varname, i, self.trainer_id)
T
typhoonzero 已提交
1423 1424
                else:
                    new_var_name = "%s.block%d" % \
T
tangwei12 已提交
1425
                                   (varname, i)
T
typhoonzero 已提交
1426
                var = program.global_block().create_var(
T
typhoonzero 已提交
1427 1428
                    name=new_var_name,
                    persistable=False,
T
typhoonzero 已提交
1429
                    dtype=orig_var.dtype,
1430
                    type=orig_var.type,
T
typhoonzero 已提交
1431
                    shape=splited_shape)  # flattend splited var
T
typhoonzero 已提交
1432
                var_mapping[varname].append(var)
W
Wu Yi 已提交
1433
            program.global_block()._sync_with_cpp()
T
typhoonzero 已提交
1434
        return var_mapping
T
done  
typhoonzero 已提交
1435

1436
    def _clone_var(self, block, var, persistable=True):
T
done  
typhoonzero 已提交
1437 1438 1439 1440 1441 1442
        return block.create_var(
            name=var.name,
            shape=var.shape,
            dtype=var.dtype,
            type=var.type,
            lod_level=var.lod_level,
1443
            persistable=persistable)
T
done  
typhoonzero 已提交
1444

Y
Yancey1989 已提交
1445
    def _insert_split_op(self, program, orig_var, index, splited_vars):
Y
update  
Yancey1989 已提交
1446 1447 1448 1449
        if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS:
            height_sections = []
            for v in splited_vars:
                height_sections.append(v.shape[0])
Q
Qiao Longfei 已提交
1450 1451 1452 1453
            sparse_param_name = self.grad_name_to_param_name[orig_var.name]
            if sparse_param_name != self.table_name:
                self.sparse_param_to_height_sections[
                    sparse_param_name] = height_sections
W
Wu Yi 已提交
1454
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1455 1456 1457 1458
                index=index + 1,
                type="split_selected_rows",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
1459 1460 1461 1462
                attrs={
                    "height_sections": height_sections,
                    RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE
                })
Y
update  
Yancey1989 已提交
1463 1464 1465 1466
        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
1467
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1468 1469 1470 1471
                index=index + 1,
                type="split_byref",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
1472 1473 1474 1475
                attrs={
                    "sections": sections,
                    RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE
                })
Y
update  
Yancey1989 已提交
1476 1477 1478
        else:
            AssertionError("Variable type should be in set "
                           "[LOD_TENSOR, SELECTED_ROWS]")
T
done  
typhoonzero 已提交
1479

T
typhoonzero 已提交
1480 1481 1482 1483
    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
1484
        Param and Grad is split to multiple servers.
T
typhoonzero 已提交
1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496
        """
        # 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
1497
        elif op_type in ["momentum", "lars_momentum"]:
T
typhoonzero 已提交
1498 1499
            if varkey == "Velocity":
                return param_shape
W
Wu Yi 已提交
1500 1501
        elif op_type == "rmsprop":
            if varkey in ["Moment", "MeanSquare"]:
T
typhoonzero 已提交
1502
                return param_shape
1503 1504 1505
        elif op_type == "decayed_adagrad":
            if varkey == "Moment":
                return param_shape
1506 1507 1508
        elif op_type == "ftrl":
            if varkey in ["SquaredAccumulator", "LinearAccumulator"]:
                return param_shape
T
typhoonzero 已提交
1509 1510
        elif op_type == "sgd":
            pass
1511 1512 1513 1514
        else:
            raise ValueError(
                "Not supported optimizer for distributed training: %s" %
                op_type)
T
typhoonzero 已提交
1515 1516
        return orig_shape

1517 1518
    def _get_varname_parts(self, varname):
        # returns origin, blockid, trainerid
T
typhoonzero 已提交
1519
        orig_var_name = ""
1520 1521 1522 1523 1524 1525 1526 1527 1528 1529
        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
1530
        else:
1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552
            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
1553
            return None
1554 1555 1556 1557
        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
1558
        else:
1559
            merged_var_name = orig_varname
1560 1561

        merged_var = pserver_block.vars[merged_var_name]
1562 1563 1564
        grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx))
        if self.sync_mode and self.trainer_num > 1:
            vars2merge = []
1565
            for i in range(self.trainer_num):
1566
                per_trainer_name = "%s.trainer_%d" % \
T
tangwei12 已提交
1567
                                   (merged_var_name, i)
1568 1569 1570 1571
                vars2merge.append(pserver_block.vars[per_trainer_name])
            optimize_block.append_op(
                type="sum",
                inputs={"X": vars2merge},
1572 1573
                outputs={"Out": merged_var},
                attrs={"use_mkldnn": False})
Q
qiaolongfei 已提交
1574 1575 1576 1577 1578
            optimize_block.append_op(
                type="scale",
                inputs={"X": merged_var},
                outputs={"Out": merged_var},
                attrs={"scale": 1.0 / float(self.trainer_num)})
1579
        return merged_var
T
typhoonzero 已提交
1580

W
Wu Yi 已提交
1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642
    def _append_dc_asgd_ops(self, block, param_var, grad_var):
        # NOTE: can not use grammar candy here, should put ops in specific block
        local_param_bak = block.create_var(
            name="%s.local_bak" % param_var.name,
            shape=param_var.shape,
            type=param_var.type,
            dtype=param_var.dtype,
            persistable=False)
        # trainer_id_var is block local
        trainer_id_var = block.create_var(
            name="@TRAINER_ID@",
            type=core.VarDesc.VarType.LOD_TENSOR,
            dtype=core.VarDesc.VarType.INT64,
            shape=[1],
            persistable=False)

        # ref_inputs = [x[1] for x in self.param_bak_list]
        ref_inputs = []
        for p, p_bak in self.param_bak_list:
            if p.name == param_var.name:
                ref_inputs.append(p_bak)
        block.append_op(
            type="ref_by_trainer_id",
            inputs={"X": ref_inputs,
                    "TrainerId": trainer_id_var},
            outputs={"Out": local_param_bak})

        def __create_temp_var__():
            return block.create_var(
                name=unique_name.generate("tmp_dc_output"),
                shape=param_var.shape,
                type=param_var.type,
                dtype=param_var.dtype,
                persistable=False)

        o1 = __create_temp_var__()
        block.append_op(
            type="elementwise_sub",
            inputs={"X": param_var,
                    "Y": local_param_bak},
            outputs={"Out": o1})
        o2 = __create_temp_var__()
        block.append_op(
            type="elementwise_mul",
            inputs={"X": o1,
                    "Y": grad_var},
            outputs={"Out": o2})
        o3 = __create_temp_var__()
        block.append_op(
            type="elementwise_mul",
            inputs={"X": o2,
                    "Y": grad_var},
            outputs={"Out": o3})
        # TODO(typhoonzero): append scale
        o4 = __create_temp_var__()
        block.append_op(
            type="elementwise_add",
            inputs={"X": grad_var,
                    "Y": o3},
            outputs={"Out": o4})
        return o4

1643
    def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
1644
                            grad_to_block_id, origin_program, merged_var):
1645
        program = optimize_block.program
T
typhoonzero 已提交
1646
        pserver_block = program.global_block()
1647
        new_inputs = collections.OrderedDict()
W
Wu Yi 已提交
1648 1649 1650 1651 1652 1653 1654 1655 1656 1657

        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

W
Wu Yi 已提交
1658 1659 1660 1661
        if self.config.enable_dc_asgd:
            param_var = _get_param_block(opt_op)
            dc = self._append_dc_asgd_ops(optimize_block, param_var, merged_var)

T
typhoonzero 已提交
1662
        for key in opt_op.input_names:
T
typhoonzero 已提交
1663
            if key == "Grad":
W
Wu Yi 已提交
1664 1665 1666 1667
                if self.config.enable_dc_asgd:
                    new_inputs[key] = dc
                else:
                    new_inputs[key] = merged_var
T
typhoonzero 已提交
1668
            elif key == "Param":
W
Wu Yi 已提交
1669
                param_block = _get_param_block(opt_op)
T
typhoonzero 已提交
1670 1671
                if not param_block:
                    return
T
typhoonzero 已提交
1672
                tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1673
                    name=param_block.name,
T
typhoonzero 已提交
1674
                    persistable=True,
T
typhoonzero 已提交
1675 1676 1677
                    dtype=param_block.dtype,
                    shape=param_block.shape)
                new_inputs[key] = tmpvar
1678
            elif key == "LearningRate":
1679
                # learning rate variable has already be created by non-optimize op,
1680
                # don't create it once again.
1681
                lr_varname = opt_op.input(key)[0]
1682
                if lr_varname in pserver_block.vars:
1683 1684 1685 1686 1687 1688 1689 1690 1691
                    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
1692

T
typhoonzero 已提交
1693
        for key in opt_op.input_names:
1694
            new_shape = None
W
Wu Yi 已提交
1695
            if key in ["Param", "Grad", "LearningRate"]:
T
typhoonzero 已提交
1696
                continue
1697
            var = self.origin_program.global_block().vars[opt_op.input(key)[0]]
T
typhoonzero 已提交
1698 1699 1700 1701
            # 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
1702
            tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1703 1704 1705 1706 1707
                name=var.name,
                persistable=var.persistable,
                dtype=var.dtype,
                shape=new_shape)
            new_inputs[key] = tmpvar
T
typhoonzero 已提交
1708

1709
        # change output's ParamOut variable
1710 1711
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1712
        outputs["ParamOut"] = new_inputs["Param"]
1713
        optimize_block.append_op(
T
typhoonzero 已提交
1714 1715
            type=opt_op.type,
            inputs=new_inputs,
T
typhoonzero 已提交
1716
            outputs=outputs,
G
gongweibao 已提交
1717
            attrs=opt_op.all_attrs())
T
typhoonzero 已提交
1718

1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729
    def _get_pserver_grad_param_var(self, var, var_dict):
        """
        Return pserver side grad/param variable, return None
        if the variable is not grad/param, e.g.

            a@GRAD -> a@GRAD.block0
            a@GRAD -> a@GRAD (a is not splited)
            fc_0.w_0 -> fc_0.w_0.block_0
            fc_0.w_0 -> fc_0.w_0 (weight is not splited)
            _generated_var_123 -> None
        """
1730
        grad_block = None
M
minqiyang 已提交
1731
        for _, g in six.iteritems(var_dict):
1732
            if self._orig_varname(g.name) == self._orig_varname(var.name):
1733
                # skip per trainer vars
1734
                if g.name.find(".trainer_") == -1:
1735 1736 1737 1738 1739
                    # only param or grads have splited blocks
                    if self._orig_varname(g.name) in self.grad_name_to_param_name or\
                        self._orig_varname(g.name) in self.param_name_to_grad_name:
                        grad_block = g
                        break
1740 1741
        return grad_block

Q
Qiyang Min 已提交
1742 1743 1744
    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
1745
        for key, varlist in six.iteritems(inputs):
Q
Qiyang Min 已提交
1746 1747 1748 1749
            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
1750
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1751 1752 1753

        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, op)
M
minqiyang 已提交
1754
        for key, varlist in six.iteritems(outputs):
Q
Qiyang Min 已提交
1755 1756 1757 1758
            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
1759
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1760

Y
Yancey1989 已提交
1761
        return block.append_op(
G
gongweibao 已提交
1762
            type=op.type, inputs=inputs, outputs=outputs, attrs=op.all_attrs())
Q
Qiyang Min 已提交
1763 1764

    def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
1765
        program = optimize_block.program
1766
        # Append the ops for parameters that do not need to be optimized/updated
1767 1768
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, opt_op)
M
minqiyang 已提交
1769
        for key, varlist in six.iteritems(inputs):
1770 1771
            if not isinstance(varlist, list):
                varlist = [varlist]
1772 1773 1774
            for i in range(len(varlist)):
                var = varlist[i]
                # for ops like clipping and weight decay, get the splited var (xxx.block0)
1775
                # for inputs/outputs
1776
                grad_block = self._get_pserver_grad_param_var(
1777 1778
                    var, program.global_block().vars)
                if grad_block:
1779
                    varlist[i] = grad_block
1780
                elif var.name not in program.global_block().vars:
1781 1782 1783 1784 1785
                    tmpvar = program.global_block()._clone_variable(var)
                    varlist[i] = tmpvar
                else:
                    varlist[i] = program.global_block().vars[var.name]
            inputs[key] = varlist
T
typhoonzero 已提交
1786

1787 1788
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
M
minqiyang 已提交
1789
        for key, varlist in six.iteritems(outputs):
1790 1791
            if not isinstance(varlist, list):
                varlist = [varlist]
1792 1793 1794
            for i in range(len(varlist)):
                var = varlist[i]
                grad_block = self._get_pserver_grad_param_var(
1795 1796
                    var, program.global_block().vars)
                if grad_block:
1797
                    varlist[i] = grad_block
1798
                elif var.name not in program.global_block().vars:
1799 1800 1801 1802 1803
                    tmpvar = program.global_block()._clone_variable(var)
                    varlist[i] = tmpvar
                else:
                    varlist[i] = program.global_block().vars[var.name]
            outputs[key] = varlist
1804

Y
Yancey1989 已提交
1805
        return optimize_block.append_op(
T
typhoonzero 已提交
1806
            type=opt_op.type,
T
typhoonzero 已提交
1807 1808
            inputs=inputs,
            outputs=outputs,
G
gongweibao 已提交
1809
            attrs=opt_op.all_attrs())
T
typhoonzero 已提交
1810

1811 1812 1813 1814
    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
1815
        if set(op1.desc.output_arg_names()) & set(op2.desc.input_arg_names()) or \
T
tangwei12 已提交
1816
                set(op1.desc.input_arg_names()) & set(op2.desc.output_arg_names()):
1817 1818 1819 1820 1821 1822
            return True
        return False

    def _create_ufind(self, optimize_ops):
        # Create a unit find data struct by optimize ops
        ufind = UnionFind(optimize_ops)
1823 1824
        for i in range(len(optimize_ops)):
            for j in range(i, len(optimize_ops)):
1825 1826 1827 1828 1829 1830
                op1 = optimize_ops[i]
                op2 = optimize_ops[j]
                if self._is_op_connected(op1, op2):
                    ufind.union(op1, op2)
        return ufind

1831
    def _is_optimizer_op(self, op):
T
typhoonzero 已提交
1832
        if "Param" in op.input_names and \
T
tangwei12 已提交
1833
                "LearningRate" in op.input_names:
1834 1835 1836 1837 1838 1839 1840
            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
1841
        if op.input("Param")[0] in param_names:
1842 1843 1844
            return True
        else:
            for n in param_names:
T
typhoonzero 已提交
1845
                param = op.input("Param")[0]
T
typhoonzero 已提交
1846
                if same_or_split_var(n, param) and n != param:
1847 1848 1849
                    return True
            return False

T
typhoonzero 已提交
1850
    def _get_input_map_from_op(self, varmap, op):
1851
        """Returns a dict from op input name to the vars in varmap."""
1852
        iomap = collections.OrderedDict()
T
typhoonzero 已提交
1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863
        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):
1864
        """Returns a dict from op output name to the vars in varmap."""
1865
        iomap = collections.OrderedDict()
T
typhoonzero 已提交
1866 1867 1868 1869 1870 1871 1872 1873 1874
        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
1875 1876

    def _get_lr_ops(self):
1877 1878 1879
        lr_ops = []
        block = self.origin_program.global_block()
        for op in block.ops:
X
fix  
Xin Pan 已提交
1880 1881 1882 1883
            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):
1884 1885 1886 1887 1888
                lr_ops.append(op)
                log("append lr op: ", op.type)
        return lr_ops

    def _get_lr_ops_deprecated(self):
1889 1890 1891 1892
        lr_ops = []
        # find learning rate variables by optimize op
        lr_vars = set()
        for op in self.optimize_ops:
1893
            if self._is_optimizer_op(op):
1894 1895 1896 1897
                lr_vars.add(op.input("LearningRate")[0])

        find_ops = []
        # find ops which output is lr var
1898
        block = self.origin_program.global_block()
1899 1900 1901 1902 1903
        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)
1904

1905 1906 1907 1908 1909
        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
1910
                        not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2):
1911 1912 1913 1914 1915 1916
                    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)
1917 1918
                    # we only need to append op for once
                    break
1919
        return lr_ops
Y
Yancey1989 已提交
1920

W
Wu Yi 已提交
1921 1922 1923 1924 1925
    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
1926 1927
        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
1928 1929 1930
            return True
        return False

Y
Yancey1989 已提交
1931
    def _get_optimize_pass(self):
1932
        """
1933
        Get optimizer operators, parameters and gradients from origin_program
1934 1935 1936 1937
        Returns:
            opt_ops (list): optimize operators.
            params_grads (dict): paramter->gradient.
        """
Y
Yancey1989 已提交
1938 1939 1940
        block = self.origin_program.global_block()
        opt_ops = []
        params_grads = []
1941 1942
        # tmp set to dedup
        optimize_params = set()
1943
        origin_var_dict = self.origin_program.global_block().vars
Y
Yancey1989 已提交
1944
        for op in block.ops:
W
Wu Yi 已提交
1945
            if self._is_opt_role_op(op):
Y
Yancey1989 已提交
1946
                opt_ops.append(op)
1947 1948 1949 1950 1951 1952
                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)
1953 1954
                        params_grads.append([
                            origin_var_dict[param_name],
1955
                            origin_var_dict[grad_name]
1956
                        ])
Y
Yancey1989 已提交
1957 1958 1959
            else:
                pass
        return opt_ops, params_grads