public.py

# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import print_function
from functools import reduce

import collections
import math
import os
import warnings

import six
import paddle.fluid as fluid
from paddle.fluid import core
from paddle.fluid.core import CommContext
import paddle.fluid.framework as framework
from paddle.fluid.incubate.fleet.parameter_server.mode import DistributedMode
from paddle.fluid.incubate.fleet.parameter_server.ir import vars_metatools
from paddle.fluid.incubate.fleet.parameter_server.ir.ps_dispatcher import RoundRobin, PSDispatcher
from paddle.fluid.transpiler.details.program_utils import delete_ops

OP_NAME_SCOPE = "op_namescope"
CLIP_OP_NAME_SCOPE = "@CLIP"
STEP_COUNTER = "@PS_STEP_COUNTER@"
LEARNING_RATE_DECAY_COUNTER = "@LR_DECAY_COUNTER@"

OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
RPC_OP_ROLE_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleAttrName()
RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
LR_SCHED_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.LRSched
OPT_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Optimize

SPARSE_OP_LIST = ["lookup_table", "lookup_table_v2"]
SPARSE_OP_TYPE_DICT = {"lookup_table": "W", "lookup_table_v2": "W"}


def _get_lr_ops(program):
    lr_ops = []
    for index, op in enumerate(program.global_block().ops):
        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):
            lr_ops.append(op)
    return lr_ops


def _has_global_step(lr_ops):
    if len(lr_ops) > 0:
        for idx, op in enumerate(lr_ops):
            if op.type != 'increment':
                continue
            counter = op.input("X")[0]
            if counter == LEARNING_RATE_DECAY_COUNTER:
                return True
    return False


def is_sparse_op(op):
    if op.type in SPARSE_OP_LIST and op.attr('is_sparse') is True and op.attr(
            'is_distributed') is False:
        return True

    if op.type == "distributed_lookup_table" and op.attr(
            'is_distributed') is False:
        return True

    return False


def is_distributed_sparse_op(op):
    if op.type in SPARSE_OP_LIST and op.attr('is_distributed') is True:
        return True

    if op.type == "distributed_lookup_table" and op.attr(
            'is_distributed') is True:
        return True

    return False


def get_sparse_tablename(op):
    return op.input("W")[0]


def get_sparse_tablenames(program, is_distributed):
    tablenames = set()
    if is_distributed:
        for op in program.global_block().ops:
            if is_distributed_sparse_op(op):
                tablenames.add(get_sparse_tablename(op))
    else:
        for op in program.global_block().ops:
            if is_sparse_op(op):
                tablenames.add(get_sparse_tablename(op))
    return list(tablenames)


class MergedVariable:
    def __init__(self, merged, ordered, offsets):
        self.merged_var = merged
        self.ordered_vars = ordered
        self.offsets = offsets


def Singleton(cls):
    _instance = {}

    def _singleton(*args, **kargs):
        if cls not in _instance:
            _instance[cls] = cls(*args, **kargs)
        return _instance[cls]

    return _singleton


@Singleton
class CompileTimeStrategy(object):
    def __init__(self, main_program, startup_program, strategy, role_maker):
        self.min_block_size = 8192

        self.origin_main_program = main_program
        self.origin_startup_program = startup_program
        self.origin_ps_main_program = main_program
        self.origin_ps_startup_program = startup_program

        self.strategy = strategy
        self.role_maker = role_maker

        self.origin_sparse_pairs = []
        self.origin_dense_pairs = []

        self.merged_variables_pairs = []
        self.merged_dense_pairs = []
        self.merged_sparse_pairs = []

        self.merged_variable_map = {}
        self.param_name_to_grad_name = {}
        self.grad_name_to_param_name = {}

        self.param_grad_ep_mapping = collections.OrderedDict()
        self.grad_param_mapping = collections.OrderedDict()

        self._build_var_distributed()

        # for heter-ps save variables
        self.origin_merged_variables_pairs = list(self.merged_variables_pairs)
        self.origin_merged_dense_pairs = list(self.merged_dense_pairs)
        self.origin_merged_sparse_pairs = list(self.merged_sparse_pairs)

    def get_distributed_mode(self):
        trainer = self.strategy.get_trainer_runtime_config()
        return trainer.mode

    def is_sync_mode(self):
        trainer = self.strategy.get_trainer_runtime_config()
        return trainer.mode == DistributedMode.SYNC

    def is_geo_mode(self):
        trainer = self.strategy.get_trainer_runtime_config()
        return trainer.mode == DistributedMode.GEO

    def is_async_mode(self):
        trainer = self.strategy.get_trainer_runtime_config()
        return trainer.mode == DistributedMode.ASYNC

    def get_role_id(self):
        try:
            return self.role_maker._role_id()
        except Exception:
            return self.role_maker.role_id()

    def get_trainers(self):
        try:
            return self.role_maker._worker_num()
        except Exception:
            return self.role_maker.worker_num()

    def get_ps_endpoint(self):
        try:
            return self.role_maker._get_pserver_endpoints()[self.get_role_id()]
        except Exception:
            return self.role_maker.get_pserver_endpoints()[self.get_role_id()]

    def get_ps_endpoints(self):
        try:
            return self.role_maker._get_pserver_endpoints()
        except Exception:
            return self.role_maker.get_pserver_endpoints()

    def get_heter_worker_endpoints(self):
        try:
            return self.role_maker._get_heter_worker_endpoints()
        except Exception:
            return self.role_maker.get_heter_worker_endpoints()

    def get_heter_worker_endpoint(self):
        try:
            return self.role_maker._get_heter_worker_endpoint()
        except Exception:
            return self.role_maker.get_heter_worker_endpoint()

    def get_origin_programs(self):
        return self.origin_main_program, self.origin_startup_program

    def get_origin_main_program(self):
        return self.origin_main_program

    def get_origin_startup_program(self):
        return self.origin_startup_program

    def set_origin_ps_main_program(self, program):
        self.origin_ps_main_program = program

    def set_origin_ps_startup_program(self, program):
        self.origin_ps_startup_program = program

    def get_origin_ps_main_program(self):
        return self.origin_ps_main_program

    def get_origin_ps_startup_program(self):
        return self.origin_ps_startup_program

    def get_sparse_varname_on_ps(self, is_distributed, endpoint=None):
        if not endpoint:
            endpoint = self.get_ps_endpoint()

        varnames = get_sparse_tablenames(self.get_origin_main_program(),
                                         is_distributed)
        ps_sparse_varnames = []
        for varname in varnames:
            tables = self.get_var_distributed(varname, True)
            for i in range(len(tables)):
                table, ep, _ = tables[i]
                if ep == endpoint:
                    ps_sparse_varnames.append(table)
        return ps_sparse_varnames

    def build_ctx(self,
                  vars,
                  mapping,
                  is_grad,
                  is_sparse,
                  is_send,
                  is_distributed=False):
        def get_grad_var_ep(slices):
            names = []
            eps = []
            sections = []

            for slice in slices:
                if self.is_geo_mode():
                    if is_send:
                        names.append("{}.delta".format(slice.name))
                    else:
                        names.append(slice.name)
                elif is_grad and self.is_sync_mode() and self.get_trainers(
                ) > 1:
                    names.append("{}.trainer_{}".format(slice.name,
                                                        self.get_role_id()))
                else:
                    names.append(slice.name)

                sections.append(slice.shape[0])

                for ep, pairs in self.param_grad_ep_mapping.items():
                    params, grads = pairs["params"], pairs["grads"]

                    for var in params + grads:
                        if slice.name == var.name:
                            eps.append(ep)
                            break
            return names, eps, sections

        if isinstance(vars, MergedVariable):
            name = vars.merged_var.name
            slices = mapping[name]
            names, eps, sections = get_grad_var_ep(slices)
            origin_varnames = [var.name for var in vars.ordered_vars]
        else:
            name = vars.name
            slices = mapping[name]
            names, eps, sections = get_grad_var_ep(slices)
            origin_varnames = [vars.name]

        trainer_id = self.get_role_id()
        aggregate = True
        ctx = CommContext(name, names, eps, sections, origin_varnames,
                          trainer_id, aggregate, is_sparse, is_distributed)
        return ctx

    def get_trainer_send_context(self):
        send_ctx = {}
        distibuted_varnames = get_sparse_tablenames(self.origin_main_program,
                                                    True)

        if not self.is_geo_mode():
            for merged in self.merged_dense_pairs:
                grad = merged[1]
                ctx = self.build_ctx(grad, self.grad_var_mapping, True, False,
                                     True)
                send_ctx[ctx.var_name()] = ctx

            for merged in self.merged_sparse_pairs:
                param = merged[0]
                grad = merged[1]

                param_name = param.merged_var.name

                is_distributed = True if param_name in distibuted_varnames else False

                ctx = self.build_ctx(grad, self.grad_var_mapping, True, True,
                                     True, is_distributed)
                send_ctx[ctx.var_name()] = ctx

            if self.is_async_mode():
                name, ctx = self._step_ctx()
                send_ctx[name] = ctx
        else:
            for pairs in self.origin_sparse_pairs:
                param, grad = pairs
                param_name = param.name
                is_distributed = True if param_name in distibuted_varnames else False

                param_ctx = self.build_ctx(param, self.param_var_mapping, False,
                                           True, True, is_distributed)
                grad_ctx = self.build_ctx(grad, self.grad_var_mapping, True,
                                          True, True, is_distributed)

                ctx = CommContext(param_ctx.var_name(),
                                  param_ctx.split_varnames(),
                                  param_ctx.split_endpoints(),
                                  param_ctx.sections(),
                                  grad_ctx.origin_varnames(),
                                  param_ctx.trainer_id(),
                                  param_ctx.aggregate(),
                                  param_ctx.is_sparse(),
                                  param_ctx.is_distributed())

                send_ctx[ctx.var_name()] = ctx
            name, ctx = self._step_ctx()
            send_ctx[name] = ctx
        return send_ctx

    def get_communicator_send_context(self):
        send_ctx = {}
        distibuted_varnames = get_sparse_tablenames(self.origin_main_program,
                                                    True)

        if self.is_geo_mode():
            for pairs in self.merged_dense_pairs:
                param = pairs[0]
                ctx = self.build_ctx(param, self.param_var_mapping, False,
                                     False, True)
                send_ctx[ctx.var_name()] = ctx

            for pairs in self.merged_sparse_pairs:
                param = pairs[0]
                param_name = param.merged_var.name
                is_distributed = True if param_name in distibuted_varnames else False

                ctx = self.build_ctx(param, self.param_var_mapping, False, True,
                                     True, is_distributed)
                send_ctx[ctx.var_name()] = ctx
            name, ctx = self._step_ctx()
            send_ctx[name] = ctx
        else:
            for merged in self.merged_dense_pairs:
                grad = merged[1]
                ctx = self.build_ctx(grad, self.grad_var_mapping, True, False,
                                     True)
                send_ctx[ctx.var_name()] = ctx

            for merged in self.merged_sparse_pairs:
                param, grad = merged
                param_name = param.merged_var.name

                is_distributed = True if param_name in distibuted_varnames else False

                ctx = self.build_ctx(grad, self.grad_var_mapping, True, True,
                                     True, is_distributed)
                send_ctx[ctx.var_name()] = ctx

            name, ctx = self._step_ctx()
            send_ctx[name] = ctx
        return send_ctx

    def get_communicator_recv_context(self,
                                      recv_type=1,
                                      use_origin_program=False):
        # recv_type
        # 1 : DENSE 2. SPARSE 3. DISTRIBUTED 4. ALL
        distibuted_varnames = get_sparse_tablenames(self.origin_main_program,
                                                    True)
        sparse_varnames = []
        for pairs in self.origin_sparse_pairs:
            param, grad = pairs
            sparse_varnames.append(param.name)

        dense_recv_ctx = {}
        sparse_recv_ctx = {}
        distributed_recv_ctx = {}

        variables_pairs = self.merged_variables_pairs if not use_origin_program else self.origin_merged_variables_pairs
        for merged in variables_pairs:
            params = merged[0]
            if params.merged_var.name in sparse_varnames:
                continue

            ctx = self.build_ctx(params, self.param_var_mapping, False, False,
                                 False)
            dense_recv_ctx[ctx.var_name()] = ctx

        for pairs in self.origin_sparse_pairs:
            param, grad = pairs

            if param.name in distibuted_varnames:
                ctx = self.build_ctx(param, self.param_var_mapping, False, True,
                                     False, True)
                distributed_recv_ctx[ctx.var_name()] = ctx
            else:
                ctx = self.build_ctx(param, self.param_var_mapping, False, True,
                                     False, False)
                sparse_recv_ctx[ctx.var_name()] = ctx

        if recv_type == 1:
            return dense_recv_ctx
        if recv_type == 2:
            return sparse_recv_ctx
        if recv_type == 3:
            return distributed_recv_ctx
        if recv_type == 4:
            dense_recv_ctx.update(sparse_recv_ctx)
            dense_recv_ctx.update(distributed_recv_ctx)
            return dense_recv_ctx
        assert ValueError(
            "recv_type can only be 1/2/3/4, 1 : DENSE 2. SPARSE 3. DISTRIBUTED 4. ALL"
        )

    def get_server_runtime_config(self):
        return self.strategy.get_server_runtime_config()

    def get_var_distributed(self, varname, is_param):
        var_distributed = []
        offset = 0
        if is_param:
            params = self.param_var_mapping[varname]
            param_varnames = [var.name for var in params]
            for ep, pairs in self.param_grad_ep_mapping.items():
                for p in pairs["params"]:
                    if p.name in param_varnames:
                        offset += p.shape[0]
                        var_distributed.append((p.name, ep, p.shape[0]))
        else:
            grads = self.grad_var_mapping[varname]
            grad_varnames = [var.name for var in grads]
            for ep, pairs in self.param_grad_ep_mapping.items():
                for g in pairs["grads"]:
                    if g.name in grad_varnames:
                        var_distributed.append((g.name, ep, g.shape[0]))
        return var_distributed

    def _step_ctx(self):
        name = STEP_COUNTER
        trainer_id = self.get_role_id()
        endpoints = self.get_ps_endpoints()
        sections = [1] * len(endpoints)
        names = [name] * len(endpoints)
        ctx = CommContext(name, names, endpoints, sections, [name], trainer_id,
                          True, False, False)
        return name, ctx

    def _create_vars_from_blocklist(self, block_list):
        """
        Create vars for each split.
        NOTE: only grads need to be named for different trainers, use
              add_trainer_suffix to rename the grad vars.
        Args:
            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.
        Returns:
            var_mapping (collections.OrderedDict(varname->[new_varname_variable])):A dict mapping
                from original var name to each var split.
        """

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

        for block_str in block_list:
            varname, offset, size = block_str.split(":")
            if varname not in block_map:
                block_map[varname] = []
            block_map[varname].append((int(offset), int(size)))

        for varname, split in six.iteritems(block_map):
            orig_var = self.merged_variable_map[varname]

            if len(split) == 1:
                var_mapping[varname] = [orig_var]
                self.var_distributed.add_distributed_var(
                    origin_var=orig_var,
                    slice_var=orig_var,
                    block_id=0,
                    offset=0,
                    is_slice=False,
                    vtype="Param")
            else:
                var_mapping[varname] = []
                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:])

                for i, block in enumerate(split):
                    size = block[1]
                    rows = size // orig_dim1_flatten
                    splited_shape = [rows]
                    if len(orig_shape) >= 2:
                        splited_shape.extend(orig_shape[1:])

                    new_var_name = "%s.block%d" % (varname, i)
                    slice_var = vars_metatools.VarStruct(
                        name=new_var_name,
                        shape=splited_shape,
                        dtype=orig_var.dtype,
                        type=orig_var.type,
                        lod_level=orig_var.lod_level,
                        persistable=False)
                    var_mapping[varname].append(slice_var)

                    self.var_distributed.add_distributed_var(
                        origin_var=orig_var,
                        slice_var=slice_var,
                        block_id=i,
                        offset=-1,
                        is_slice=False,
                        vtype="Param")

        return var_mapping

    def _dispatcher(self):
        ps_dispatcher = RoundRobin(self.get_ps_endpoints())
        ps_dispatcher.reset()
        grad_var_mapping_items = list(six.iteritems(self.grad_var_mapping))

        sparse_gradnames = [grad.name for _, grad in self.origin_sparse_pairs]

        for grad_varname, splited_vars in grad_var_mapping_items:
            if grad_varname in sparse_gradnames:
                continue

            send_vars = []
            for _, var in enumerate(splited_vars):
                send_vars.append(var)

            recv_vars = []
            for _, var in enumerate(send_vars):
                recv_vars.append(self.grad_param_mapping[var])

            eps = ps_dispatcher.dispatch(recv_vars)

            for i, ep in enumerate(eps):
                self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
                self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])

        for grad_varname, splited_vars in grad_var_mapping_items:
            if grad_varname not in sparse_gradnames:
                continue

            ps_dispatcher.reset()

            send_vars = []
            for _, var in enumerate(splited_vars):
                send_vars.append(var)

            recv_vars = []
            for _, var in enumerate(send_vars):
                recv_vars.append(self.grad_param_mapping[var])

            eps = ps_dispatcher.dispatch(recv_vars)

            for i, ep in enumerate(eps):
                self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
                self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])

    def _slice_variable(self,
                        var_list,
                        slice_count,
                        min_block_size,
                        uniform=False):
        """
        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
        minimum block size 8K elements (maybe 16bit or 32bit or 64bit).

        Args:
            var_list (list): List of variables.
            slice_count (int): Numel of count that variables will be sliced, which
                could be the pserver services' count.
            min_block_size (int): Minimum split block size.
        Returns:
            blocks (list[(varname, block_id, current_block_size)]): A list
                of VarBlocks. Each VarBlock specifies a shard of the var.
        """
        blocks = []
        for var in var_list:
            if not uniform:
                var_numel = reduce(lambda x, y: x * y, var.shape)

                split_count = 1

                # if min_block_size == -1:
                #     split_count = 1
                # else:
                #     split_count = slice_count
                #     max_pserver_count = int(
                #         math.floor(var_numel / float(min_block_size)))
                #     if max_pserver_count == 0:
                #         max_pserver_count = 1
                #     if max_pserver_count < slice_count:
                #         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
                        # update split_count after aligning
                split_count = int(math.ceil(var_numel / float(block_size)))
                for block_id in range(split_count):
                    curr_block_size = min(block_size, var_numel - (
                        (block_id) * block_size))
                    block = vars_metatools.VarBlock(var.name, block_id,
                                                    curr_block_size)
                    blocks.append(str(block))
            else:
                block_size = var.shape[0] / slice_count
                remainder = var.shape[0] % slice_count

                if block_size == 0:
                    dim0s = [block_size] * remainder
                else:
                    dim0s = [block_size] * slice_count
                for i in range(remainder):
                    dim0s[i] = dim0s[i] + 1

                dim1 = reduce(lambda x, y: x * y, var.shape[1:])

                for block_id in range(len(dim0s)):
                    numel = dim0s[block_id] * dim1
                    block = vars_metatools.VarBlock(var.name, block_id, numel)
                    blocks.append(str(block))
        return blocks

    def _get_param_grad_blocks(self, pairs, min_block_size, uniform=False):
        param_list = []
        grad_list = []
        param_grad_set = set()
        for p, g in pairs:
            # todo(tangwei12) skip parameter marked not trainable
            # if type(p) == Parameter and p.trainable == False:
            # continue
            p = p.merged_var
            g = g.merged_var

            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)

                # 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.
        grad_blocks = self._slice_variable(grad_list,
                                           len(self.get_ps_endpoints()),
                                           min_block_size, uniform)

        param_blocks = self._slice_variable(param_list,
                                            len(self.get_ps_endpoints()),
                                            min_block_size, uniform)
        return param_blocks, grad_blocks

    def _var_slice_and_distribute(self):
        # update these mappings for further transpile:
        # 1. param_var_mapping : param var name->[split params vars]
        # 2. grad_var_mapping : grad var name->[split grads vars]
        # 3. grad_param_mapping : grad.blockx->param.blockx
        # 4. param_grad_ep_mapping : ep->{"params" : [], "grads" : [] }

        dps, dgs = self._get_param_grad_blocks(self.merged_dense_pairs, -1,
                                               False)
        sps, sgs = self._get_param_grad_blocks(self.merged_sparse_pairs,
                                               self.min_block_size, True)

        param_blocks = dps + sps
        grad_blocks = dgs + sgs

        assert (len(grad_blocks) == len(param_blocks))

        # origin_param_name->[splited_param_vars]
        self.param_var_mapping = self._create_vars_from_blocklist(param_blocks)
        self.grad_var_mapping = self._create_vars_from_blocklist(grad_blocks)

        # dict(grad_splited_var->param_splited_var)
        self.grad_param_mapping = collections.OrderedDict()
        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)]

        print_maps = {}
        for k, v in self.grad_param_mapping.items():
            print_maps[str(k)] = str(v)

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

    def _build_var_distributed(self):
        self.var_distributed = vars_metatools.VarsDistributed()

        sparse_pairs, dense_pairs = self.get_param_grads()
        origin_for_sparse = []
        origin_for_dense = []
        param_name_grad_name = dict()
        grad_name_to_param_name = dict()

        for param, grad in sparse_pairs:
            param = vars_metatools.create_var_struct(param)
            grad = vars_metatools.create_var_struct(grad)
            origin_for_sparse.append((param, grad))

        for param, grad in dense_pairs:
            param = vars_metatools.create_var_struct(param)
            grad = vars_metatools.create_var_struct(grad)
            origin_for_dense.append((param, grad))

        for dense_pair in origin_for_dense:
            param, grad = dense_pair

            m_param = MergedVariable(param, [param], [0])
            m_grad = MergedVariable(grad, [grad], [0])
            self.merged_variables_pairs.append((m_param, m_grad))
            self.merged_dense_pairs.append((m_param, m_grad))

        for sparse_pair in origin_for_sparse:
            param, grad = sparse_pair

            m_param = MergedVariable(param, [param], [0])
            m_grad = MergedVariable(grad, [grad], [0])
            self.merged_variables_pairs.append((m_param, m_grad))
            self.merged_sparse_pairs.append((m_param, m_grad))

        for merged in self.merged_variables_pairs:
            m_param, m_grad = merged
            self.merged_variable_map[
                m_param.merged_var.name] = m_param.merged_var
            self.merged_variable_map[m_grad.merged_var.name] = m_grad.merged_var

        param_merges = []
        param_merges.extend(origin_for_sparse)
        param_merges.extend(origin_for_dense)

        for param, grad in param_merges:
            param_name_grad_name[param.name] = grad.name
            grad_name_to_param_name[grad.name] = param.name

        self.origin_sparse_pairs = origin_for_sparse
        self.origin_dense_pairs = origin_for_dense
        self.param_name_to_grad_name = param_name_grad_name
        self.grad_name_to_param_name = grad_name_to_param_name

        sparse_pair_map = collections.OrderedDict()

        for pair in self.origin_sparse_pairs + self.origin_dense_pairs:
            param, grad = pair
            sparse_pair_map[param.name] = str(param)
            sparse_pair_map[grad.name] = str(grad)

        self._var_slice_and_distribute()
        self._dispatcher()

    def get_param_grads(self):
        origin_program = self.origin_main_program

        def _get_params_grads(sparse_varnames):
            block = origin_program.global_block()

            dense_param_grads = []
            sparse_param_grads = []

            optimize_params = set()
            origin_var_dict = origin_program.global_block().vars
            role_id = int(core.op_proto_and_checker_maker.OpRole.Backward)
            for op in block.ops:
                if _is_opt_role_op(op):
                    # delete clip op from opt_ops when run in Parameter Server mode
                    if OP_NAME_SCOPE in op.all_attrs() \
                            and CLIP_OP_NAME_SCOPE in op.attr(OP_NAME_SCOPE):
                        op._set_attr("op_role", role_id)
                        continue
                    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 param_name not in optimize_params:
                            optimize_params.add(param_name)
                            param_grad = (origin_var_dict[param_name],
                                          origin_var_dict[grad_name])

                            if param_name in sparse_varnames:
                                sparse_param_grads.append(param_grad)
                            else:
                                dense_param_grads.append(param_grad)
            return sparse_param_grads, dense_param_grads

        def _get_sparse_varnames():
            varnames = []
            for op in origin_program.global_block().ops:
                if op.type in SPARSE_OP_TYPE_DICT.keys() \
                        and op.attr('remote_prefetch') is True:
                    param_name = op.input(SPARSE_OP_TYPE_DICT[op.type])[0]
                    varnames.append(param_name)

            return list(set(varnames))

        sparse_varnames = _get_sparse_varnames()
        sparse_param_grads, dense_param_grads = _get_params_grads(
            sparse_varnames)

        return sparse_param_grads, dense_param_grads

    def remove_var_pair_by_grad(self, var_name):

        for index, pair in enumerate(self.merged_variables_pairs):
            var = pair[0]
            var_grad = pair[1]
            if var_grad.merged_var.name == var_name:
                del self.merged_variables_pairs[index]

        for index, pair in enumerate(self.merged_dense_pairs):
            var = pair[0]
            var_grad = pair[1]
            if var_grad.merged_var.name == var_name:
                del self.merged_dense_pairs[index]
                return

        for index, pair in enumerate(self.merged_sparse_pairs):
            var = pair[0]
            var_grad = pair[1]
            if var_grad.merged_var.name == var_name:
                del self.merged_sparse_pairs[index]
                return

        print("Not find {} in self.merge_pairs".format(var_name))


def _is_opt_role_op(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.attr_names and \
            int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(optimize_role):
        return True
    return False


def _get_optimize_ops(_program):
    block = _program.global_block()
    opt_ops = []
    for op in block.ops:
        if _is_opt_role_op(op):
            # delete clip op from opt_ops when run in Parameter Server mode
            if OP_NAME_SCOPE in op.all_attrs() \
                    and CLIP_OP_NAME_SCOPE in op.attr(OP_NAME_SCOPE):
                op._set_attr(
                    "op_role",
                    int(core.op_proto_and_checker_maker.OpRole.Backward))
                continue
            opt_ops.append(op)
    return opt_ops


def _get_varname_parts(varname):
    # returns origin, blockid, trainerid
    orig_var_name = ""
    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]
    else:
        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(varname):
    orig, _, _ = _get_varname_parts(varname)
    return orig