基于 Qwen2-1.5B Lora 微调训练医疗问答任务

一、Qwen2 Lora 微调

Qwen是阿里巴巴集团Qwen团队研发的大语言模型和大型多模态模型系列。Qwen2 是 Qwen1.5 的重大升级。无论是语言模型还是多模态模型，均在大规模多语言和多模态数据上进行预训练，并通过高质量数据进行后期微调以贴近人类偏好。Qwen具备自然语言理解、文本生成、视觉理解、音频理解、工具使用、角色扮演、作为AI Agent进行互动等多种能力。

Qwen2有以下特点：

• • 5种模型规模，包括0.5B、1.5B、7B、57B-A14B和72B；
• • 针对每种尺寸提供基础模型和指令微调模型，并确保指令微调模型按照人类偏好进行校准；
• • 基础模型和指令微调模型的多语言支持；
• • 所有模型均稳定支持32K长度上下文；Qwen2-7B-Instruct与Qwen2-72B-Instruct可支持128K上下文（需额外配置）
• • 支持工具调用、RAG（检索增强文本生成）、角色扮演、AI Agent等；

更多详细的介绍可以参考官方文档：

https://qwen.readthedocs.io/zh-cn/latest/

下面实验所使用的核心依赖版本如下：

torch==1.13.1+cu116
peft==0.12.0
transformers==4.37.0
tensorboard==2.17.1

二、构建 Qwen2-1.5B Lora 模型

LoRA 微调技术的思想很简单，在原始 PLM (Pre-trained Language Model) 增加一个旁路，一般是在 transformer 层，做一个降维再升维的操作，模型的输入输出维度不变，来模拟 intrinsic rank，如下图的 A 和 B。训练时冻结 PLM 的参数，只训练 A 和 B ，，输出时将旁路输出与 PLM 的参数叠加，进而影响原始模型的效果。该方式，可以大大降低训练的参数量，而性能可以优于其它参数高效微调方法，甚至和全参数微调（Fine-Tuning）持平甚至超过。

对于 A 和 B 参数的初始化，A 使用随机高斯分布，B 使用 0 矩阵，这样在最初时可以保证旁路为一个 0 矩阵，最开始时使用原始模型的能力。

在构建 Qwen2-1.5B Lora 结构模型前，先了解下现在 Qwen2-1.5B 的结构：

这里直接使用 PyTorch 的模型打印方式，主要看模型的组成：

from transformers import AutoModelForCausalLM

model_path = "model/Qwen2-1.5B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype="auto", device_map="auto")
print(model)

输出结果：

Qwen2ForCausalLM(
  (model): Qwen2Model(
    (embed_tokens): Embedding(151936, 1536)
    (layers): ModuleList(
      (0): Qwen2DecoderLayer(
        (self_attn): Qwen2Attention(
          (q_proj): Linear(in_features=1536, out_features=1536, bias=True)
          (k_proj): Linear(in_features=1536, out_features=256, bias=True)
          (v_proj): Linear(in_features=1536, out_features=256, bias=True)
          (o_proj): Linear(in_features=1536, out_features=1536, bias=False)
          (rotary_emb): Qwen2RotaryEmbedding()
        )
        (mlp): Qwen2MLP(
          (gate_proj): Linear(in_features=1536, out_features=8960, bias=False)
          (up_proj): Linear(in_features=1536, out_features=8960, bias=False)
          (down_proj): Linear(in_features=8960, out_features=1536, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): Qwen2RMSNorm()
        (post_attention_layernorm): Qwen2RMSNorm()
      )
      .
      . 省略中间结构
      .
      (27): Qwen2DecoderLayer(
        (self_attn): Qwen2Attention(
          (q_proj): Linear(in_features=1536, out_features=1536, bias=True)
          (k_proj): Linear(in_features=1536, out_features=256, bias=True)
          (v_proj): Linear(in_features=1536, out_features=256, bias=True)
          (o_proj): Linear(in_features=1536, out_features=1536, bias=False)
          (rotary_emb): Qwen2RotaryEmbedding()
        )
        (mlp): Qwen2MLP(
          (gate_proj): Linear(in_features=1536, out_features=8960, bias=False)
          (up_proj): Linear(in_features=1536, out_features=8960, bias=False)
          (down_proj): Linear(in_features=8960, out_features=1536, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): Qwen2RMSNorm()
        (post_attention_layernorm): Qwen2RMSNorm()
      )
    )
    (norm): Qwen2RMSNorm()
  )
  (lm_head): Linear(in_features=1536, out_features=151936, bias=False)
)

从上面的结构可以看出 Qwen2-1.5B 的结构其实并不复杂，由 27 层 DecoderLayer 构成，每个 Decoder 主要的核心是 self_attention 和 mlp，因此可以尝试在 q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj 层添加 Lora 结构，下面使用 PEFT 库实现，这里 r 使用 8，lora_alpha 使用 32 ：

from transformers import AutoModelForCausalLM
from peft import LoraConfig, get_peft_model, TaskType

model_path = "model/Qwen2-1.5B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype="auto", device_map="auto")
peft_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    inference_mode=False,
    r=8,
    lora_alpha=32,
    lora_dropout=0.1
)
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()
print(model)

输出结果：

trainable params: 9,232,384 || all params: 1,552,946,688 || trainable%: 0.5945
PeftModelForCausalLM(
  (base_model): LoraModel(
    (model): Qwen2ForCausalLM(
      (model): Qwen2Model(
        (embed_tokens): Embedding(151936, 1536)
        (layers): ModuleList(
          (0): Qwen2DecoderLayer(
            (self_attn): Qwen2Attention(
              (q_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=1536, bias=True)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=1536, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (k_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=256, bias=True)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=256, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (v_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=256, bias=True)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=256, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (o_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=1536, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=1536, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (rotary_emb): Qwen2RotaryEmbedding()
            )
            (mlp): Qwen2MLP(
              (gate_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=8960, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=8960, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (up_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=8960, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=8960, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (down_proj): lora.Linear(
                (base_layer): Linear(in_features=8960, out_features=1536, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=8960, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=1536, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (act_fn): SiLU()
            )
            (input_layernorm): Qwen2RMSNorm()
            (post_attention_layernorm): Qwen2RMSNorm()
          )
       .
       . 省略中间结构
       .  
          (27): Qwen2DecoderLayer(
            (self_attn): Qwen2Attention(
              (q_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=1536, bias=True)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=1536, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (k_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=256, bias=True)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=256, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (v_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=256, bias=True)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=256, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (o_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=1536, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=1536, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (rotary_emb): Qwen2RotaryEmbedding()
            )
            (mlp): Qwen2MLP(
              (gate_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=8960, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=8960, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (up_proj): lora.Linear(
                (base_layer): Linear(in_features=1536, out_features=8960, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=1536, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=8960, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (down_proj): lora.Linear(
                (base_layer): Linear(in_features=8960, out_features=1536, bias=False)
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.1, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=8960, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=1536, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
                (lora_magnitude_vector): ModuleDict()
              )
              (act_fn): SiLU()
            )
            (input_layernorm): Qwen2RMSNorm()
            (post_attention_layernorm): Qwen2RMSNorm()
          )
        )
        (norm): Qwen2RMSNorm()
      )
      (lm_head): Linear(in_features=1536, out_features=151936, bias=False)
    )
  )
)

从结果可以看出，Lora 之后在每一层都增加了一个 lora_A 和 lora_B 结构来实现降维升维的作用。

三、准备训练数据集

数据集采用 GitHub 上的 Chinese-medical-dialogue-data 中文医疗对话数据集。

GitHub 地址如下：

https://github.com/Toyhom/Chinese-medical-dialogue-data

数据分了 6 个科目类型：

数据格式如下所示：

其中 ask 为病症的问题描述，answer 为病症的回答。

由于整体数据比较多，这里为了演示效果，选取 内科、肿瘤科、儿科、外科 四个科目的数据进行实验，并且每个科目取前 10000 条数据进行训练、2000 条数据进行验证。

首先将数据集转为 json 格式方便后续读取：

import json
import pandas as pd

data_path = [
    "./data/Chinese-medical-dialogue-data-master/Data_数据/IM_内科/内科5000-33000.csv",
    "./data/Chinese-medical-dialogue-data-master/Data_数据/Oncology_肿瘤科/肿瘤科5-10000.csv",
    "./data/Chinese-medical-dialogue-data-master/Data_数据/Pediatric_儿科/儿科5-14000.csv",
    "./data/Chinese-medical-dialogue-data-master/Data_数据/Surgical_外科/外科5-14000.csv",
]

train_json_path = "./data/train.json"
val_json_path = "./data/val.json"
# 每个数据取 10000 条作为训练
train_size = 10000
# 每个数据取 2000 条作为验证
val_size = 2000


def main():
    train_f = open(train_json_path, "a", encoding='utf-8')
    val_f = open(val_json_path, "a", encoding='utf-8')
    for path in data_path:
        data = pd.read_csv(path, encoding='ANSI')
        train_count = 0
        val_count = 0
        for index, row in data.iterrows():
            question = row["ask"]
            answer = row["answer"]
            line = {
                "question": question,
                "answer": answer
            }
            line = json.dumps(line, ensure_ascii=False)
            if train_count < train_size:
                train_f.write(line + "\n")
                train_count = train_count + 1
            elif val_count < val_size:
                val_f.write(line + "\n")
                val_count = val_count + 1
            else:
                break
    print("数据处理完毕！")
    train_f.close()
    val_f.close()


if __name__ == '__main__':
    main()

处理之后可以看到两个生成的文件：

四、微调训练

解析数据，构建 Dataset 数据集

qa_dataset.py：

# -*- coding: utf-8 -*-
from torch.utils.data import Dataset
import torch
import json
import numpy as np


class QADataset(Dataset):
    def __init__(self, data_path, tokenizer, max_source_length, max_target_length) -> None:
        super().__init__()
        self.tokenizer = tokenizer
        self.max_source_length = max_source_length
        self.max_target_length = max_target_length
        self.max_seq_length = self.max_source_length + self.max_target_length

        self.data = []
        if data_path:
            with open(data_path, "r", encoding='utf-8') as f:
                for line in f:
                    if not line or line == "":
                        continue
                    json_line = json.loads(line)
                    question = json_line["question"]
                    answer = json_line["answer"]
                    self.data.append({
                        "question": question,
                        "answer": answer
                    })
        print("data load ， size：", len(self.data))

    def preprocess(self, question, answer):
        messages = [
            {"role": "system", "content": "你是一个医疗方面的专家，可以根据患者的问题进行解答。"},
            {"role": "user", "content": question}
        ]
        prompt = self.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        instruction = self.tokenizer(prompt, add_special_tokens=False, max_length=self.max_source_length)
        response = self.tokenizer(answer, add_special_tokens=False, max_length=self.max_target_length)
        input_ids = instruction["input_ids"] + response["input_ids"] + [self.tokenizer.pad_token_id]
        attention_mask = (instruction["attention_mask"] + response["attention_mask"] + [1])
        labels = [-100] * len(instruction["input_ids"]) + response["input_ids"] + [self.tokenizer.pad_token_id]
        if len(input_ids) > self.max_seq_length:
            input_ids = input_ids[:self.max_seq_length]
            attention_mask = attention_mask[:self.max_seq_length]
            labels = labels[:self.max_seq_length]
        return input_ids, attention_mask, labels

    def __getitem__(self, index):
        item_data = self.data[index]

        input_ids, attention_mask, labels = self.preprocess(**item_data)

        return {
            "input_ids": torch.LongTensor(np.array(input_ids)),
            "attention_mask": torch.LongTensor(np.array(attention_mask)),
            "labels": torch.LongTensor(np.array(labels))
        }

    def __len__(self):
        return len(self.data)

训练：

# -*- coding: utf-8 -*-
import torch
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import LoraConfig, get_peft_model, TaskType
import pandas as pd
from qa_dataset import QADataset
from tqdm import tqdm
import os, time, sys


def train_model(model, train_loader, val_loader, optimizer, gradient_accumulation_steps,
                device, num_epochs, model_output_dir, writer):
    batch_step = 0
    for epoch in range(num_epochs):
        time1 = time.time()
        model.train()
        for index, data in enumerate(tqdm(train_loader, file=sys.stdout, desc="Train Epoch: " + str(epoch))):
            input_ids = data['input_ids'].to(device, dtype=torch.long)
            attention_mask = data['attention_mask'].to(device, dtype=torch.long)
            labels = data['labels'].to(device, dtype=torch.long)
            # 前向传播
            outputs = model(
                input_ids=input_ids,
                attention_mask=attention_mask,
                labels=labels,
            )
            loss = outputs.loss
            # 反向传播，计算当前梯度
            loss.backward()
            # 梯度累积步数
            if (index % gradient_accumulation_steps == 0 and index != 0) or index == len(train_loader) - 1:
                # 更新网络参数
                optimizer.step()
                # 清空过往梯度
                optimizer.zero_grad()
                writer.add_scalar('Loss/train', loss, batch_step)
                batch_step += 1
            # 100轮打印一次 loss
            if index % 100 == 0 or index == len(train_loader) - 1:
                time2 = time.time()
                tqdm.write(
                    f"{index}, epoch: {epoch} -loss: {str(loss)} ; each step's time spent: {(str(float(time2 - time1) / float(index + 0.0001)))}")
        # 验证
        model.eval()
        val_loss = validate_model(model, val_loader, device)
        writer.add_scalar('Loss/val', val_loss, epoch)
        print(f"val loss: {val_loss} , epoch: {epoch}")
        print("Save Model To ", model_output_dir)
        model.save_pretrained(model_output_dir)


def validate_model(model, device, val_loader):
    running_loss = 0.0
    with torch.no_grad():
        for _, data in enumerate(tqdm(val_loader, file=sys.stdout, desc="Validation Data")):
            input_ids = data['input_ids'].to(device, dtype=torch.long)
            attention_mask = data['attention_mask'].to(device, dtype=torch.long)
            labels = data['labels'].to(device, dtype=torch.long)
            outputs = model(
                input_ids=input_ids,
                attention_mask=attention_mask,
                labels=labels,
            )
            loss = outputs.loss
            running_loss += loss.item()
    return running_loss / len(val_loader)


def main():
    # 基础模型位置
    model_name = "model/Qwen2-1.5B-Instruct"
    # 训练集
    train_json_path = "./data/train.json"
    # 验证集
    val_json_path = "./data/val.json"
    max_source_length = 128
    max_target_length = 256
    epochs = 10
    batch_size = 1
    lr = 1e-4
    gradient_accumulation_steps = 16
    lora_rank = 8
    lora_alpha = 32
    model_output_dir = "output"
    logs_dir = "logs"
    # 设备
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    # 加载分词器和模型
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
    # setup peft
    peft_config = LoraConfig(
        task_type=TaskType.CAUSAL_LM,
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
        inference_mode=False,
        r=lora_rank,
        lora_alpha=lora_alpha,
        lora_dropout=0.1
    )
    model = get_peft_model(model, peft_config)
    model.is_parallelizable = True
    model.model_parallel = True
    model.print_trainable_parameters()
    print("Start Load Train Data...")
    train_params = {
        "batch_size": batch_size,
        "shuffle": True,
        "num_workers": 0,
    }
    training_set = QADataset(train_json_path, tokenizer, max_source_length, max_target_length)
    training_loader = DataLoader(training_set, **train_params)
    print("Start Load Validation Data...")
    val_params = {
        "batch_size": batch_size,
        "shuffle": False,
        "num_workers": 0,
    }
    val_set = QADataset(val_json_path, tokenizer, max_source_length, max_target_length)
    val_loader = DataLoader(val_set, **val_params)
    # 日志记录
    writer = SummaryWriter(logs_dir)
    # 优化器
    optimizer = torch.optim.AdamW(params=model.parameters(), lr=lr)
    model = model.to(device)
    # 开始训练
    print("Start Training...")
    train_model(
        model=model,
        train_loader=training_loader,
        val_loader=val_loader,
        optimizer=optimizer,
        gradient_accumulation_steps=gradient_accumulation_steps,
        device=device,
        num_epochs=epochs,
        model_output_dir=model_output_dir,
        writer=writer
    )
    writer.close()


if __name__ == '__main__':
    main()

训练过程：

训练结束后，可以在 output 中看到 lora 模型：

五、模型测试

# -*- coding: utf-8 -*-
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

model_path = "model/Qwen2-1.5B-Instruct"
lora_dir = "output"

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = PeftModel.from_pretrained(model, lora_dir)
model.to(device)

prompt = """
5月至今上腹靠右隐痛，右背隐痛带酸，便秘，喜睡，时有腹痛，头痛，腰酸症状？
"""
messages = [
    {"role": "system", "content": "你是一个医疗方面的专家，可以根据患者的问题进行解答。"},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
print(text)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=258)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

模型回答：根据你的叙述，胃炎胆汁反流性胃炎的可能性大，建议口服奥美拉唑，吗丁啉救治，清淡易消化饮食，忌辛辣打击食物，留意歇息，不要加班除了正规救治胃痛外，患者还需要有看重护理方面，比如恰当饮食，始终保持心情愉快。与此同时患者还要留意决定一家专业医院诊病，这样才能获得良好的治疗效果。

六、模型合并

上面测试还是分开加载的基础模型和lora模型，可以将两个合并为一个，方便后续部署：

# -*- coding: utf-8 -*-
import time
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from peft import PeftModel

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model_path = "model/Qwen2-1.5B-Instruct"
lora_dir = "output"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
model = PeftModel.from_pretrained(model, lora_dir).to(device)
print(model)
# 合并model, 同时保存 token
model = model.merge_and_unload()
model.save_pretrained("lora_output")
tokenizer.save_pretrained("lora_output")

合并后的结构：

后面就不需要再通过 PeftModel 直接加载模型既可使用：

# -*- coding: utf-8 -*-
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "lora_output"

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model.to(device)

prompt = """
5月至今上腹靠右隐痛，右背隐痛带酸，便秘，喜睡，时有腹痛，头痛，腰酸症状？
"""
messages = [
    {"role": "system", "content": "你是一个医疗方面的专家，可以根据患者的问题进行解答。"},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
print(text)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=258)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

​

如何学习大模型

现在社会上大模型越来越普及了，已经有很多人都想往这里面扎，但是却找不到适合的方法去学习。

作为一名资深码农，初入大模型时也吃了很多亏，踩了无数坑。现在我想把我的经验和知识分享给你们，帮助你们学习AI大模型，能够解决你们学习中的困难。

下面这些都是我当初辛苦整理和花钱购买的资料，现在我已将重要的AI大模型资料包括市面上AI大模型各大白皮书、AGI大模型系统学习路线、AI大模型视频教程、实战学习，等录播视频免费分享出来，需要的小伙伴可以扫取。

一、AGI大模型系统学习路线

很多人学习大模型的时候没有方向，东学一点西学一点，像只无头苍蝇乱撞，我下面分享的这个学习路线希望能够帮助到你们学习AI大模型。

二、AI大模型视频教程

三、AI大模型各大学习书籍

四、AI大模型各大场景实战案例

五、结束语

学习AI大模型是当前科技发展的趋势，它不仅能够为我们提供更多的机会和挑战，还能够让我们更好地理解和应用人工智能技术。通过学习AI大模型，我们可以深入了解深度学习、神经网络等核心概念，并将其应用于自然语言处理、计算机视觉、语音识别等领域。同时，掌握AI大模型还能够为我们的职业发展增添竞争力，成为未来技术领域的领导者。

再者，学习AI大模型也能为我们自己创造更多的价值，提供更多的岗位以及副业创收，让自己的生活更上一层楼。

因此，学习AI大模型是一项有前景且值得投入的时间和精力的重要选择。