distribute_transpiler.py 59.9 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 16 17 18
"""
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
19
4. append send_op to send splited variables to server and
20 21
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.
22 23 24 25 26 27 28 29

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
30

T
typhoonzero 已提交
31
from __future__ import print_function
32

T
typhoonzero 已提交
33
import math
S
seiriosPlus 已提交
34
import random
35
import numpy as np
36

Y
Yancey1989 已提交
37
from ps_dispatcher import RoundRobin, HashName, PSDispatcher
Y
Yancey 已提交
38
from .. import core, framework
T
typhoonzero 已提交
39
from ..framework import Program, default_main_program, \
Q
Qiyang Min 已提交
40
                        default_startup_program, Block, \
W
Wu Yi 已提交
41
                        Parameter, grad_var_name
42
from details import *
43 44 45

LOOKUP_TABLE_TYPE = "lookup_table"
LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
46
OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
Y
Yancey1989 已提交
47 48 49
RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName(
)
RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
T
done  
typhoonzero 已提交
50 51


T
typhoonzero 已提交
52 53 54 55 56 57
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
58

T
typhoonzero 已提交
59 60
    def __str__(self):
        return "%s:%d:%d" % (self.varname, self.offset, self.size)
T
done  
typhoonzero 已提交
61 62


63 64 65 66
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
67
def slice_variable(var_list, slice_count, min_block_size):
T
typhoonzero 已提交
68
    """
69 70 71 72 73 74
    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
75
    minimum block size 8K elements (maybe 16bit or 32bit or 64bit).
76 77 78

    Args:
        var_list (list): List of variables.
79 80
        slice_count (int): Numel of count that variables will be sliced, which
            could be the pserver services' count.
81 82
        min_block_size (int): Minimum splitted block size.
    Returns:
83
        blocks (list[(varname, block_id, current_block_size)]): A list
84
            of VarBlocks. Each VarBlock specifies a shard of the var.
T
typhoonzero 已提交
85 86 87
    """
    blocks = []
    for var in var_list:
88
        split_count = slice_count
T
typhoonzero 已提交
89 90 91 92
        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
93
        if max_pserver_count < slice_count:
T
typhoonzero 已提交
94 95 96 97 98 99 100 101 102
            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
103
        # update split_count after aligning
T
typhoonzero 已提交
104 105 106 107 108 109 110 111 112
        split_count = int(math.ceil(var_numel / float(block_size)))
        for block_id in xrange(split_count):
            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
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128
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
        We can use bandwidth effiently when data size is larger than 2MB.If you 
        want to change it, please be sure you see the slice_variable function.
    """

    slice_var_up = True
    split_method = None
    min_block_size = 8192


Y
gen rst  
yi.wu 已提交
129
class DistributeTranspiler(object):
Y
yi.wu 已提交
130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
    """
    **DistributeTranspiler**

    Convert the fluid program to distributed data-parallelism programs.

    The main_program will be transformed to use a remote parameter server
    to do parameter optimization. And the optimization graph will be put
    into a parameter server program.

    Examples:
        .. code-block:: python

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

           t = distribute_transpiler.DistributeTranspiler()
           t.transpile(
                trainer_id, pservers=pserver_endpoints, trainers=trainers)
           if role == "PSERVER":
                pserver_program = t.get_pserver_program(current_endpoint)
                pserver_startup_program = t.get_startup_program(current_endpoint,
                                                                pserver_program)
           elif role == "TRAINER":
                trainer_program = t.get_trainer_program()
    """
Y
Yancey1989 已提交
164

G
gongweibao 已提交
165 166 167 168 169 170 171 172 173 174 175 176
    def __init__(self, config=None):
        if config is not None:
            self.config = config
        else:
            self.config = DistributeTranspilerConfig()

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

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

177 178 179 180 181 182 183
    def transpile(self,
                  trainer_id,
                  program=None,
                  pservers="127.0.0.1:6174",
                  trainers=1,
                  sync_mode=True):
        """
Y
yi.wu 已提交
184 185 186 187 188 189 190 191 192 193 194
        Run the transpiler.

        Args:
            trainer_id (int): id for current trainer worker, if you have
                n workers, the id may range from 0 ~ n-1
            program (Program|None): program to transpile,
                default is fluid.default_main_program().
            pservers (str): comma separated ip:port string for the pserver
                list.
            trainers (int): number of trainers in the distributed job.
            sync_mode (bool): Do sync training or not, default is True.
195 196 197 198 199 200 201 202 203 204 205
        """
        if program is None:
            program = default_main_program()
        self.origin_program = program
        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
206
        ps_dispatcher = self.config.split_method(self.pserver_endpoints)
207 208 209
        self.has_distributed_lookup_table = self._has_distributed_lookup_table()

        # split and create vars, then put splited vars in dicts for later use.
G
gongweibao 已提交
210
        self._init_splited_vars()
211

Y
Yancey1989 已提交
212 213
        # step 3.1: insert send op to send gradient vars to parameter servers
        ps_dispatcher.reset()
Y
update  
Yancey1989 已提交
214
        send_vars = []
215 216 217 218 219 220

        # 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
221
        grad_var_mapping_items = self.grad_var_mapping.items()
G
gongweibao 已提交
222
        if not self.config.slice_var_up:
S
seiriosPlus 已提交
223 224
            random.seed(self.trainer_num)
            random.shuffle(grad_var_mapping_items)
225 226

        for orig_varname, splited_vars in grad_var_mapping_items:
Y
update  
Yancey1989 已提交
227
            eplist = ps_dispatcher.dispatch(splited_vars)
228

G
gongweibao 已提交
229
            if not self.config.slice_var_up:
230 231
                assert (len(splited_vars) == 1)

Y
Yancey1989 已提交
232 233 234 235 236 237 238 239 240
            if len(splited_vars) == 1:
                orig_varname = splited_vars[0].name
                index = find_op_by_output_arg(program.global_block(),
                                              orig_varname)
            elif len(splited_vars) > 1:
                orig_var = program.global_block().vars[orig_varname]
                index = find_op_by_output_arg(program.global_block(),
                                              orig_varname)
                self._insert_split_op(program, orig_var, index, splited_vars)
Y
update  
Yancey1989 已提交
241
                index += 1
Y
Yancey1989 已提交
242 243 244 245
            else:
                AssertionError("Can not insert the send op by original "
                               "variable name :", orig_varname)

W
Wu Yi 已提交
246
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
247
                index=index + 1,
248
                type="send",
Y
update  
Yancey1989 已提交
249
                inputs={"X": splited_vars},
Y
Yancey1989 已提交
250 251 252 253 254
                outputs={},
                attrs={
                    "epmap": eplist,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })
Y
update  
Yancey1989 已提交
255 256
            for _, var in enumerate(splited_vars):
                send_vars.append(var)
Y
Yancey1989 已提交
257 258 259 260 261

        if self.sync_mode:
            program.global_block().append_op(
                type="send_barrier",
                inputs={},
Y
Yancey1989 已提交
262
                outputs={},
Y
Yancey1989 已提交
263 264
                attrs={
                    "endpoints": pserver_endpoints,
Y
Yancey1989 已提交
265 266
                    "sync_mode": self.sync_mode,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Y
Yancey1989 已提交
267
                })
Y
Yancey1989 已提交
268 269 270

        # step 3.2: insert recv op to receive parameters from parameter server
        recv_vars = []
Y
update  
Yancey1989 已提交
271
        for _, var in enumerate(send_vars):
272
            recv_vars.append(self.grad_param_mapping[var])
Y
update  
Yancey1989 已提交
273
        ps_dispatcher.reset()
Y
Yancey1989 已提交
274 275
        eplist = ps_dispatcher.dispatch(recv_vars)

T
typhoonzero 已提交
276
        for i, ep in enumerate(eplist):
Y
Yancey1989 已提交
277 278
            self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
            self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
279

Y
Yancey1989 已提交
280
        # step4: Concat the parameters splits together after recv.
281
        for varname, splited_var in self.param_var_mapping.iteritems():
Y
Yancey1989 已提交
282 283 284 285 286 287 288 289
            eps = []
            for var in splited_var:
                index = [v.name for v in recv_vars].index(var.name)
                eps.append(eplist[index])

            program.global_block().append_op(
                type="recv",
                inputs={},
Y
Yancey1989 已提交
290 291 292 293 294
                outputs={"Out": splited_var},
                attrs={
                    "epmap": eps,
                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                })
T
typhoonzero 已提交
295

T
typhoonzero 已提交
296
        program.global_block().append_op(
Y
Yancey1989 已提交
297 298
            type="fetch_barrier",
            inputs={},
Y
Yancey1989 已提交
299
            outputs={},
Q
qiaolongfei 已提交
300 301
            attrs={
                "endpoints": pserver_endpoints,
Y
Yancey1989 已提交
302
                RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Q
qiaolongfei 已提交
303
            })
Y
Yancey1989 已提交
304

305
        for varname, splited_var in self.param_var_mapping.iteritems():
T
typhoonzero 已提交
306 307
            if len(splited_var) <= 1:
                continue
T
typhoonzero 已提交
308
            orig_param = program.global_block().vars[varname]
T
typhoonzero 已提交
309
            program.global_block().append_op(
T
typhoonzero 已提交
310
                type="concat",
T
typhoonzero 已提交
311
                inputs={"X": splited_var},
T
typhoonzero 已提交
312
                outputs={"Out": [orig_param]},
T
typhoonzero 已提交
313
                attrs={"axis": 0})
T
typhoonzero 已提交
314

315
        if self.has_distributed_lookup_table:
Q
update  
qiaolongfei 已提交
316 317
            self._replace_lookup_table_op_with_prefetch(program,
                                                        pserver_endpoints)
Y
Yancey1989 已提交
318
            self._split_table_grad_and_add_send_vars(program, pserver_endpoints)
319

T
typhoonzero 已提交
320
    def get_trainer_program(self):
Y
yi.wu 已提交
321 322 323 324 325 326
        """
        Get transpiled trainer side program.

        Returns:
            Program: trainer side program.
        """
T
typhoonzero 已提交
327
        # remove optimize ops and add a send op to main_program
X
Xin Pan 已提交
328
        # FIXME(typhoonzero): Also ops like clip_gradient, lrn_decay?
329
        delete_ops(self.origin_program.global_block(), self.optimize_ops)
330 331
        self.origin_program.__str__()
        return self.origin_program
T
typhoonzero 已提交
332 333 334

    def get_pserver_program(self, endpoint):
        """
Y
yi.wu 已提交
335
        Get parameter server side program.
336

Y
yi.wu 已提交
337 338
        Args:
            endpoint (str): current parameter server endpoint.
339

Y
yi.wu 已提交
340 341
        Returns:
            Program: the program for current parameter server to run.
T
typhoonzero 已提交
342
        """
Y
yi.wu 已提交
343 344 345 346 347
        # 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.

T
typhoonzero 已提交
348 349
        # step1
        pserver_program = Program()
350
        # step2: Create vars to receive vars at parameter servers.
T
typhoonzero 已提交
351 352 353 354 355 356 357 358
        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
359 360 361 362 363
            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
364 365 366 367 368 369 370 371 372
            # 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)
373
            if self.sync_mode and self.trainer_num > 1:
374
                for trainer_id in xrange(self.trainer_num):
T
typhoonzero 已提交
375 376 377 378 379 380 381 382 383
                    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)
384

Q
qiaolongfei 已提交
385
        # step 3
386
        # Create a union-find data structure from optimize ops,
T
typhoonzero 已提交
387 388 389
        # 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
390
        # step 3.2
T
typhoonzero 已提交
391 392 393 394
        # Iterate through the ops and append optimize op which
        # located on current pserver
        opt_op_on_pserver = []
        for _, op in enumerate(self.optimize_ops):
395 396
            if self._is_optimizer_op(op) and self._is_opt_op_on_pserver(
                    endpoint, op):
T
typhoonzero 已提交
397
                opt_op_on_pserver.append(op)
Q
qiaolongfei 已提交
398
        # step 3.3
T
typhoonzero 已提交
399
        # Iterate through the ops, and if an op and the optimize ops
400
        # which located on current pserver are in one set, then
T
typhoonzero 已提交
401
        # append it into the sub program.
T
typhoonzero 已提交
402 403 404

        global_ops = []

Y
wip  
yi.wu 已提交
405 406
        def __append_optimize_op__(op, block, grad_to_block_id, merged_var,
                                   lr_ops):
407
            if self._is_optimizer_op(op):
Q
qiaolongfei 已提交
408
                self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
409
                                         self.origin_program, merged_var)
Y
wip  
yi.wu 已提交
410
            elif op not in lr_ops:
Q
Qiyang Min 已提交
411
                self._append_pserver_non_opt_ops(block, op)
412 413 414 415 416 417

        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
418

Y
Yancey1989 已提交
419
        def __clone_lr_op_sub_block__(op, program, lr_block):
Q
Qiyang Min 已提交
420 421 422 423 424 425 426 427
            if not op.has_attr('sub_block'):
                return

            origin_block_desc = op.attr('sub_block')
            origin_block = self.origin_program.block(origin_block_desc.id)
            assert isinstance(origin_block, Block)
            # we put the new sub block to new block to follow the block
            # hierarchy of the original blocks
Y
Yancey1989 已提交
428
            new_sub_block = program.create_block(lr_block.idx)
Q
Qiyang Min 已提交
429 430 431

            # clone vars
            for var in origin_block.vars:
W
Wu Yi 已提交
432
                new_sub_block._clone_variable(var)
Q
Qiyang Min 已提交
433 434

            # clone ops
Y
Yancey1989 已提交
435 436
            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
437
                # clone sub_block of op
Y
Yancey1989 已提交
438
                __clone_lr_op_sub_block__(cloned_op, program, new_sub_block)
Q
Qiyang Min 已提交
439 440 441 442

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

443
        # append lr decay ops to the child block if exists
444
        lr_ops = self._get_lr_ops()
445 446
        # record optimize blocks and we can run them on pserver parallel
        optimize_blocks = []
447
        if len(lr_ops) > 0:
Q
qiaolongfei 已提交
448 449
            lr_decay_block = pserver_program.create_block(
                pserver_program.num_blocks - 1)
450
            optimize_blocks.append(lr_decay_block)
451
            for _, op in enumerate(lr_ops):
Y
Yancey1989 已提交
452
                cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op)
Q
Qiyang Min 已提交
453
                # append sub blocks to pserver_program in lr_decay_op
Y
Yancey1989 已提交
454 455
                __clone_lr_op_sub_block__(cloned_op, pserver_program,
                                          lr_decay_block)
456

T
typhoonzero 已提交
457
        # append op to the current block
Q
qiaolongfei 已提交
458
        grad_to_block_id = []
Q
qiaolongfei 已提交
459
        pre_block_idx = pserver_program.num_blocks - 1
T
typhoonzero 已提交
460
        for idx, opt_op in enumerate(opt_op_on_pserver):
461
            per_opt_block = pserver_program.create_block(pre_block_idx)
462
            optimize_blocks.append(per_opt_block)
463
            # append grad merging ops before clip and weight decay
T
typhoonzero 已提交
464 465
            # cases may like: 
            # L2Decay op -> clip op -> optimize
466 467 468 469 470 471 472
            for _, op in enumerate(self.optimize_ops):
                # find the origin @GRAD var before clipping
                grad_varname_for_block = __op_have_grad_input__(op)
                if ufind.is_connected(op, opt_op) and grad_varname_for_block:
                    merged_var = self._append_pserver_grad_merge_ops(
                        per_opt_block, grad_varname_for_block, endpoint,
                        grad_to_block_id, self.origin_program)
T
typhoonzero 已提交
473
                    break  # append optimize op once then append other ops.
T
typhoonzero 已提交
474 475
            for _, op in enumerate(self.optimize_ops):
                # optimizer is connected to itself
476
                if ufind.is_connected(op, opt_op) and op not in global_ops:
477
                    __append_optimize_op__(op, per_opt_block, grad_to_block_id,
Y
wip  
yi.wu 已提交
478
                                           merged_var, lr_ops)
T
typhoonzero 已提交
479

W
Wu Yi 已提交
480 481
        # dedup grad to ids list
        grad_to_block_id = list(set(grad_to_block_id))
T
typhoonzero 已提交
482
        # append global ops
483
        if global_ops:
Q
qiaolongfei 已提交
484 485
            opt_state_block = pserver_program.create_block(
                pserver_program.num_blocks - 1)
486
            optimize_blocks.append(opt_state_block)
Q
qiaolongfei 已提交
487
            for glb_op in global_ops:
X
Xi Chen 已提交
488
                __append_optimize_op__(glb_op, opt_state_block,
Y
wip  
yi.wu 已提交
489
                                       grad_to_block_id, None, lr_ops)
T
typhoonzero 已提交
490

491
        # process distributed lookup_table
Q
qiaolongfei 已提交
492
        prefetch_var_name_to_block_id = []
493 494
        if self.has_distributed_lookup_table:
            pserver_index = self.pserver_endpoints.index(endpoint)
495
            table_opt_block = self._create_table_optimize_block(
496
                pserver_index, pserver_program, pre_block_idx, grad_to_block_id)
Q
qiaolongfei 已提交
497
            prefetch_var_name_to_block_id = self._create_prefetch_block(
498
                pserver_index, pserver_program, table_opt_block)
T
tangwei12 已提交
499 500
            checkpoint_block_id = self._create_checkpoint_save_block(
                pserver_program, table_opt_block.idx)
501 502 503 504

        # NOTE: if has_distributed_lookup_table is False, then prefetch_block will
        # not be executed, so it's safe to use optimize_block to hold the place
        if self.has_distributed_lookup_table:
Q
qiaolongfei 已提交
505
            assert len(prefetch_var_name_to_block_id) > 0
506
        else:
Q
qiaolongfei 已提交
507
            assert len(prefetch_var_name_to_block_id) == 0
508

509
        attrs = {
510
            "optimize_blocks": optimize_blocks,
511 512 513
            "endpoint": endpoint,
            "Fanin": self.trainer_num,
            "sync_mode": self.sync_mode,
Y
Yancey1989 已提交
514
            "grad_to_block_id": grad_to_block_id,
515 516 517 518
        }
        if len(prefetch_var_name_to_block_id) > 0:
            attrs['prefetch_var_name_to_block_id'] \
                = prefetch_var_name_to_block_id
T
tangwei12 已提交
519
            attrs['checkpint_block_id'] = checkpoint_block_id
520

T
typhoonzero 已提交
521 522 523 524 525
        # step5 append the listen_and_serv op
        pserver_program.global_block().append_op(
            type="listen_and_serv",
            inputs={'X': recv_inputs},
            outputs={},
526
            attrs=attrs)
527

W
Wu Yi 已提交
528
        pserver_program._sync_with_cpp()
T
typhoonzero 已提交
529 530 531 532 533 534 535
        return pserver_program

    def get_startup_program(self, endpoint, pserver_program):
        """
        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
536 537 538 539 540

        Args:
            endpoint (str): current pserver endpoint.
            pserver_program (Program): call get_pserver_program first and
                pass the result here.
541

Y
yi.wu 已提交
542 543
        Returns:
            Program: parameter server side startup program.
T
typhoonzero 已提交
544 545
        """
        s_prog = Program()
T
typhoonzero 已提交
546
        orig_s_prog = default_startup_program()
T
typhoonzero 已提交
547 548 549 550 551 552 553 554 555 556 557 558 559
        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
        created_var_map = dict()
        for _, var in pserver_vars.iteritems():
W
Wu Yi 已提交
560
            tmpvar = s_prog.global_block()._clone_variable(var)
T
typhoonzero 已提交
561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577
            created_var_map[var.name] = tmpvar

        # 2. rename op outputs
        for op in orig_s_prog.global_block().ops:
            new_outputs = dict()
            # do not append startup op if var is not on this pserver
            op_on_pserver = False
            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]]

            if op_on_pserver:
578 579 580
                # most startup program ops have no inputs
                new_inputs = self._get_input_map_from_op(pserver_vars, op)

T
typhoonzero 已提交
581 582 583 584 585 586 587 588 589 590 591
                if op.type in [
                        "gaussian_random", "fill_constant", "uniform_random"
                ]:
                    op.attrs["shape"] = new_outputs["Out"].shape
                s_prog.global_block().append_op(
                    type=op.type,
                    inputs=new_inputs,
                    outputs=new_outputs,
                    attrs=op.attrs)
        return s_prog

592 593
    # ====================== private transpiler functions =====================

Y
yi.wu 已提交
594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654
    def _has_distributed_lookup_table(self):
        # process lookup_table_op
        # 1. check all lookup_table_op is distributed
        # 2. check all lookup_table_op share the same table.
        distributed_lookup_table_ops = []
        # support only one distributed_lookup_table now
        self.table_name = None
        for op in self.origin_program.global_block().ops:
            if op.type == LOOKUP_TABLE_TYPE:
                if op.attrs['is_distributed'] is True:
                    if self.table_name is None:
                        self.table_name = op.input("W")[0]
                    if self.table_name != op.input("W")[0]:
                        raise RuntimeError("all distributed lookup_table_ops"
                                           " should have only one table")
                    distributed_lookup_table_ops.append(op)
                else:
                    if self.table_name is not None:
                        assert op.input("W")[0] != self.table_name

        return len(distributed_lookup_table_ops) > 0

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

G
gongweibao 已提交
655
    def _init_splited_vars(self):
Y
yi.wu 已提交
656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678
        # 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
679
        if self.config.slice_var_up:
Y
yi.wu 已提交
680 681
            # 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
682 683 684
            grad_blocks = slice_variable(grad_list,
                                         len(self.pserver_endpoints),
                                         self.config.min_block_size)
Y
yi.wu 已提交
685
            param_blocks = slice_variable(param_list,
G
gongweibao 已提交
686 687
                                          len(self.pserver_endpoints),
                                          self.config.min_block_size)
Y
yi.wu 已提交
688 689 690
        else:
            # when we do NOT slice var up into blocks, we will always slice params
            # grads into one block.
G
gongweibao 已提交
691 692 693 694
            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
695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721
        assert (len(grad_blocks) == len(param_blocks))

        # origin_varname -> [splited_var]
        self.param_var_mapping = self._create_vars_from_blocklist(
            self.origin_program, param_blocks)
        self.grad_var_mapping = self._create_vars_from_blocklist(
            self.origin_program,
            grad_blocks,
            add_trainer_suffix=self.trainer_num > 1)
        self.grad_param_mapping = dict()
        for g, p in zip(grad_blocks, param_blocks):
            g_name, g_bid, _ = g.split(":")
            p_name, p_bid, _ = p.split(":")
            self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] =  \
                    self.param_var_mapping[p_name][int(p_bid)]

        # create mapping of endpoint -> split var to create pserver side program
        self.param_grad_ep_mapping = dict()
        [
            self.param_grad_ep_mapping.update({
                ep: {
                    "params": [],
                    "grads": []
                }
            }) for ep in self.pserver_endpoints
        ]

722
    # transpiler function for dis lookup_table
Q
update  
qiaolongfei 已提交
723 724
    def _replace_lookup_table_op_with_prefetch(self, program,
                                               pserver_endpoints):
725
        # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
Q
qiaolongfei 已提交
726 727 728 729 730 731 732 733 734
        # self.all_prefetch_input_vars =
        #       [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
        #        [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
        self.all_prefetch_input_vars = []

        # self.all_prefetch_input_vars =
        #       [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
        #        [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
        self.all_prefetch_output_vars = []
735 736 737 738 739 740 741 742 743

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

744
                    lookup_table_op_index = list(all_ops).index(op)
745 746 747
                    ids_name = op.input("Ids")
                    out_name = op.output("Out")

Q
qiaolongfei 已提交
748 749 750 751 752 753 754 755 756 757 758 759 760
                    ids_var = program.global_block().vars[ids_name[0]]
                    prefetch_input_vars = self.create_splited_vars(
                        source_var=ids_var,
                        block=program.global_block(),
                        tag="_prefetch_in_")
                    self.all_prefetch_input_vars.append(prefetch_input_vars)

                    out_var = program.global_block().vars[out_name[0]]
                    prefetch_output_vars = self.create_splited_vars(
                        source_var=out_var,
                        block=program.global_block(),
                        tag="_prefetch_out_")
                    self.all_prefetch_output_vars.append(prefetch_output_vars)
761 762

                    # insert split_ids_op
W
Wu Yi 已提交
763
                    program.global_block()._insert_op(
764
                        index=lookup_table_op_index,
765 766 767 768 769 770 771
                        type="split_ids",
                        inputs={
                            'Ids': [
                                program.global_block().vars[varname]
                                for varname in ids_name
                            ]
                        },
Q
qiaolongfei 已提交
772
                        outputs={"Out": prefetch_input_vars})
773 774

                    # insert prefetch_op
W
Wu Yi 已提交
775
                    program.global_block()._insert_op(
776
                        index=lookup_table_op_index + 1,
777
                        type="prefetch",
Q
qiaolongfei 已提交
778 779
                        inputs={'X': prefetch_input_vars},
                        outputs={"Out": prefetch_output_vars},
Y
Yancey1989 已提交
780
                        attrs={
781
                            "epmap": pserver_endpoints,
782 783 784
                            # FIXME(qiao) temporarily disable this config because prefetch
                            # is not act as other rpc op, it's more like a forward op
                            # RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
Y
Yancey1989 已提交
785
                        })
786 787

                    # insert concat_op
W
Wu Yi 已提交
788
                    program.global_block()._insert_op(
789 790 791 792 793 794 795
                        index=lookup_table_op_index + 2,
                        type="merge_ids",
                        inputs={
                            'Ids': [
                                program.global_block().vars[varname]
                                for varname in ids_name
                            ],
796
                            'X': prefetch_output_vars
797
                        },
798 799 800 801 802
                        outputs={
                            "Out": [
                                program.global_block().vars[varname]
                                for varname in out_name
                            ]
803
                        })
804 805

                    # delete lookup_table_op
806
                    delete_ops(program.global_block(), [op])
807 808 809
                    # break for loop
                    break

Y
Yancey1989 已提交
810
    def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
811
        # 2. add split_ids_op and send_op to send gradient to pservers
812 813
        # there should only be one table_name
        all_ops = program.global_block().ops
T
typhoonzero 已提交
814
        table_grad_name = grad_var_name(self.table_name)
815 816 817 818
        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
819
                program.global_block()._insert_op(
820 821 822 823 824
                    index=op_index + 1,
                    type="split_ids",
                    inputs={
                        'Ids': [program.global_block().vars[table_grad_name]]
                    },
825
                    outputs={"Out": self.trainer_side_table_grad_list})
W
Wu Yi 已提交
826
                program.global_block()._insert_op(
827
                    index=op_index + 2,
828
                    type="send",
829
                    inputs={'X': self.trainer_side_table_grad_list},
Y
Yancey1989 已提交
830 831
                    outputs={},
                    attrs={
832
                        "sync_mode": True,
Y
Yancey1989 已提交
833 834 835
                        "epmap": pserver_endpoints,
                        RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                    })
836 837 838 839 840 841
                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
842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869
        prefetch_var_name_to_block_id = []
        for index in range(len(self.all_prefetch_input_vars)):
            prefetch_block = pserver_program.create_block(optimize_block.idx)
            trainer_ids = self.all_prefetch_input_vars[index][pserver_index]
            pserver_ids = pserver_program.global_block().create_var(
                name=trainer_ids.name,
                type=trainer_ids.type,
                shape=trainer_ids.shape,
                dtype=trainer_ids.dtype)
            trainer_out = self.all_prefetch_output_vars[index][pserver_index]
            pserver_out = pserver_program.global_block().create_var(
                name=trainer_out.name,
                type=trainer_out.type,
                shape=trainer_out.shape,
                dtype=trainer_out.dtype)
            prefetch_block.append_op(
                type="lookup_sparse_table",
                inputs={'Ids': pserver_ids,
                        "W": table_var},
                outputs={"Out": pserver_out},
                attrs={
                    "is_sparse": True,  # has no effect on lookup_table op
                    "is_distributed": True,
                    "padding_idx": -1
                })
            prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str(
                prefetch_block.idx))
        return prefetch_var_name_to_block_id
870 871

    def _create_table_optimize_block(self, pserver_index, pserver_program,
872
                                     pre_block_idx, grad_to_block_id):
873 874
        # STEP: create table optimize block
        # create table param and grad var in pserver program
Y
Yancey1989 已提交
875 876 877 878 879 880 881 882
        origin_param_var = self.origin_program.global_block().vars[
            self.table_name]
        param_var = pserver_program.global_block().create_var(
            name=origin_param_var.name,
            shape=origin_param_var.shape,
            dtype=origin_param_var.dtype,
            type=core.VarDesc.VarType.SELECTED_ROWS,
            persistable=True)
883 884
        # parameter must be selected rows
        param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS)
W
Wu Yi 已提交
885
        grad_var = pserver_program.global_block()._clone_variable(
T
typhoonzero 已提交
886
            self.origin_program.global_block().vars[grad_var_name(
887
                self.table_name)])
888 889 890 891

        # create table optimize block in pserver program
        table_opt_op = [
            op for op in self.optimize_ops
892 893
            if 'Param' in op.input_names and op.input("Param")[0] ==
            self.table_name
894
        ][0]
Q
qiaolongfei 已提交
895
        table_opt_block = pserver_program.create_block(pre_block_idx)
896 897 898
        # only support sgd now
        assert table_opt_op.type == "sgd"

899 900 901
        if self.sync_mode:
            # create grad vars in pserver program
            table_grad_var = self.table_param_grad[1]
902
            pserver_side_table_grad_list = [
903 904 905 906 907 908 909 910 911
                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)
            ]

912
            # append sum op for pserver_side_table_grad_list
913 914
            table_opt_block.append_op(
                type="sum",
915
                inputs={"X": pserver_side_table_grad_list},
916 917
                outputs={"Out": [grad_var]},
                attrs={"use_mkldnn": False})
918 919
        else:
            # in async_mode, for table gradient, it also need to be splited to each parameter server
920
            origin_grad_name = grad_var.name
921 922
            splited_grad_name = self.trainer_side_table_grad_list[
                pserver_index].name
923 924
            if not splited_grad_name.startswith(origin_grad_name):
                raise ValueError("origin_grad_var: " + splited_grad_name +
925
                                 " grad_var:" + grad_var.name)
W
Wu Yi 已提交
926
            grad_var = pserver_program.global_block()._rename_var(
927
                origin_grad_name, splited_grad_name)
928 929 930 931 932 933 934 935 936 937 938 939 940 941 942

        lr_var = pserver_program.global_block().vars[table_opt_op.input(
            "LearningRate")[0]]
        inputs = {
            "Param": [param_var],
            "Grad": [grad_var],
            "LearningRate": [lr_var]
        }
        outputs = {"ParamOut": [param_var]}
        table_opt_block.append_op(
            type=table_opt_op.type,
            inputs=inputs,
            outputs=outputs,
            attrs=table_opt_op.attrs)

943 944 945
        # 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))

946 947
        return table_opt_block

T
tangwei12 已提交
948 949 950 951 952 953
    def _create_checkpoint_save_block(self, pserver_program, pre_block_idx):
        """
        create a new block to handle save checkpoint.
        """
        import os

T
tangwei12 已提交
954
        pserver_program.global_block().create_var(
T
tangwei12 已提交
955
            name="kLookupTablePath",
T
tangwei12 已提交
956 957
            persistable=True,
            type=core.VarDesc.VarType.RAW)
T
tangwei12 已提交
958

T
tangwei12 已提交
959
        checkpoint_save_block = pserver_program.create_block(pre_block_idx)
T
tangwei12 已提交
960
        # this 'file_path' do not be used in save lookup table variable
T
tangwei12 已提交
961 962 963 964
        checkpoint_save_block.append_op(
            type='save',
            inputs={'X': [self.table_name]},
            outputs={},
T
tangwei12 已提交
965
            attrs={'file_path': "none"})
T
tangwei12 已提交
966 967 968

        return checkpoint_save_block.idx

T
typhoonzero 已提交
969 970 971 972 973
    def _create_vars_from_blocklist(self,
                                    program,
                                    block_list,
                                    add_trainer_suffix=False):
        """
974
        Create vars for each split.
T
typhoonzero 已提交
975 976
        NOTE: only grads need to be named for different trainers, use
              add_trainer_suffix to rename the grad vars.
977 978 979 980
        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.
981 982
        Returns:
            var_mapping (dict(varname->[new_varname_variable])):A dict mapping
983
                from original var name to each var split.
T
typhoonzero 已提交
984
        """
985 986

        # varname->[(block_id, current_block_size)]
T
typhoonzero 已提交
987
        block_map = dict()
988

T
typhoonzero 已提交
989
        var_mapping = dict()
T
typhoonzero 已提交
990 991 992 993 994
        for block_str in block_list:
            varname, offset, size = block_str.split(":")
            if not block_map.has_key(varname):
                block_map[varname] = []
            block_map[varname].append((long(offset), long(size)))
Y
yi.wu 已提交
995

T
typhoonzero 已提交
996
        for varname, splited in block_map.iteritems():
T
typhoonzero 已提交
997
            orig_var = program.global_block().var(varname)
T
typhoonzero 已提交
998
            if len(splited) == 1:
999
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1000 1001
                    new_var_name = "%s.trainer_%d" % \
                        (orig_var.name, self.trainer_id)
W
Wu Yi 已提交
1002
                    program.global_block()._rename_var(varname, new_var_name)
T
typhoonzero 已提交
1003 1004 1005 1006 1007
                    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
1008
                continue
T
typhoonzero 已提交
1009 1010

            var_mapping[varname] = []
T
typhoonzero 已提交
1011 1012 1013 1014
            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
1015

T
typhoonzero 已提交
1016
            for i, block in enumerate(splited):
T
typhoonzero 已提交
1017
                size = block[1]
T
typhoonzero 已提交
1018 1019 1020 1021
                rows = size / orig_dim1_flatten
                splited_shape = [rows]
                if len(orig_shape) >= 2:
                    splited_shape.extend(orig_shape[1:])
T
typhoonzero 已提交
1022
                new_var_name = ""
1023
                if self.sync_mode and add_trainer_suffix:
T
typhoonzero 已提交
1024 1025 1026 1027 1028
                    new_var_name = "%s.block%d.trainer_%d" % \
                        (varname, i, self.trainer_id)
                else:
                    new_var_name = "%s.block%d" % \
                        (varname, i)
T
typhoonzero 已提交
1029
                var = program.global_block().create_var(
T
typhoonzero 已提交
1030 1031
                    name=new_var_name,
                    persistable=False,
T
typhoonzero 已提交
1032
                    dtype=orig_var.dtype,
1033
                    type=orig_var.type,
T
typhoonzero 已提交
1034
                    shape=splited_shape)  # flattend splited var
T
typhoonzero 已提交
1035
                var_mapping[varname].append(var)
W
Wu Yi 已提交
1036
            program.global_block()._sync_with_cpp()
T
typhoonzero 已提交
1037
        return var_mapping
T
done  
typhoonzero 已提交
1038

1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049
    def create_splited_vars(self, source_var, block, tag):
        return [
            block.create_var(
                name=str(source_var.name + tag + str(index)),
                type=source_var.type,
                shape=source_var.shape,
                dtype=source_var.dtype)
            for index in range(len(self.pserver_endpoints))
        ]

    def _clone_var(self, block, var, persistable=True):
T
done  
typhoonzero 已提交
1050 1051 1052 1053 1054 1055
        return block.create_var(
            name=var.name,
            shape=var.shape,
            dtype=var.dtype,
            type=var.type,
            lod_level=var.lod_level,
1056
            persistable=persistable)
T
done  
typhoonzero 已提交
1057

Y
Yancey1989 已提交
1058
    def _insert_split_op(self, program, orig_var, index, splited_vars):
Y
update  
Yancey1989 已提交
1059 1060 1061 1062
        if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS:
            height_sections = []
            for v in splited_vars:
                height_sections.append(v.shape[0])
W
Wu Yi 已提交
1063
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1064 1065 1066 1067 1068 1069 1070 1071 1072
                index=index + 1,
                type="split_selected_rows",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
                attrs={"height_sections": height_sections})
        elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR:
            sections = []
            for v in splited_vars:
                sections.append(v.shape[0])
W
Wu Yi 已提交
1073
            program.global_block()._insert_op(
Y
update  
Yancey1989 已提交
1074 1075 1076 1077 1078 1079 1080 1081 1082
                index=index + 1,
                type="split_byref",
                inputs={"X": orig_var},
                outputs={"Out": splited_vars},
                attrs={"sections": sections}  # assume split evenly
            )
        else:
            AssertionError("Variable type should be in set "
                           "[LOD_TENSOR, SELECTED_ROWS]")
T
done  
typhoonzero 已提交
1083

T
typhoonzero 已提交
1084 1085 1086 1087
    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
1088
        Param and Grad is split to multiple servers.
T
typhoonzero 已提交
1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110
        """
        # HACK(typhoonzero): Should use functions of corresponding optimizer in
        # optimizer.py to get the shape, do not  bind this in the transpiler.
        if op_type == "adam":
            if varkey in ["Moment1", "Moment2"]:
                return param_shape
        elif op_type == "adagrad":
            if varkey == "Moment":
                return param_shape
        elif op_type == "adamax":
            if varkey in ["Moment", "InfNorm"]:
                return param_shape
        elif op_type == "momentum":
            if varkey == "Velocity":
                return param_shape
        elif op_type == "":
            if varkey == "Moment":
                return param_shape
        elif op_type == "sgd":
            pass
        return orig_shape

1111 1112
    def _get_varname_parts(self, varname):
        # returns origin, blockid, trainerid
T
typhoonzero 已提交
1113
        orig_var_name = ""
1114 1115 1116 1117 1118 1119 1120 1121 1122 1123
        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
1124
        else:
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151
            block_index = len(varname)
        orig_var_name = varname[0:min(block_index, trainer_idx)]
        return orig_var_name, block_part, trainer_part

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

    def _append_pserver_grad_merge_ops(self, optimize_block,
                                       grad_varname_for_block, endpoint,
                                       grad_to_block_id, origin_program):
        program = optimize_block.program
        pserver_block = program.global_block()
        grad_block = None
        for g in self.param_grad_ep_mapping[endpoint]["grads"]:
            if self._orig_varname(g.name) == \
                    self._orig_varname(grad_varname_for_block):
                grad_block = g
                break
        if not grad_block:
            # do not append this op if current endpoint
            # is not dealing with this grad block
            return
        orig_varname, block_name, trainer_name = self._get_varname_parts(
            grad_block.name)
        if block_name:
            merged_var_name = '.'.join([orig_varname, block_name])
T
typhoonzero 已提交
1152
        else:
1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166
            merged_var_name = orig_varname
        merged_var = \
            pserver_block.vars[merged_var_name]
        grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx))
        if self.sync_mode and self.trainer_num > 1:
            vars2merge = []
            for i in xrange(self.trainer_num):
                per_trainer_name = "%s.trainer_%d" % \
                (merged_var_name, i)
                vars2merge.append(pserver_block.vars[per_trainer_name])

            optimize_block.append_op(
                type="sum",
                inputs={"X": vars2merge},
1167 1168
                outputs={"Out": merged_var},
                attrs={"use_mkldnn": False})
1169 1170 1171 1172 1173 1174 1175 1176
            # TODO(panyx0718): What if it's SELECTED_ROWS.
            if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS:
                optimize_block.append_op(
                    type="scale",
                    inputs={"X": merged_var},
                    outputs={"Out": merged_var},
                    attrs={"scale": 1.0 / float(self.trainer_num)})
        return merged_var
T
typhoonzero 已提交
1177

1178
    def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
1179
                            grad_to_block_id, origin_program, merged_var):
1180
        program = optimize_block.program
T
typhoonzero 已提交
1181
        pserver_block = program.global_block()
T
typhoonzero 已提交
1182
        new_inputs = dict()
T
typhoonzero 已提交
1183 1184
        # update param/grad shape first, then other inputs like
        # moment can use the updated shape
T
typhoonzero 已提交
1185
        for key in opt_op.input_names:
T
typhoonzero 已提交
1186 1187 1188 1189 1190 1191
            if key == "Grad":
                new_inputs[key] = merged_var
            elif key == "Param":
                # param is already created on global program
                param_block = None
                for p in self.param_grad_ep_mapping[endpoint]["params"]:
T
typhoonzero 已提交
1192
                    if same_or_split_var(p.name, opt_op.input(key)[0]):
T
typhoonzero 已提交
1193 1194 1195 1196
                        param_block = p
                        break
                if not param_block:
                    return
T
typhoonzero 已提交
1197
                tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1198
                    name=param_block.name,
T
typhoonzero 已提交
1199
                    persistable=True,
T
typhoonzero 已提交
1200 1201 1202
                    dtype=param_block.dtype,
                    shape=param_block.shape)
                new_inputs[key] = tmpvar
1203
            elif key == "LearningRate":
1204
                # learning rate variable has already be created by non-optimize op,
1205
                # don't create it once again.
1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216
                lr_varname = opt_op.input(key)[0]
                if pserver_block.vars.has_key(lr_varname):
                    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
1217

T
typhoonzero 已提交
1218
        for key in opt_op.input_names:
1219 1220
            new_shape = None
            if key in ["Param", "Grad", "LearningRate"]:
T
typhoonzero 已提交
1221
                continue
1222
            var = self.origin_program.global_block().vars[opt_op.input(key)[0]]
T
typhoonzero 已提交
1223 1224 1225 1226
            # 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
1227
            tmpvar = pserver_block.create_var(
T
typhoonzero 已提交
1228 1229 1230 1231 1232
                name=var.name,
                persistable=var.persistable,
                dtype=var.dtype,
                shape=new_shape)
            new_inputs[key] = tmpvar
T
typhoonzero 已提交
1233

1234
        # change output's ParamOut variable
1235 1236
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1237
        outputs["ParamOut"] = new_inputs["Param"]
T
typhoonzero 已提交
1238

1239
        optimize_block.append_op(
T
typhoonzero 已提交
1240 1241
            type=opt_op.type,
            inputs=new_inputs,
T
typhoonzero 已提交
1242
            outputs=outputs,
T
typhoonzero 已提交
1243 1244
            attrs=opt_op.attrs)

1245 1246 1247 1248 1249 1250 1251 1252 1253
    def _is_splited_grad_var(self, var, var_dict):
        grad_block = None
        for _, g in var_dict.iteritems():
            if self._orig_varname(g.name) == self._orig_varname(var.name):
                if g.name.find(".trainer_") == -1:
                    grad_block = g
                    break
        return grad_block

Q
Qiyang Min 已提交
1254 1255 1256 1257 1258 1259 1260 1261
    def _clone_lr_op(self, program, block, op):
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, op)
        for key, varlist in inputs.iteritems():
            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
1262
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1263 1264 1265 1266 1267 1268 1269 1270

        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, op)
        for key, varlist in outputs.iteritems():
            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
1271
                    block._clone_variable(var)
Q
Qiyang Min 已提交
1272

Y
Yancey1989 已提交
1273
        return block.append_op(
Q
Qiyang Min 已提交
1274 1275 1276
            type=op.type, inputs=inputs, outputs=outputs, attrs=op.attrs)

    def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
1277
        program = optimize_block.program
1278
        # Append the ops for parameters that do not need to be optimized/updated
1279 1280
        inputs = self._get_input_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1281
        for key, varlist in inputs.iteritems():
1282 1283
            if not isinstance(varlist, list):
                varlist = [varlist]
T
typhoonzero 已提交
1284
            for var in varlist:
1285 1286 1287 1288 1289 1290 1291
                # for ops like clipping and weight decay, get the splited var
                # for inputs/outputs
                grad_block = self._is_splited_grad_var(
                    var, program.global_block().vars)
                if grad_block:
                    inputs[key] = grad_block
                elif not program.global_block().vars.has_key(var.name):
1292
                    program.global_block().create_var(
T
typhoonzero 已提交
1293 1294 1295 1296 1297
                        name=var.name,
                        persistable=var.persistable,
                        dtype=var.dtype,
                        shape=var.shape)

1298 1299
        outputs = self._get_output_map_from_op(
            self.origin_program.global_block().vars, opt_op)
1300
        for key, varlist in outputs.iteritems():
1301 1302 1303
            if not isinstance(varlist, list):
                varlist = [varlist]
            for var in varlist:
1304 1305 1306 1307 1308
                grad_block = self._is_splited_grad_var(
                    var, program.global_block().vars)
                if grad_block:
                    outputs[key] = grad_block
                elif not program.global_block().vars.has_key(var.name):
W
Wu Yi 已提交
1309
                    program.global_block()._clone_variable(var)
1310

Y
Yancey1989 已提交
1311
        return optimize_block.append_op(
T
typhoonzero 已提交
1312
            type=opt_op.type,
T
typhoonzero 已提交
1313 1314
            inputs=inputs,
            outputs=outputs,
T
typhoonzero 已提交
1315 1316
            attrs=opt_op.attrs)

1317 1318 1319 1320
    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
1321 1322
        if set(op1.desc.output_arg_names()) & set(op2.desc.input_arg_names()) or \
           set(op1.desc.input_arg_names()) & set(op2.desc.output_arg_names()):
1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336
            return True
        return False

    def _create_ufind(self, optimize_ops):
        # Create a unit find data struct by optimize ops
        ufind = UnionFind(optimize_ops)
        for i in xrange(len(optimize_ops)):
            for j in xrange(i, len(optimize_ops)):
                op1 = optimize_ops[i]
                op2 = optimize_ops[j]
                if self._is_op_connected(op1, op2):
                    ufind.union(op1, op2)
        return ufind

1337
    def _is_optimizer_op(self, op):
T
typhoonzero 已提交
1338 1339
        if "Param" in op.input_names and \
            "LearningRate" in op.input_names:
1340 1341 1342 1343 1344 1345 1346
            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
1347
        if op.input("Param")[0] in param_names:
1348 1349 1350
            return True
        else:
            for n in param_names:
T
typhoonzero 已提交
1351
                param = op.input("Param")[0]
T
typhoonzero 已提交
1352
                if same_or_split_var(n, param) and n != param:
1353 1354 1355
                    return True
            return False

T
typhoonzero 已提交
1356
    def _get_input_map_from_op(self, varmap, op):
1357
        """Returns a dict from op input name to the vars in varmap."""
T
typhoonzero 已提交
1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369
        iomap = dict()
        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):
1370
        """Returns a dict from op output name to the vars in varmap."""
T
typhoonzero 已提交
1371 1372 1373 1374 1375 1376 1377 1378 1379 1380
        iomap = dict()
        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
1381 1382 1383 1384 1385 1386

    def _get_lr_ops(self):
        lr_ops = []
        # find learning rate variables by optimize op
        lr_vars = set()
        for op in self.optimize_ops:
1387
            if self._is_optimizer_op(op):
1388 1389 1390 1391
                lr_vars.add(op.input("LearningRate")[0])

        find_ops = []
        # find ops which output is lr var
1392
        block = self.origin_program.global_block()
1393 1394 1395 1396 1397
        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)
1398

1399 1400 1401 1402 1403
        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 \
1404
                    not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2):
1405 1406 1407 1408 1409 1410
                    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)
1411 1412
                    # we only need to append op for once
                    break
1413
        return lr_ops
Y
Yancey1989 已提交
1414

W
Wu Yi 已提交
1415 1416 1417 1418 1419 1420 1421 1422 1423 1424
    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
        if op_maker.kOpRoleAttrName() in op.attrs and \
            int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role):
            return True
        return False

Y
Yancey1989 已提交
1425
    def _get_optimize_pass(self):
1426
        """
1427
        Get optimizer operators, parameters and gradients from origin_program
1428 1429 1430 1431
        Returns:
            opt_ops (list): optimize operators.
            params_grads (dict): paramter->gradient.
        """
Y
Yancey1989 已提交
1432 1433 1434
        block = self.origin_program.global_block()
        opt_ops = []
        params_grads = []
1435
        origin_var_dict = self.origin_program.global_block().vars
Y
Yancey1989 已提交
1436
        for op in block.ops:
W
Wu Yi 已提交
1437
            if self._is_opt_role_op(op):
Y
Yancey1989 已提交
1438
                opt_ops.append(op)
1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449
                # HACK(wuyi): if we find grad vars from input of optimize
                # ops, we may get the output of clip op. Use syntax "@GRAD"
                # and op_role_var to get the pair.
                for input_name in op.input_arg_names:
                    if input_name.find("@GRAD") != -1 and \
                        op.attrs[RPC_OP_ROLE_ATTR_NAME]:
                        param_name = op.attrs[OP_ROLE_VAR_ATTR_NAME][0]
                        params_grads.append([
                            origin_var_dict[param_name],
                            origin_var_dict[input_name]
                        ])
Y
Yancey1989 已提交
1450 1451 1452
            else:
                pass
        return opt_ops, params_grads