[大模型]Qwen-7B-Chat Ptuning 微调

Qwen-7B-Chat Ptuning 微调

Ptuning原理

参考transformers-code

数据加载与模型配置与LoRa一致,在此具体讲一下Ptuning的细节:
基本原理为冻结主模型全部参数,在训练数据前加入一小段Prompt,之训练Prompt的嵌入层。在Ptuning中，只有soft prompt,是自动学习的,不用人工设置。

• 在Prompt-tuning的时候，前面只是增加了一个Prompt Embedding.
• 在Ptuning这里替换为了Prompt Encoder，新加了LSTM/MLP+Embedding.

Ptuning PEFT模块

• PromptEncoderReparameterizationType来决定到底是用LSTM还是MLP
• 可调的参数有：num_virtual_tokens、encoder_reparameterization_type、encoder_dropout、encoder_num_layers、encoder_hidden_size
• 注意,encoder_num_layers设置层数,但MLP设置层数固定为2,此参数只对LSTM时有效。

from peft import PromptEncoderConfig, TaskType, get_peft_model, PromptEncoderReparameterizationType

config = PromptEncoderConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10,
                             encoder_reparameterization_type=PromptEncoderReparameterizationType.MLP,
                             encoder_dropout=0.1, encoder_num_layers=5, encoder_hidden_size=1024)
config
'''
PromptEncoderConfig(peft_type=<PeftType.P_TUNING: 'P_TUNING'>, auto_mapping=None, base_model_name_or_path=None, revision=None, task_type=<TaskType.CAUSAL_LM: 'CAUSAL_LM'>, inference_mode=False, num_virtual_tokens=10, token_dim=None, num_transformer_submodules=None, num_attention_heads=None, num_layers=None, encoder_reparameterization_type=<PromptEncoderReparameterizationType.MLP: 'MLP'>, encoder_hidden_size=1024, encoder_num_layers=5, encoder_dropout=0.1)
'''

其余可见目录下微调脚本

from datasets import Dataset
import pandas as pd
from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer

import torch
from peft import PromptEncoderConfig, TaskType, get_peft_model, PromptEncoderReparameterizationType
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '1'

# 用于处理数据集的函数
def process_func(example):
    MAX_LENGTH = 384    # Llama分词器会将一个中文字切分为多个token，因此需要放开一些最大长度，保证数据的完整性
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer("\n".join(["<|im_start|>system", "现在你要扮演皇帝身边的女人--甄嬛.<|im_end|>" + "\n<|im_start|>user\n" + example["instruction"] + example["input"] + "<|im_end|>\n"]).strip(), add_special_tokens=False)  # add_special_tokens 不在开头加 special_tokens
    response = tokenizer("<|im_start|>assistant\n" + example["output"] + "<|im_end|>\n", add_special_tokens=False)
    input_ids = instruction["input_ids"] + response["input_ids"] + [tokenizer.pad_token_id]
    attention_mask = instruction["attention_mask"] + response["attention_mask"] + [1]  # 因为eos token咱们也是要关注的所以 补充为1
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"] + [tokenizer.pad_token_id]  # Qwen的特殊构造就是这样的
    if len(input_ids) > MAX_LENGTH:  # 做一个截断
        input_ids = input_ids[:MAX_LENGTH]
        attention_mask = attention_mask[:MAX_LENGTH]
        labels = labels[:MAX_LENGTH]
    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels
    }

#  loraConfig
config = PromptEncoderConfig(
    task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10,
    encoder_reparameterization_type=PromptEncoderReparameterizationType.MLP,
    encoder_dropout=0.1, encoder_num_layers=5, encoder_hidden_size=1024)

# 配置训练参数
args = TrainingArguments(
    output_dir="./output/Qwen",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=2,
    logging_steps=10,
    num_train_epochs=3,
    # gradient_checkpointing=True,
    save_steps=100,
    learning_rate=1e-4,
    save_on_each_node=True
)


if "__main__" == __name__:
    os.chdir('/root/self-llm')  # /root/self-llm 需改成自己 self-llm 项目的绝对路径
    # 处理数据集
    # 将JSON文件转换为CSV文件
    df = pd.read_json('./dataset/huanhuan.json')
    ds = Dataset.from_pandas(df)
    # 加载tokenizer
    tokenizer = AutoTokenizer.from_pretrained('/root/autodl-tmp/qwen/Qwen-7B-Chat', use_fast=False, trust_remote_code=True)
    tokenizer.pad_token_id = tokenizer.eod_id
    # 将数据集变化为token形式
    tokenized_id = ds.map(process_func, remove_columns=ds.column_names)

    # 创建模型并以半精度形式加载
    model = AutoModelForCausalLM.from_pretrained('/root/autodl-tmp/qwen/Qwen-7B-Chat', trust_remote_code=True, torch_dtype=torch.half, device_map="auto")
    # model.enable_input_require_grads()  # 开启梯度检查点时，要执行该方法
    # 加载lora参数
    model = get_peft_model(model, config)
    # 使用trainer训练
    trainer = Trainer(
        model=model,
        args=args,
        train_dataset=tokenized_id,
        data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
        )
    trainer.train() # 开始训练
    response, history = model.chat(tokenizer, "你是谁", history=[], system="现在你要扮演皇帝身边的女人--甄嬛.")
    print(response)