pipeline_parallel.py 10.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
#   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

import paddle
import paddle.fluid as fluid
from .meta_parallel_base import MetaParallelBase
17
from .pp_utils.utils import is_float_tensor
18
from .parallel_layers.pp_layers import PipelineLayer
19 20 21

from ..utils.hybrid_parallel_util import broadcast_mp_parameters
from ..utils.hybrid_parallel_util import broadcast_dp_parameters
22
from ..utils.log_util import logger
23
from ..meta_optimizers.dygraph_optimizer import HybridParallelOptimizer, HybridParallelGradScaler
S
ShenLiang 已提交
24
from .pp_utils import p2p_communication as p2p
25

26 27
__all__ = []

28 29 30

class PipelineParallel(MetaParallelBase):
    def __init__(self, layers, hcg, strategy):
31 32 33
        if not isinstance(layers, PipelineLayer):
            raise TypeError(
                "The Layer should be a derived class of PipelineLayer.")
34 35 36 37 38 39 40 41 42 43 44 45 46
        super(PipelineParallel, self).__init__(layers, hcg, strategy)
        self.use_data_parallel = self._hcg.get_data_parallel_world_size() > 1
        self.use_model_parallel = self._hcg.get_model_parallel_world_size() > 1

        self.total_loss = None

        self.micro_batch_size = self._strategy.pipeline_configs[
            'micro_batch_size']
        self.accumulate_steps = self._strategy.pipeline_configs[
            'accumulate_steps']

        self.num_stages = self._hcg.get_pipe_parallel_world_size()
        self.stage_id = self._hcg.get_stage_id()
47
        self.pp_group = self._hcg.get_pipe_parallel_group()
48

S
ShenLiang 已提交
49
        p2p.initialize_p2p_groups(hcg)
50 51 52 53

        self.is_first_stage = self.stage_id == 0
        self.is_last_stage = (self.stage_id == (self.num_stages - 1))
        self.global_rank = self._hcg.get_global_rank()
54
        self.micro_batch_id = 0
55

56 57 58 59 60 61 62 63
        logger.info("Pipeline Info -- num_stages: {}, stage_id: {}".format(
            self.num_stages, self.stage_id))

        if self.use_model_parallel:
            logger.info("start broadcast mp parameters")
            broadcast_mp_parameters(self._layers, self._hcg)

        if self.use_data_parallel:
64
            logger.info("start broadcast dp parameters")
65
            broadcast_dp_parameters(self._layers, self._hcg)
66

67
    def train_batch(self, data, optimizer, lr_scheduler=None, scaler=None):
68 69
        assert isinstance(optimizer, HybridParallelOptimizer), (
            'optimizer should be HybridParallelOptimizer subclass.')
70 71 72
        if scaler is not None:
            assert isinstance(scaler, HybridParallelGradScaler), (
                'scaler should be HybridParallelGradScaler subclass or None.')
73 74 75
        assert fluid.framework._dygraph_tracer()._has_grad, (
            'Please enable the generation of gradients.')

76 77
        if self.is_first_stage or self.is_last_stage:
            assert data is not None, (
78
                "For the first and the last stage, the data must be set.")
79
        else:
80 81
            data = None

82 83 84
        self.optimizer = optimizer
        self.lr_scheduler = lr_scheduler
        self.scaler = scaler
85
        self.data = data
86

87 88
        self._layers.train()

89 90 91
        # store total loss of entire batch
        self.total_loss = None

92 93
        # store data id for micro_batch
        self.micro_batch_id = 0
94

95 96 97
        # Next, use the 1f1b scheduling strategy.
        # this strategy is inspired by:
        # https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/schedules.py
98

99 100 101
        startup_steps = (self.num_stages - self.stage_id - 1)
        startup_steps = min(startup_steps, self.accumulate_steps)
        steady_steps = self.accumulate_steps - startup_steps
102

103 104
        input_buffers = []
        output_buffers = []
105

106 107
        for step_id in range(startup_steps):
            input_tensor = p2p.recv_forward()
108

109 110
            output_tensor = self._forward_step(input_tensor)
            p2p.send_forward(output_tensor)
111

112 113
            input_buffers.append(input_tensor)
            output_buffers.append(output_tensor)
114

115 116
        if steady_steps > 0:
            input_tensor = p2p.recv_forward()
117

118 119
        for i in range(steady_steps):
            last_iter = (i == (steady_steps - 1))
120

121
            output_tensor = self._forward_step(input_tensor)
122

123
            output_tensor_grad = p2p.send_forward_recv_backward(output_tensor)
124

125 126
            input_buffers.append(input_tensor)
            output_buffers.append(output_tensor)
127

128 129
            input_tensor, output_tensor = input_buffers.pop(
                0), output_buffers.pop(0)
130

131 132 133 134 135 136
            input_tensor_grad = self._backward_step(input_tensor, output_tensor,
                                                    output_tensor_grad)

            if last_iter:
                input_tensor = None
                p2p.send_backward(input_tensor_grad)
137
            else:
138
                input_tensor = p2p.send_backward_recv_forward(input_tensor_grad)
139

140 141 142
        for i in range(startup_steps):
            input_tensor = input_buffers.pop(0)
            output_tensor = output_buffers.pop(0)
143

144
            output_tensor_grad = p2p.recv_backward()
145

146 147 148
            input_tensor_grad = self._backward_step(input_tensor, output_tensor,
                                                    output_tensor_grad)
            p2p.send_backward(input_tensor_grad)
149

150
        self._layers.allreduce_shared_weight_gradients()
151

152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
        self.train_loss = self._reduce_final_loss()

        # optimizer
        self._optimizer_step()
        return self.train_loss

    def _forward_step(self, input_tensor):
        if self.stage_id == 0:
            input_tensor = self._load_micro_batch(self.micro_batch_id)

        output_tensor = self._layers.forward(input_tensor)

        if self.is_last_stage:
            labels = self._load_micro_batch(self.micro_batch_id)
            output_tensor = self._layers._loss_fn(output_tensor, labels)
            assert isinstance(
                output_tensor, paddle.
                Tensor), "Currently, loss_fn should obtain Paddle.Tensor dtype"

            if self.accumulate_steps > 1:
                output_tensor = output_tensor / self.accumulate_steps

            if self.total_loss is None:
                self.total_loss = paddle.zeros_like(output_tensor)
            self.total_loss += output_tensor.detach()

        self.micro_batch_id += 1
        return output_tensor

    def _backward_step(self, input_tensor, output_tensor, output_tensor_grad):
        if self.is_last_stage:
            assert output_tensor_grad is None
            if self.scaler:
                paddle.autograd.backward(self.scaler.scale(output_tensor))
            else:
                paddle.autograd.backward(output_tensor)
188
        else:
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
            if isinstance(output_tensor, tuple):
                outputs = [t for t in output_tensor if not t.stop_gradient]
                assert len(outputs) == len(output_tensor_grad)
                paddle.autograd.backward(
                    tensors=outputs,
                    grad_tensors=[t for t in output_tensor_grad])
            else:
                paddle.autograd.backward(
                    tensors=[output_tensor], grad_tensors=[output_tensor_grad])

        input_tensor_grad = None
        if input_tensor is not None:
            if isinstance(input_tensor, tuple):
                input_tensor_grad = tuple(
                    [t.grad for t in input_tensor if not t.stop_gradient])
            else:
                input_tensor_grad = input_tensor.grad
        return input_tensor_grad
207 208

    def _load_micro_batch(self, cache_id):
209 210 211 212 213 214 215 216 217 218 219 220 221
        inputs = self.data
        begin = cache_id * self.micro_batch_size
        end = begin + self.micro_batch_size

        if self.is_first_stage:
            assert len(inputs) == 2, "length of input should be 2"
            if isinstance(inputs[0], tuple):
                batch_size = inputs[0][0].shape[0]
                assert self.micro_batch_size * self.accumulate_steps == batch_size, (
                    "batch_size needs to be divisible by micro_batch_size. Currently, "
                    "batch_size = %d, micro_batch_size = %d, accumulate_steps = %d."
                    %
                    (batch_size, self.micro_batch_size, self.accumulate_steps))
222 223
                data = [input[begin:end, :].detach() for input in inputs[0]]
                return tuple(data)
224 225 226
            else:
                batch_size = inputs[0].shape[0]
                assert self.micro_batch_size * self.accumulate_steps == batch_size
227
                return inputs[0][begin:end, :].detach()
228 229 230 231 232
        elif self.is_last_stage:
            assert len(inputs) == 2, "length of input should be 2"
            if isinstance(inputs[1], tuple):
                batch_size = inputs[1][0].shape[0]
                assert self.micro_batch_size * self.accumulate_steps == batch_size
233 234
                data = [input[begin:end, :].detach() for input in inputs[1]]
                return tuple(data)
235
            else:
236 237
                batch_size = inputs[1].shape[0]
                assert self.micro_batch_size * self.accumulate_steps == batch_size
238
                return inputs[1][begin:end, :].detach()
239 240 241
        else:
            # No data input is required for other stages
            inputs = None
242

243 244 245 246 247 248 249 250 251
    def _reduce_final_loss(self):
        if self.is_last_stage:
            assert self.total_loss is not None, "train_batch() in last stage should obtain vaild loss"
            loss = self.total_loss.detach()
            paddle.distributed.broadcast(
                loss,
                src=self.global_rank,
                use_calc_stream=True,
                group=self.pp_group)
252
        else:
253 254 255 256 257 258 259
            loss = paddle.zeros(shape=[1], dtype="float32")
            paddle.distributed.broadcast(
                loss,
                src=self._hcg.get_rank_from_stage(self.num_stages - 1),
                use_calc_stream=True,
                group=self.pp_group)
        return loss
260

261
    def _optimizer_step(self):
262 263 264 265
        if self.scaler:
            self.scaler.minimize(self.optimizer, self.train_loss)
        else:
            self.optimizer.step()
266

267 268 269
        self.optimizer.clear_grad()
        if self.lr_scheduler:
            self.lr_scheduler.step()