Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
pretrain_megatron_chatglm.py 3.37 KB
Newer Older
hepj's avatar
Initial commit  
hepj committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98 
# Copyright (c) 2023 Alibaba PAI and Nvidia Megatron-LM Team.
#
# 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
from functools import partial
import torch

from megatron.core.enums import ModelType
from megatron import get_args
from megatron.utils import average_losses_across_data_parallel_group

from megatron_patch.data import \
    build_pretrain_dataset_from_original, build_pretrain_dataset_from_idxmap
from megatron_patch.model.chatglm.gpt_model import GPTModel
from megatron_patch.tokenizer import build_tokenizer
from megatron_patch.training import pretrain
from megatron_patch.arguments import get_patch_args
from megatron.arguments import core_transformer_config_from_args

from megatron_patch.tokenizer import get_tokenizer
from megatron_patch.data.finetune_dataset import ChatGLMDataset
from megatron.core import tensor_parallel
def model_provider(pre_process=True, post_process=True):
    args = get_args()
    build_tokenizer(args)
    config = core_transformer_config_from_args(get_args())
    model = GPTModel(config,
                     num_tokentypes=0,
                     parallel_output=True,
                     pre_process=pre_process,
                     post_process=post_process)
    return model


def train_valid_test_datasets_provider(train_val_test_num_samples):
    """Build train, valid, and test datasets."""
    args = get_args()

    train_ds, valid_ds, test_ds = \
        build_pretrain_dataset_from_original(args.dataset)
    return train_ds, valid_ds, test_ds



def forward_step(data_iterator, model):
    """Forward step."""

    # try:
    #     data = next(data_iterator)
    # except BaseException:
    #     data = data_iterator
    # datatype = torch.int64
    # keys = ['input_ids', 'labels']
    # data_b = tensor_parallel.broadcast_data(keys, data, datatype)
    # input_ids = data_b['input_ids'].long().cuda()
    # labels = data_b['labels'].long().cuda()
    # lm_logits = model(input_ids=input_ids)



    try:
        data_iterator = next(data_iterator)
    except BaseException:
        data_iterator = data_iterator
    input_ids = data_iterator['input_ids'].long().cuda()
    labels = data_iterator['labels'].long().cuda()
    lm_logits = model(input_ids=input_ids)

    shift_logits = lm_logits[..., :-1, :].contiguous()
    shift_labels = labels[..., 1:].contiguous()

    loss_fct = torch.nn.CrossEntropyLoss(ignore_index=-100)
    def loss_func(shift_logits):
        loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                        shift_labels.view(-1))
        averaged_loss = average_losses_across_data_parallel_group([loss])
        return loss, {'lm loss': averaged_loss[0]}
    
    return shift_logits, partial(loss_func)

if __name__ == '__main__':
    pretrain(train_valid_test_datasets_provider,
             model_provider,
             ModelType.encoder_or_decoder,
             forward_step,
             extra_args_provider=get_patch_args)