hybrid_parallel_util.py 7.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
#   Copyright (c) 2021 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.
import os
import six
import numpy as np
import warnings

from paddle import framework
import paddle
from paddle.fluid import core
22
from paddle.fluid.dygraph.parallel import _split_tensors, sync_params_buffers, build_groups
23
from paddle.fluid.framework import in_dygraph_mode, _in_legacy_dygraph
24
from collections import OrderedDict
25
from .log_util import logger
26

27 28
__all__ = []

29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

def _apply_collective_grads(parameters, comm_group):
    grad_var_set = set()
    grad_vars = []
    sparse_grad_vars = []

    for param in parameters:
        if param.trainable and (param._grad_ivar() is not None):
            g_var = param._grad_ivar()
            assert not g_var._is_sparse(
            ), "Now, it doesn't support sparse parameters"
            grad_vars.append(g_var)
            assert g_var not in grad_var_set
            grad_var_set.add(g_var)

44
    coalesced_grads_and_vars = build_groups(grad_vars, 128 * 1024 * 1024)
45 46 47

    for coalesced_grad, _, _ in coalesced_grads_and_vars:
        # need to div nranks
H
Haohongxiang 已提交
48
        nranks = paddle.distributed.get_world_size(
49 50
        ) if comm_group is None else comm_group.nranks
        div_factor = paddle.to_tensor(nranks, dtype=coalesced_grad.dtype)
51
        paddle.distributed.all_reduce(coalesced_grad, group=comm_group)
52 53 54 55 56 57 58
        paddle.fluid.framework._dygraph_tracer().trace_op(
            type="elementwise_div",
            inputs={'X': coalesced_grad,
                    'Y': div_factor},
            outputs={'Out': coalesced_grad},
            attrs={'axis': -1})

59 60 61
    _split_tensors(coalesced_grads_and_vars)


62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
def _apply_collective_grads_eager(parameters, comm_group):
    grad_var_set = set()
    grad_vars = []

    for param in parameters:
        if param.trainable and (param._grad_ivar() is not None):
            g_var = param._grad_ivar()
            assert not g_var.is_sparse(
            ), "Now, it doesn't support sparse parameters"
            grad_vars.append(g_var)
            assert g_var not in grad_var_set
            grad_var_set.add(g_var)

    coalesced_grads_and_vars = build_groups(grad_vars, 128 * 1024 * 1024)

    div_factor = 1.0 / comm_group.nranks
    for coalesced_grad, _, _ in coalesced_grads_and_vars:
        # need to div nranks 
        coalesced_grad.scale_(div_factor)
        paddle.distributed.all_reduce(coalesced_grad, group=comm_group)

    _split_tensors(coalesced_grads_and_vars)


86
def _broadcast_data_help(data, shape, dtype, hcg):
87 88
    model_parallel_group = hcg.get_model_parallel_group()
    src_rank = hcg.get_model_parallel_group_src_rank()
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
    mp_rank = hcg.get_model_parallel_rank()

    shape_gpu = paddle.to_tensor(shape, dtype="int32")
    paddle.distributed.broadcast(
        shape_gpu,
        src=src_rank,
        group=model_parallel_group,
        use_calc_stream=True)

    if mp_rank != 0:
        input_data = paddle.zeros(shape_gpu, dtype=dtype)
    else:
        input_data = data

    paddle.distributed.broadcast(
        input_data,
        src=src_rank,
        group=model_parallel_group,
        use_calc_stream=True)
108

109 110 111 112

def broadcast_input_data(hcg, *inputs, **kwargs):
    for v in inputs:
        if isinstance(v, core.VarBase):
113
            with framework.no_grad():
114
                _broadcast_data_help(v, v.shape, v.dtype, hcg)
115
        else:
116
            logger.error("it doesn't support data type {}".format(type(v)))
117 118 119 120

    for k, v in kwargs.items():
        if isinstance(v, core.VarBase):
            with framework.no_grad():
121
                _broadcast_data_help(v, v.shape, v.dtype, hcg)
122 123
            kwargs[k] = v
        else:
124
            logger.error("it doesn't support data type {}".format(type(v)))
125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142
    return inputs, kwargs


def broadcast_mp_parameters(model, hcg):
    model_parallel_group = hcg.get_model_parallel_group()
    src_rank = hcg.get_model_parallel_group_src_rank()
    sync_params_buffers(
        model, model_parallel_group, src_rank, is_model_parallel=True)


def broadcast_dp_parameters(model, hcg):
    data_parallel_group = hcg.get_data_parallel_group()
    src_rank = hcg.get_data_parallel_group_src_rank()
    sync_params_buffers(
        model, data_parallel_group, src_rank, is_model_parallel=False)


def fused_allreduce_gradients(parameter_list, hcg):
143 144 145 146 147 148 149 150 151 152 153
    if _in_legacy_dygraph():
        data_parallel_group = None if hcg is None else hcg.get_data_parallel_group(
        )
        logger.debug("dp start fuse allreduce gradients")
        with framework.no_grad():
            _apply_collective_grads(parameter_list, data_parallel_group)
    elif in_dygraph_mode():
        assert hcg is None, "It's not support to use hcg in EagerDygraph now."
        data_parallel_group = paddle.distributed.collective._get_default_group()
        with framework.no_grad():
            _apply_collective_grads_eager(parameter_list, data_parallel_group)
J
JZ-LIANG 已提交
154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196


def sharding_reduce_gradients(parameter_list, hcg):
    # TODO allreduce --> reduce
    # TODO merge grad / nrank with dp 
    logger.debug("sharding start gradients sync")
    with framework.no_grad():

        sharding_nrank = hcg.get_sharding_parallel_group().nranks
        for param in parameter_list:
            if param.trainable and (param._grad_ivar() is not None):

                g_var = param._grad_ivar()

                # need use trace_op to allreduce 
                # paddle.distributed.all_reduce(
                #     g_var, group=hcg.get_sharding_parallel_group(), use_calc_stream=True)
                paddle.fluid.framework._dygraph_tracer().trace_op(
                    type="c_allreduce_sum",
                    inputs={'X': g_var},
                    outputs={'Out': g_var},
                    attrs={
                        'ring_id': hcg.get_sharding_parallel_group().id,
                        'use_calc_stream': True
                    })

                # grad / sharding_rank
                div_factor = paddle.to_tensor(sharding_nrank, dtype=g_var.dtype)
                paddle.fluid.framework._dygraph_tracer().trace_op(
                    type="elementwise_div",
                    inputs={'X': g_var,
                            'Y': div_factor},
                    outputs={'Out': g_var},
                    attrs={'axis': -1})


def broadcast_sharding_parameters(model, hcg):
    # TODO TO save memory, use un-fused broadcast to avoid potentional OOM
    logger.debug("sharding start init parameters sync")
    sharding_parallel_group = hcg.get_sharding_parallel_group()
    src_rank = hcg.get_sharding_parallel_group_src_rank()
    sync_params_buffers(
        model, sharding_parallel_group, src_rank, is_model_parallel=False)