ChatGLM3 自己训练微调制作数据代码，与训练、训练完成后模型合并、解译代码完整版_chatglm3 int8微调

ChatGLM3 自己训练微调制作数据代码，与训练完成后模型合并解译代码

import json


keyword = '这年轻人'

# 摘自百度百科
description = "这年轻人，男，1993年出生于陕西省湖北市潼关县。2015年毕业于中国背景大学。2016年加入西安旧东方，当选（旧东方）当时最年轻的英语教研主管；2019年加入旧东方在线，是高三英语名师并成为高三英语学科最年轻的负责人，被称为“中关村王杰伦”。现是东方甄选高级合伙人、旧东方教育科技集团董事长文化助理，兼任新东方文旅集团副总裁。"

#对 prompt 使用一些简单的数据增强的方法，以便更好地收敛。
def get_prompt_list(keyword):
    return [
        f'{keyword}', 
        f'你知道{keyword}吗?',
        f'{keyword}是谁？',
        f'介绍一下{keyword}',
        f'你听过{keyword}吗?',
        f'谁是{keyword}？',
        f'{keyword}是？',
        f'你认识{keyword}吗？',
        f'{keyword}的资料',
        f'{keyword}简介'
    ]

# ChatGLM3 自己训练微调制作数据代码，与训练完成后模型合并解译代码


# 对话数据格式
data = [
    {
        "conversations": [
            {
                "role": "system",
                "content": "You are ChatGLM3, a large language model trained by Zhipu.AI. Follow the user's instructions carefully. Respond using markdown."
            },
            {
                "role": "user",
                "content": x
            },
            {
                "role": "assistant",
                "content": description
            }
        ]
    }
    for x in get_prompt_list(keyword)
]

# 保存到 formatted_data/my_data_qa.jsonl
with open("formatted_data/my_data_qa.jsonl", "w") as f:
    for e in data:
        f.write(json.dumps(e, ensure_ascii=False) + "\n")

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

模型合并代码

import torch
from peft import PeftModel
from transformers import AutoTokenizer, AutoModel
#加载原模型
base_model = '/media/DATA/XXX/large_model/weights'
base_model = AutoModel.from_pretrained(base_model, trust_remote_code=True).cuda(3)
#加载微调的模型
lora_model_path = '/media/DATA/XXX/large_model/Chat_weitiao/ChatGLM3/finetune_demo/output/checkpoint-3000'
lora_model = PeftModel.from_pretrained(base_model,lora_model_path, torch_dtype=torch.float16)
lora_model.to("cpu")
#合并
merged_model = lora_model.merge_and_unload()
#合并的模型存储
new_model_directory = '/media/DATA/XXX/large_model/Chat_weitiao/ChatGLM3/finetune_demo/output/fintrue_chatglm3'
merged_model.save_pretrained(new_model_directory, max_shard_size="2048MB", safe_serialization=True)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

后推理代码

from transformers import AutoModel, AutoTokenizer  # 导入transformers库的AutoModel和AutoTokenizer

#加载模型
new_model_directory = '/media/DATA/XXX/large_model/Chat_weitiao/ChatGLM3/finetune_demo/output/fintrue_chatglm3'
tokenizer = AutoTokenizer.from_pretrained(new_model_directory, trust_remote_code=True)
model = AutoModel.from_pretrained(new_model_directory, trust_remote_code=True).cuda(3)
model.eval()
#输入
#instruction = "你现在是一个信息抽取模型，请你帮我抽取出关系内容为\"性能故障\", \"部件故障\", \"组成\"和 \"检测工具\"的相关三元组，三元组内部用\"_\"连接，三元组之间用\\n分割。文本："
input = "被称为“中关村周杰伦"
#验证
response, _ = model.chat(tokenizer, input, history=None)
print(response)

1
2
3
4
5
6
7
8
9
10
11
12
13
14

第二种

#!/usr/bin/env python
# -*- coding: utf-8 -*-

from pathlib import Path
from typing import Annotated, Union

import typer
from peft import AutoPeftModelForCausalLM, PeftModelForCausalLM
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    PreTrainedModel,
    PreTrainedTokenizer,
    PreTrainedTokenizerFast,
)

ModelType = Union[PreTrainedModel, PeftModelForCausalLM]
TokenizerType = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]

app = typer.Typer(pretty_exceptions_show_locals=False)


def _resolve_path(path: Union[str, Path]) -> Path:
    return Path(path).expanduser().resolve()


def load_model_and_tokenizer(model_dir: Union[str, Path]) -> tuple[ModelType, TokenizerType]:
    model_dir = _resolve_path(model_dir)
    if (model_dir / 'adapter_config.json').exists():
        model = AutoPeftModelForCausalLM.from_pretrained(
            model_dir, trust_remote_code=True, device_map='auto'
        )
        tokenizer_dir = model.peft_config['default'].base_model_name_or_path
    else:
        model = AutoModelForCausalLM.from_pretrained(
            model_dir, trust_remote_code=True, device_map='auto'
        )
        tokenizer_dir = model_dir
    tokenizer = AutoTokenizer.from_pretrained(
        tokenizer_dir, trust_remote_code=True
    )
    return model, tokenizer


@app.command()
def main(
        model_dir: Annotated[str, typer.Argument(help='')],
        prompt: Annotated[str, typer.Option(help='')],
):
    model, tokenizer = load_model_and_tokenizer(model_dir)
    response, _ = model.chat(tokenizer, prompt)
    print(response)


if __name__ == '__main__':
    app()

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

训练解译指令

CUDA_VISIBLE_DEVICES=3  python finetune_hf.py formatted_data/ /media/DATA/XXX/large_model/weights/ configs/lora.yaml
1

CUDA_VISIBLE_DEVICES=2  python finetune_hf.py formatted_data/ /media/DATA/XXX/large_model/weights/ configs/ptuning_v2.yaml

1
2

训练代码

# -*- coding: utf-8 -*-

import dataclasses as dc
import functools
from collections.abc import Callable, Mapping, Sequence
from pathlib import Path
from typing import Annotated, Any, Optional, Union

import jieba
import numpy as np
import ruamel.yaml as yaml
import torch
import typer
from datasets import Dataset, DatasetDict, NamedSplit, Split, load_dataset
from nltk.translate.bleu_score import SmoothingFunction, sentence_bleu
from peft import (
    PeftConfig,
    PeftModelForCausalLM,
    get_peft_config,
    get_peft_model
)
from rouge_chinese import Rouge
from torch import nn
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    EvalPrediction,
    GenerationConfig,
    PreTrainedModel,
    PreTrainedTokenizer,
    PreTrainedTokenizerFast,
    Seq2SeqTrainingArguments, AutoConfig,
)
from transformers import DataCollatorForSeq2Seq as _DataCollatorForSeq2Seq

from transformers import Seq2SeqTrainer as _Seq2SeqTrainer
import os

ModelType = Union[PreTrainedModel, PeftModelForCausalLM]
TokenizerType = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
app = typer.Typer(pretty_exceptions_show_locals=False)


class DataCollatorForSeq2Seq(_DataCollatorForSeq2Seq):
    def __call__(self, features, return_tensors=None):
        output_ids = (
            [feature['output_ids'] for feature in features]
            if 'output_ids' in features[0].keys()
            else None
        )
        if output_ids is not None:
            max_output_length = max(len(out) for out in output_ids)
            if self.pad_to_multiple_of is not None:
                max_output_length = (
                        (
                                max_output_length + self.pad_to_multiple_of - 1) //
                        self.pad_to_multiple_of * self.pad_to_multiple_of
                )
            for feature in features:
                remainder = [self.tokenizer.pad_token_id] * (
                        max_output_length - len(feature['output_ids'])
                )
                if isinstance(feature['output_ids'], list):
                    feature['output_ids'] = feature['output_ids'] + remainder
                else:
                    feature['output_ids'] = np.concatenate(
                        [feature['output_ids'], remainder]
                    ).astype(np.int64)
        return super().__call__(features, return_tensors)


class Seq2SeqTrainer(_Seq2SeqTrainer):
    def prediction_step(
            self,
            model: nn.Module,
            inputs: dict[str, Any],
            prediction_loss_only: bool,
            ignore_keys=None,
            **gen_kwargs,
    ) -> tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]:

        if self.args.predict_with_generate:
            output_ids = inputs.pop('output_ids')
        input_ids = inputs['input_ids']
        loss, generated_tokens, labels = super().prediction_step(
            model, inputs, prediction_loss_only, ignore_keys, **gen_kwargs
        )
        generated_tokens = generated_tokens[:, input_ids.size()[1]:]
        if self.args.predict_with_generate:
            labels = output_ids
        return loss, generated_tokens, labels


def _resolve_path(path: Union[str, Path]) -> Path:
    return Path(path).expanduser().resolve()


def _sanity_check(
        input_ids: Sequence[int],
        output_ids: Sequence[int],
        tokenizer: PreTrainedTokenizer,
):
    print('--> Sanity check')
    for in_id, out_id in zip(input_ids, output_ids):
        if in_id == 0:
            continue
        if in_id in tokenizer.tokenizer.index_special_tokens:
            in_text = tokenizer.tokenizer.index_special_tokens[in_id]
        else:
            in_text = tokenizer.decode([in_id])
        print(f'{repr(in_text):>20}: {in_id} -> {out_id}')


@functools.cache
def _get_yaml_parser() -> yaml.YAML:
    parser = yaml.YAML(typ='safe', pure=True)
    parser.indent(mapping=2, offset=2, sequence=4)
    parser.default_flow_style = False
    return parser


@dc.dataclass
class DataConfig(object):
    train_file: str
    val_file: Optional[str] = None
    test_file: Optional[str] = None

    num_proc: Optional[int] = None

    @property
    def data_format(self) -> str:
        return Path(self.train_file).suffix

    @property
    def data_files(self) -> dict[NamedSplit, str]:
        return {
            split: data_file
            for split, data_file in zip(
                [Split.TRAIN, Split.VALIDATION, Split.TEST],
                [self.train_file, self.val_file, self.test_file],
            )
            if data_file is not None
        }


@dc.dataclass
class FinetuningConfig(object):
    data_config: DataConfig

    max_input_length: int
    max_output_length: int

    training_args: Seq2SeqTrainingArguments = dc.field(
        default=Seq2SeqTrainingArguments(output_dir='./output')
    )
    peft_config: Optional[PeftConfig] = None

    def __post_init__(self):
        if not self.training_args.do_eval or self.data_config.val_file is None:
            # skips the evaluation stage when `do_eval` or `eval_file` is not provided
            self.training_args.do_eval = True
            self.training_args.evaluation_strategy = 'no'
            self.data_config.val_file = None
        else:
            self.training_args.per_device_eval_batch_size = (
                    self.training_args.per_device_eval_batch_size
                    or self.training_args.per_device_train_batch_size
            )

    @classmethod
    def from_dict(cls, **kwargs) -> 'FinetuningConfig':
        training_args = kwargs.get('training_args', None)
        if training_args is not None and not isinstance(
                training_args, Seq2SeqTrainingArguments
        ):
            gen_config = training_args.get('generation_config')
            # TODO: a bit hacky
            if not isinstance(gen_config, GenerationConfig):
                training_args['generation_config'] = GenerationConfig(
                    **gen_config
                )
            kwargs['training_args'] = Seq2SeqTrainingArguments(**training_args)

        data_config = kwargs.get('data_config')
        if not isinstance(data_config, DataConfig):
            kwargs['data_config'] = DataConfig(**data_config)

        peft_config = kwargs.get('peft_config', None)
        if peft_config is not None and not isinstance(peft_config, PeftConfig):
            kwargs['peft_config'] = get_peft_config(peft_config)
        return cls(**kwargs)

    @classmethod
    def from_file(cls, path: Union[str, Path]) -> 'FinetuningConfig':
        path = _resolve_path(path)
        kwargs = _get_yaml_parser().load(path)
        return cls.from_dict(**kwargs)


def _load_datasets(
        data_dir: Path,
        data_format: str,
        data_files: dict[NamedSplit, str],
        num_proc: Optional[int],
) -> DatasetDict:
    if data_format in ('.csv', '.json', '.jsonl'):
        dataset_dct = load_dataset(
            data_format[1:],
            data_dir=data_dir,
            data_files=data_files,
            num_proc=num_proc,
        )
    else:
        err_msg = f"Cannot load dataset in the '{data_format}' format."
        raise NotImplementedError(err_msg)

    return dataset_dct


class DataManager(object):
    def __init__(self, data_dir: str, data_config: DataConfig):
        self._num_proc = data_config.num_proc

        self._dataset_dct = _load_datasets(
            _resolve_path(data_dir),
            data_config.data_format,
            data_config.data_files,
            self._num_proc,
        )

    def _get_dataset(self, split: NamedSplit) -> Optional[Dataset]:
        return self._dataset_dct.get(split, None)

    def get_dataset(
            self,
            split: NamedSplit,
            process_fn: Callable[[dict[str, Any]], dict[str, Any]],
            batched: bool = True,
            remove_orig_columns: bool = True,
    ) -> Optional[Dataset]:
        orig_dataset = self._get_dataset(split)
        if orig_dataset is None:
            return

        if remove_orig_columns:
            remove_columns = orig_dataset.column_names
        else:
            remove_columns = None
        return orig_dataset.map(
            process_fn,
            batched=batched,
            remove_columns=remove_columns,
            num_proc=self._num_proc,
        )
def print_model_size(model: PreTrainedModel):
    print("--> Model")
    total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    print(f"\n--> model has {total_params / 1e6}M params\n")


def process_batch(
        batch: Mapping[str, Sequence],
        tokenizer: PreTrainedTokenizer,
        max_input_length: int,
        max_output_length: int,
) -> dict[str, list]:
    batched_tools = batch.get('tools', None)
    batched_conv = batch['conversations']
    batched_input_ids = []
    batched_labels = []

    if batched_tools is None:
        batched_tools = [None] * len(batched_conv)

    for tools, conv in zip(batched_tools, batched_conv):
        input_ids, loss_masks = [
            tokenizer.get_command('[gMASK]'),
            tokenizer.get_command('sop'),
        ], [False, False]

        if tools is not None:
            raise NotImplementedError()

        for message in conv:
            if message['role'] in ('system', 'user'):
                loss_mask_val = False
            else:
                loss_mask_val = True

            if message['role'] == 'tool':
                raise NotImplementedError()
            else:
                new_input_ids = tokenizer.build_single_message(
                    message['role'], '', message['content']
                )
                new_loss_masks = [loss_mask_val] * len(new_input_ids)

            input_ids += new_input_ids
            loss_masks += new_loss_masks

        input_ids.append(tokenizer.eos_token_id)
        loss_masks = [False, *loss_masks]
        labels = []
        for input_id, mask in zip(input_ids, loss_masks):
            if mask:
                labels.append(input_id)
            else:
                labels.append(-100)
        max_length = max_input_length + max_output_length + 1
        batched_input_ids.append(input_ids[:max_length])
        batched_labels.append(labels[:max_length])
    return {'input_ids': batched_input_ids, 'labels': batched_labels}


def process_batch_eval(
        batch: Mapping[str, Sequence],
        tokenizer: PreTrainedTokenizer,
        max_input_length: int,
        max_output_length: int,
) -> dict[str, list]:
    batched_tools = batch.get('tools', None)
    batched_conv = batch['conversations']
    batched_input_ids = []
    # To avoid computing loss, we do not provide the `labels` field in the input dictionary.
    batched_output_ids = []

    if batched_tools is None:
        batched_tools = [None] * len(batched_conv)

    for tools, conv in zip(batched_tools, batched_conv):
        input_ids = [
            tokenizer.get_command('[gMASK]'),
            tokenizer.get_command('sop'),
        ]

        if tools is not None:
            raise NotImplementedError()

        for message in conv:
            if len(input_ids) >= max_input_length:
                break
            if message['role'] == 'tool':
                raise NotImplementedError()
            else:
                new_input_ids = tokenizer.build_single_message(
                    message['role'], '', message['content']
                )
                if message['role'] == 'assistant':
                    output_prompt, output_ids = (
                        new_input_ids[:1],
                        new_input_ids[1:],
                    )
                    output_ids.append(tokenizer.eos_token_id)
                    batched_input_ids.append(
                        input_ids[:max_input_length] + output_prompt[:1]
                    )
                    batched_output_ids.append(output_ids[:max_output_length])
                input_ids += new_input_ids
    return {'input_ids': batched_input_ids, 'output_ids': batched_output_ids}


# TODO: Not sure if this is necessary, can set it to half
def _prepare_model_for_training(model: nn.Module, use_cpu: bool):
    for param in model.parameters():
        if param.requires_grad or use_cpu:
	    # if train with cpu, cast all params to fp32 instead of trainable ones.
            param.data = param.data.to(torch.float32)


def load_tokenizer_and_model(
        model_dir: str,
        peft_config: Optional[PeftConfig] = None,
) -> tuple[PreTrainedTokenizer, nn.Module]:
    tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
    if peft_config is not None:
        if peft_config.peft_type.name == "PREFIX_TUNING":
            config = AutoConfig.from_pretrained(model_dir, trust_remote_code=True)
            config.pre_seq_len = peft_config.num_virtual_tokens
            config.use_cache = False
            model = AutoModelForCausalLM.from_pretrained(
                model_dir,
                trust_remote_code=True,
                config=config,
            )
        if peft_config.peft_type.name == "LORA":
            model = AutoModelForCausalLM.from_pretrained(
                model_dir,
                trust_remote_code=True,
                empty_init=False,
                use_cache=False
            )
            model = get_peft_model(model, peft_config)
            model.print_trainable_parameters()
    else:
        model = AutoModelForCausalLM.from_pretrained(
            model_dir,
            trust_remote_code=True,
            empty_init=False,
            use_cache=False
        )
    print_model_size(model)
    return tokenizer, model


def compute_metrics(eval_preds: EvalPrediction, tokenizer: PreTrainedTokenizer):
    batched_pred_ids, batched_label_ids = eval_preds

    metrics_dct = {'rouge-1': [], 'rouge-2': [], 'rouge-l': [], 'bleu-4': []}
    for pred_ids, label_ids in zip(batched_pred_ids, batched_label_ids):
        pred_txt = tokenizer.decode(pred_ids).strip()
        label_txt = tokenizer.decode(label_ids).strip()
        pred_tokens = list(jieba.cut(pred_txt))
        label_tokens = list(jieba.cut(label_txt))
        rouge = Rouge()
        scores = rouge.get_scores(' '.join(pred_tokens), ' '.join(label_tokens))
        for k, v in scores[0].items():
            metrics_dct[k].append(round(v['f'] * 100, 4))
        metrics_dct['bleu-4'].append(
            sentence_bleu(
                [label_tokens],
                pred_tokens,
                smoothing_function=SmoothingFunction().method3,
            )
        )
    return {k: np.mean(v) for k, v in metrics_dct.items()}


@app.command()
def main(
        data_dir: Annotated[str, typer.Argument(help='')],
        model_dir: Annotated[
            str,
            typer.Argument(
                help='A string that specifies the model id of a pretrained model configuration hosted on huggingface.co, or a path to a directory containing a model configuration file.'
            ),
        ],
        config_file: Annotated[str, typer.Argument(help='')],
        auto_resume_from_checkpoint: str = typer.Argument(
            default='yes',
            help='If entered as yes, automatically use the latest save checkpoint. If it is a numerical example 12 15, use the corresponding save checkpoint. If the input is no, restart training'
        ),

):
    ft_config = FinetuningConfig.from_file(config_file)
    tokenizer, model = load_tokenizer_and_model(model_dir, peft_config=ft_config.peft_config)
    data_manager = DataManager(data_dir, ft_config.data_config)

    train_dataset = data_manager.get_dataset(
        Split.TRAIN,
        functools.partial(
            process_batch,
            tokenizer=tokenizer,
            max_input_length=ft_config.max_input_length,
            max_output_length=ft_config.max_output_length,
        ),
        batched=True,
    )
    print('train_dataset:', train_dataset)
    val_dataset = data_manager.get_dataset(
        Split.VALIDATION,
        functools.partial(
            process_batch_eval,
            tokenizer=tokenizer,
            max_input_length=ft_config.max_input_length,
            max_output_length=ft_config.max_output_length,
        ),
        batched=True,
    )
    if val_dataset is not None:
        print('val_dataset:', val_dataset)
    test_dataset = data_manager.get_dataset(
        Split.TEST,
        functools.partial(
            process_batch_eval,
            tokenizer=tokenizer,
            max_input_length=ft_config.max_input_length,
            max_output_length=ft_config.max_output_length,
        ),
        batched=True,
    )
    if test_dataset is not None:
        print('test_dataset:', test_dataset)

    # checks encoded dataset
    # _sanity_check(
    #     train_dataset[0]["input_ids"], train_dataset[0]["labels"], tokenizer
    # )

    # turn model to fp32
    _prepare_model_for_training(model, ft_config.training_args.use_cpu)

    ft_config.training_args.generation_config.pad_token_id = (
        tokenizer.pad_token_id
    )
    ft_config.training_args.generation_config.eos_token_id = [
        tokenizer.eos_token_id,
        tokenizer.get_command('<|user|>'),
        tokenizer.get_command('<|observation|>'),
    ]
    model.gradient_checkpointing_enable()
    model.enable_input_require_grads()
    print("111:: ",ft_config.training_args)
    # exit()
    trainer = Seq2SeqTrainer(
        model=model,
        args=ft_config.training_args,
        data_collator=DataCollatorForSeq2Seq(
            tokenizer=tokenizer,
            padding='longest',
            return_tensors='pt',
        ),
        train_dataset=train_dataset,
        eval_dataset=val_dataset.select(list(range(10))), ## 50
        tokenizer=tokenizer,
        compute_metrics=functools.partial(compute_metrics, tokenizer=tokenizer),
    )

    # Determine whether to continue training without breakpoints or if it is empty, then start training again directly
    if auto_resume_from_checkpoint.upper() == "" or auto_resume_from_checkpoint is None:
        trainer.train()
    else:
        output_dir = ft_config.training_args.output_dir
        dirlist = os.listdir(output_dir)
        checkpoint_sn = 0
        for checkpoint_str in dirlist:
            if checkpoint_str.find("eckpoint") > 0 and checkpoint_str.find("tmp") == -1:
                checkpoint = int(checkpoint_str.replace("checkpoint-", ""))
                if checkpoint > checkpoint_sn:
                    checkpoint_sn = checkpoint
        if auto_resume_from_checkpoint.upper() == "YES":
            if checkpoint_sn > 0:
                model.gradient_checkpointing_enable()
                model.enable_input_require_grads()
                checkpoint_directory = os.path.join(output_dir, "checkpoint-" + str(checkpoint_sn))
                print("resume checkpoint from  checkpoint-" + str(checkpoint_sn))
                trainer.train(resume_from_checkpoint=checkpoint_directory)
            else:
                trainer.train()
        else:
            if auto_resume_from_checkpoint.isdigit():
                if int(auto_resume_from_checkpoint) > 0:
                    checkpoint_sn = int(auto_resume_from_checkpoint)
                    model.gradient_checkpointing_enable()
                    model.enable_input_require_grads()
                    checkpoint_directory = os.path.join(output_dir, "checkpoint-" + str(checkpoint_sn))
                    print("resume checkpoint from  checkpoint-" + str(checkpoint_sn))
                    trainer.train(resume_from_checkpoint=checkpoint_directory)
            else:
                print(auto_resume_from_checkpoint,
                      "The specified checkpoint sn(" + auto_resume_from_checkpoint + ") has not been saved. Please search for the correct chkeckpoint in the model output directory")

    # test stage
    if test_dataset is not None:
        trainer.predict(test_dataset)


if __name__ == '__main__':
    app()

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
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559

训练配置文件lora.yaml

data_config:
  train_file: my_data_qa.json
  val_file: my_data_qa.json
  test_file: my_data_qa.json
  num_proc: 1  #16
max_input_length: 128
max_output_length: 64 #256
training_args:
  # see `transformers.Seq2SeqTrainingArguments`
  output_dir: ./output
  max_steps: 30000
  # settings for data loading
  per_device_train_batch_size: 1
  dataloader_num_workers: 16
  remove_unused_columns: false
  # settings for saving checkpoints
  save_strategy: steps
  save_steps: 5000
  # settings for logging
  log_level: info
  logging_strategy: steps
  logging_steps: 10
  # settings for evaluation
  per_device_eval_batch_size: 1  #@16
  evaluation_strategy: steps
  eval_steps: 500
  # settings for optimizer
  # adam_epsilon: 1e-6
  # uncomment the following line to detect nan or inf values
  # debug: underflow_overflow
  predict_with_generate: true
  # see `transformers.GenerationConfig`
  generation_config:
    max_new_tokens: 256
  # set your absolute deepspeed path here
  #deepspeed: ds_zero_2.json
  # set to true if train with cpu.
  use_cpu: false
peft_config:
  peft_type: LORA
  task_type: CAUSAL_LM
  r: 8
  lora_alpha: 32
  lora_dropout: 0.1

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

训练配置文件ptuning_v2.yaml

data_config:
  train_file: my_data_qa.json    #train.json
  val_file: my_data_qa.json  #dev.json
  test_file: my_data_qa.json  #dev.json
#  train_file: /media/DATA/XXX/large_model/Chat_weitiao/ChatGLM3/finetune_demo/formatted_data/my_data_qa.jsonl    #train.json
#  val_file: /media/DATA/XXX/large_model/Chat_weitiao/ChatGLM3/finetune_demo/formatted_data/my_data_qa.jsonl  #dev.json
#  test_file: /media/DATA/XXX/large_model/Chat_weitiao/ChatGLM3/finetune_demo/formatted_data/my_data_qa.jsonl  #dev.json
  num_proc: 4  # 16
max_input_length: 256
max_output_length: 512
training_args:
  # see `transformers.Seq2SeqTrainingArguments`
  output_dir: ./output_p2
  max_steps: 3000
  # settings for data loading
  per_device_train_batch_size: 4
  dataloader_num_workers: 16 #16
  remove_unused_columns: false
  # settings for saving checkpoints
  save_strategy: steps
  save_steps: 500
  # settings for logging
  log_level: info
  logging_strategy: steps
  logging_steps: 10
  # settings for evaluation
  per_device_eval_batch_size: 16
  evaluation_strategy: steps
  eval_steps: 500
  # settings for optimizer
  # adam_epsilon: 1e-6
  # uncomment the following line to detect nan or inf values
  # debug: underflow_overflow
  predict_with_generate: true
  # see `transformers.GenerationConfig`
  generation_config:
    max_new_tokens: 512
  # set your absolute deepspeed path here
  #deepspeed: ds_zero_3.json
  use_cpu: false
peft_config:
  peft_type: PREFIX_TUNING
  task_type: CAUSAL_LM
  num_virtual_tokens: 128

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