Hugging Face's logo

Datasets:
KashiwaByte
/
DISC-Assignment

Dataset card Files Community
DISC-Assignment / FineTune /README.md
KashiwaByte's picture
KashiwaByte
initial commit
7c93e9d
preview code
|
raw
history blame
3.78 kB

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 image.png

模型文件

模型文件上传至HuggingFace

Read_Comprehension_Chatglm3-6b_qlora image.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

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 image.png