“c7ece60e2d3715a1b37973fd503eea0848a795d5”上不存在“git@gitcode.net:RobotFutures/Paddle.git”
sharding_utils.py 8.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
#   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 contextlib
from collections import abc
from enum import Enum
from math import inf
20 21
import numpy as np
from types import MethodType
22 23

import paddle
24
from paddle import _C_ops
25
from paddle.fluid import core
26 27 28 29
from paddle.fluid import layers
from paddle.fluid.dygraph import to_variable
from paddle.fluid.framework import dygraph_only
from paddle.fluid.dygraph import base as imperative_base
30
from paddle.distributed.collective import _get_global_group
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50


class Taskflow:
    """
    Task flows, one way linked list for task acquisition.
    """

    def __init__(self, task, callback):
        self.task = task
        self.callback = callback


class Type(Enum):
    """
    Type of trainable parameters
    """
    fp16 = paddle.float16
    fp32 = paddle.float32


51
class ShardingClipGrad:
52

53
    def __init__(self, clip, device, group):
54 55
        self._clip = clip
        self._device = device
56
        self._group = group
57 58 59

    @imperative_base.no_grad
    def _dygraph_clip(self, params_grads):
B
Baibaifan 已提交
60 61
        sum_square_fp32, sum_square_fp16 = [], []
        unslice_params_fp32, unslice_params_fp16 = [], []
62 63

        for p, g in params_grads:
B
Baibaifan 已提交
64
            p_slice = True  # using for slice parameter in sharding stage3
65 66
            if g is None or getattr(p, 'need_clip', True) is False:
                continue
B
Baibaifan 已提交
67 68
            if hasattr(p, "unslice"):
                p_slice = False
69 70 71 72 73 74 75 76 77

            merge_grad = g
            if g.type == core.VarDesc.VarType.SELECTED_ROWS:
                merge_grad = layers.get_tensor_from_selected_rows(
                    layers.merge_selected_rows(g))
            square = layers.square(merge_grad)
            sum_square = layers.reduce_sum(square)

            if p.dtype == paddle.float16:
B
Baibaifan 已提交
78 79
                if p_slice: sum_square_fp16.append(sum_square)
                else: unslice_params_fp16.append(sum_square)
80
            elif p.dtype == paddle.float32:
B
Baibaifan 已提交
81 82
                if p_slice: sum_square_fp32.append(sum_square)
                else: unslice_params_fp32.append(sum_square)
83 84 85 86 87 88 89

        # global norm of non-distributed FP16 params_and_grads
        if len(sum_square_fp16) == 0:
            global_norm_fp16 = paddle.to_tensor([0.], dtype=paddle.float32)
        else:
            global_norm_fp16 = layers.concat(sum_square_fp16)
            global_norm_fp16 = layers.reduce_sum(global_norm_fp16)
90 91
            global_norm_fp16 = paddle.cast(global_norm_fp16,
                                           dtype=paddle.float32)
92

B
Baibaifan 已提交
93
        # global norm of non-distributed FP16 params_and_grads for unslice parameter
B
Baibaifan 已提交
94 95 96 97 98
        if len(unslice_params_fp16) == 0:
            global_unslice_fp16 = paddle.to_tensor([0.], dtype=paddle.float32)
        else:
            global_unslice_fp16 = layers.concat(unslice_params_fp16)
            global_unslice_fp16 = layers.reduce_sum(global_unslice_fp16)
99 100
            global_unslice_fp16 = paddle.cast(global_unslice_fp16,
                                              dtype=paddle.float32)
B
Baibaifan 已提交
101

102
        # global norm of non-distributed FP32 params_and_grads
103 104
        global_norm_fp32 = layers.concat(
            sum_square_fp32) if len(sum_square_fp32) != 0 else paddle.to_tensor(
105 106 107
                [0.], dtype=paddle.float32)
        global_norm_fp32 = layers.reduce_sum(global_norm_fp32)

B
Baibaifan 已提交
108
        # global norm of non-distributed FP32 params_and_grads for unslice parameter
B
Baibaifan 已提交
109 110 111 112 113 114
        global_unslice_fp32 = layers.concat(unslice_params_fp32) if len(
            unslice_params_fp32) != 0 else paddle.to_tensor(
                [0.], dtype=paddle.float32)
        global_unslice_fp32 = layers.reduce_sum(global_unslice_fp32)
        global_unslice_var = global_unslice_fp16 + global_unslice_fp32

B
Baibaifan 已提交
115
        global_norm_var = global_norm_fp16 + global_norm_fp32 + 1.0 / self._group.nranks * global_unslice_var
116 117 118 119 120 121 122

        # add all reduce to get global norm of distributed params_and_grads
        dev_id = int(self._device.split(":")[1])
        with device_guard(dev_id, "gpu"):
            paddle.distributed.all_reduce(global_norm_var, group=self._group)

        global_norm_var = layers.sqrt(global_norm_var)
123 124 125 126 127 128 129 130
        max_global_norm = layers.fill_constant(shape=[1],
                                               dtype=global_norm_var.dtype,
                                               value=self.clip_norm)

        clip_var = layers.elementwise_div(x=max_global_norm,
                                          y=layers.elementwise_max(
                                              x=global_norm_var,
                                              y=max_global_norm))
131 132 133
        clip_var_fp16 = paddle.cast(clip_var, paddle.float16)

        for p, g in params_grads:
134
            if getattr(p, 'need_clip', True) is False or g is None:
135
                continue
136 137
            origin_state = g.stop_gradient
            g.stop_gradient = True
138
            if p.dtype == paddle.float16:
139
                g.scale_(clip_var_fp16)
140
            else:
141
                g.scale_(clip_var)
142
            g.stop_gradient = origin_state
143
            p._reset_grad_inplace_version(True)
144

145
        return params_grads
146 147 148 149 150 151 152 153

    def __getattr__(self, item):
        return getattr(self._clip, item)

    def __call__(self, params_grads):
        return self._dygraph_clip(params_grads)


154
@contextlib.contextmanager
155
def device_guard(dev_id=0, device="cpu"):
156 157 158 159 160 161 162 163 164
    origin_device = paddle.device.get_device()
    if device == "cpu":
        paddle.set_device(device)
    elif device == "gpu":
        paddle.set_device("gpu:{}".format(dev_id))
    try:
        yield
    finally:
        paddle.set_device(origin_device)
165 166 167


@dygraph_only
168
def ShardingScaler(scaler):
169

170 171 172 173 174 175
    def unscale_method(self, optimizer):
        if not self._enable:
            return
        param_grads = []
        param_grads_fp16 = []
        param_grads_fp32 = []
B
Baibaifan 已提交
176 177 178
        if hasattr(optimizer, "update_slice"):
            optimizer.update_slice()
            optimizer.update_scaler = True
179

180 181
        if getattr(optimizer._optim, '_param_groups', None) and isinstance(
                optimizer._optim._param_groups[0], dict):
182

183
            for group in optimizer._optim._param_groups:
184 185 186
                for param in group['params']:
                    if param._grad_ivar() is not None:
                        param_grads.append(param._grad_ivar())
187 188 189
                        if param._grad_ivar().dtype in [
                                core.VarDesc.VarType.FP16, paddle.float16
                        ]:
190 191 192 193
                            param_grads_fp16.append(param._grad_ivar())
                        else:
                            param_grads_fp32.append(param._grad_ivar())
        else:
B
Baibaifan 已提交
194 195 196 197 198 199 200 201 202 203
            for param in optimizer._optim._parameter_list:
                if param.grad is not None:
                    param_grads.append(param.grad)
                    if param.grad.dtype in [
                            core.VarDesc.VarType.FP16, paddle.float16
                    ]:
                        param_grads_fp16.append(param.grad)
                    else:
                        param_grads_fp32.append(param.grad)

204 205
        temp_found_inf_fp16 = to_variable(np.array([0]).astype(np.bool))
        temp_found_inf_fp32 = to_variable(np.array([0]).astype(np.bool))
206 207

        device = "cpu" if optimizer.offload else "gpu"
208 209
        dev_id = 0 if device == "cpu" else int(
            paddle.get_device().split(":")[1])
210 211 212 213 214 215 216 217 218 219

        with device_guard(dev_id, device):
            if len(param_grads_fp16):
                _C_ops.check_finite_and_unscale(param_grads_fp16, self._scale,
                                                param_grads_fp16,
                                                temp_found_inf_fp16)
            if len(param_grads_fp32):
                _C_ops.check_finite_and_unscale(param_grads_fp32, self._scale,
                                                param_grads_fp32,
                                                temp_found_inf_fp32)
220 221 222 223

        self._found_inf = 1 if temp_found_inf_fp16 or temp_found_inf_fp32 else 0
        is_found_inf = paddle.to_tensor([self._found_inf], dtype="int32")

224 225 226
        paddle.distributed.all_reduce(is_found_inf,
                                      op=paddle.distributed.ReduceOp.MAX,
                                      group=optimizer.group)
227 228 229 230
        self._found_inf = is_found_inf.numpy()[0]

    scaler._unscale = MethodType(unscale_method, scaler)
    return scaler