# 概述
我们采取了两种的不同的策略来微调模型，分别使用LoRA方法微调了MiniCPM-2B，使用QLoRA方法微调了ChatGLM3-6B

# ChatGLM3-6B QLoRA微调

## 依赖
Xtuner 集成 DeepSpeed 安装依赖：
```
pip install -U 'xtuner[deepspeed]'
```

## 模型训练

设置好模型位置和数据集后执行Xtuner命令进行训练
```
xtuner train ${CONFIG_NAME_OR_PATH}
```
显存消耗在13.5G左右,耗时约4个小时

![image.png](https://kashiwa-pic.oss-cn-beijing.aliyuncs.com/20240412091253.png)
![image.png](https://kashiwa-pic.oss-cn-beijing.aliyuncs.com/20240412085937.png)

## 模型文件
模型文件上传至HuggingFace

[Read_Comprehension_Chatglm3-6b_qlora](https://huggingface.co/KashiwaByte/Read_Comprehension_Chatglm3-6b_qlora/tree/main)
![image.png](https://kashiwa-pic.oss-cn-beijing.aliyuncs.com/20240414191810.png)
# MiniCPM-2B LoRA微调

设置好模型位置和数据集后运行train.sh脚本进行训练，训练显存消耗在21.5G左右,耗时约6个小时
![image.png](https://kashiwa-pic.oss-cn-beijing.aliyuncs.com/20240411224008.png)


以下分别是 train脚本的argparse超参数，Deepspeed配置和LoRA配置

    #train argparse
    --deepspeed ./ds_config.json \
    --output_dir="./output/MiniCPM" \
    --per_device_train_batch_size=4 \
    --gradient_accumulation_steps=4 \
    --logging_steps=10 \
    --num_train_epochs=3 \
    --save_steps=500 \
    --learning_rate=1e-4 \
    --save_on_each_node=True \


    # deepspeed config
    {
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": "auto",
            "weight_decay": "auto"
        }
    },

    "scheduler": {
        "type": "WarmupDecayLR",
        "params": {
            "last_batch_iteration": -1,
            "total_num_steps": "auto",
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },

    "zero_optimization": {
        "stage": 2,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        },
        "offload_param": {
            "device": "cpu",
            "pin_memory": true
        },
        "allgather_partitions": true,
        "allgather_bucket_size": 5e8,
        "overlap_comm": true,
        "reduce_scatter": true,
        "reduce_bucket_size": 5e8,
        "contiguous_gradients": true
    },
    "activation_checkpointing": {
        "partition_activations": false,
        "cpu_checkpointing": false,
        "contiguous_memory_optimization": false,
        "number_checkpoints": null,
        "synchronize_checkpoint_boundary": false,
        "profile": false
    },
    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "steps_per_print": 2000,
    "train_batch_size": "auto",
    "min_lr": 5e-7,
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
    }

    # loraConfig
    config = LoraConfig(
    task_type=TaskType.CAUSAL_LM, 
    target_modules=["q_proj", "v_proj"],  # 这个不同的模型需要设置不同的参数，需要看模型中的attention层
    inference_mode=False, # 训练模式
    r=8, # Lora 秩
    lora_alpha=32, # Lora alaph，具体作用参见 Lora 原理
    lora_dropout=0.1# Dropout 比例
    )



## 模型文件
模型文件上传至HuggingFace

[Read_Comprehension_MiniCPM2B](https://huggingface.co/KashiwaByte/Read_Comprehension_MiniCPM2B/tree/main)
![image.png](https://kashiwa-pic.oss-cn-beijing.aliyuncs.com/20240414191757.png)