README.md 8.57 KB
Newer Older
zhaoying1's avatar
zhaoying1 committed
1
2
# ChatGLM-6B

zhaoying1's avatar
zhaoying1 committed
3
4
5
## 论文
`GLM: General Language Model Pretraining with Autoregressive Blank Infilling`
- [https://arxiv.org/abs/2103.10360](https://arxiv.org/abs/2103.10360)
zhaoying1's avatar
zhaoying1 committed
6

zhaoying1's avatar
zhaoying1 committed
7
## 模型结构
zhaoying1's avatar
zhaoying1 committed
8
9
ChatGLM-6B 是清华大学开源的开源的、支持中英双语的对话语言模型,基于 [General Language Model (GLM)](https://github.com/THUDM/GLM) 架构,具有 62 亿参数。ChatGLM-6B 使用了和 ChatGPT 相似的技术,针对中文问答和对话进行了优化。经过约 1T 标识符的中英双语训练,辅以监督微调、反馈自助、人类反馈强化学习等技术的加持,62 亿参数的 ChatGLM-6B 已经能生成相当符合人类偏好的回答。

zhaoying1's avatar
zhaoying1 committed
10
11
12
13
14
15
16
17
18
19
20
21
22
<div align="center">
<img src="ptuning/media/GLM.png" width="550" height="200">
</div>

以下是ChatGLM-6B的主要网络参数配置:


| 模型名称 | 隐含层维度 | 层数 | 头数 | 词表大小 | 位置编码 | 最大长 |
| -------- | -------- | -------- | -------- | -------- | -------- | -------- | 
|ChatGLM-6B | 4,096 | 28 | 32 | 130528 |  RoPE | 2048 |

## 算法原理
ChatGLM-6B基于GLM架构开发。GLM是一种基于Transformer的语言模型,以自回归空白填充为训练目标, 同时具备自回归和自编码能力。
zhaoying1's avatar
zhaoying1 committed
23
24
25


## 环境配置
yuguo960516yuguo's avatar
readme  
yuguo960516yuguo committed
26

zhaoying1's avatar
zhaoying1 committed
27
### Docker(方式一)
zhaoying1's avatar
zhaoying1 committed
28
29
30
31
32
33
34
35
推荐使用docker方式运行,提供拉取的docker镜像:
```
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:1.13.1-centos7.6-dtk-23.04-py38-latest
```

进入docker,安装docker中没有的依赖:
```
pip install transformers==4.28.0 -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
zhaoying1's avatar
zhaoying1 committed
36
pip install accelerate sentencepiece mdtex2html gradio rouge_chinese nltk jieba datasets protobuf peft -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
zhaoying1's avatar
zhaoying1 committed
37
```
zhaoying1's avatar
zhaoying1 committed
38
39
40
41
42
43
### Dockerfile(方式二)
```
docker build -t chatglm:latest .
docker run -dit --network=host --name=baichuan --privileged --device=/dev/kfd --device=/dev/dri --ipc=host --shm-size=16G  --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -u root --ulimit stack=-1:-1 --ulimit memlock=-1:-1 baichuan:latest
docker exec -it baichuan /bin/bash
```
zhaoying1's avatar
zhaoying1 committed
44

zhaoying1's avatar
zhaoying1 committed
45
46
47
48
49
### Conda(方法三)
1. 创建conda虚拟环境:
```
conda create -n chatglm python=3.8
```
zhaoying1's avatar
zhaoying1 committed
50

zhaoying1's avatar
zhaoying1 committed
51
52
53
54
2. 关于本项目DCU显卡所需的工具包、深度学习库等均可从[光合](https://developer.hpccube.com/tool/)开发者社区下载安装。
- [DTK 23.04](https://cancon.hpccube.com:65024/1/main/DTK-23.04.1)
- [Pytorch 1.13.1](https://cancon.hpccube.com:65024/4/main/pytorch/dtk23.04)
- [Deepspeed 0.9.2](https://cancon.hpccube.com:65024/4/main/deepspeed/dtk23.04)
zhaoying1's avatar
zhaoying1 committed
55

zhaoying1's avatar
zhaoying1 committed
56
    Tips:以上dtk驱动、python、deepspeed等工具版本需要严格一一对应。
zhaoying1's avatar
zhaoying1 committed
57

zhaoying1's avatar
zhaoying1 committed
58
59
60
61
3. 其它依赖库参照requirements.txt安装:
```
pip install -r requirements.txt
```
zhaoying1's avatar
zhaoying1 committed
62

zhaoying1's avatar
zhaoying1 committed
63
64
65
66
67
68
69
70
71
## 数据集
本仓库以 [ADGEN](https://aclanthology.org/D19-1321.pdf) (广告生成) 数据集为例介绍代码的使用方法,该数据集任务为根据输入(content)生成一段广告词(summary),以下为下载地址:
- [Google Drive](https://drive.google.com/file/d/13_vf0xRTQsyneRKdD1bZIr93vBGOczrk/view?usp=sharing) 或者 [Tsinghua Cloud](https://cloud.tsinghua.edu.cn/f/b3f119a008264b1cabd1/?dl=1)
下载处理好的 ADGEN 数据集,将解压后的AdvertiseGen目录放到 [ptuning](./ptuning)本目录下。数据集目录结构如下:
```
 ── AdvertiseGen
    │   ├── dev.json
    │   └── train.json
```
zhaoying1's avatar
zhaoying1 committed
72
73
74
## 模型下载
Hugging Face模型下载地址:
[ChatGLM-6B](https://huggingface.co/THUDM/chatglm-6b)
zhaoying1's avatar
zhaoying1 committed
75

zhaoying1's avatar
zhaoying1 committed
76
77
## P-tuning v2 微调训练
本仓库实现了对于ChatGLM-6B模型基于[P-Tuning v2](https://github.com/THUDM/P-tuning-v2)的微调。P-Tuning v2是由清华大学提出的一种高效参数微调方法。
zhaoying1's avatar
zhaoying1 committed
78

zhaoying1's avatar
zhaoying1 committed
79
80
81
82
83
84
### 单机多卡训练
```
    cd ptuning
    bash ptuning_train.sh
```
注意:请根据自己的需求配置其中的模型路径、数据集路径、batchsize、学习率等参数;
zhaoying1's avatar
zhaoying1 committed
85
86
87

### 推理测评
在 P-tuning v2 训练时模型只保存 PrefixEncoder 部分的参数,所以在推理时需要同时加载原 ChatGLM-6B 模型以及 PrefixEncoder 的权重,可直接运行一下命令:
zhaoying1's avatar
zhaoying1 committed
88
```
zhaoying1's avatar
zhaoying1 committed
89
    cd ptuning
zhaoying1's avatar
zhaoying1 committed
90
91
92
93
94
95
96
97
98
    bash evaluate_ptuning.sh
```
### Results
- 训练Loss
<div align="center">
<img src="./ptuning/media/6B_ds_pt_bs16_accum1_4cards_zero2_5e-3.jpg" width="400" height="300">
</div>

- 推理测试结果:
zhaoying1's avatar
zhaoying1 committed
99
100
101
102
103
104
105

| Checkpoint | Training Loss |BLEU-4 | Rouge-1 |  Rouge-2 | Rouge-l |
| :------: | :------: |:------: | :------: |:------: | :------: |
| 2000 steps |  3.57 | 7.9777 | 31.0344 |  6.981 | 24.7393 |

## Finetune全参数微调

zhaoying1's avatar
zhaoying1 committed
106
107
### 单机多卡训练
```
zhaoying1's avatar
zhaoying1 committed
108
109
    cd ptuning
    bash ft_train.sh
zhaoying1's avatar
zhaoying1 committed
110
111
```
注意:请根据自己的需求配置其中的模型路径、数据集路径、batchsize、学习率等参数;
zhaoying1's avatar
zhaoying1 committed
112

zhaoying1's avatar
zhaoying1 committed
113
114
115
116
117
118
### 集群训练
```
    cd ptuning/slurm_scripts
    bash run.shi
```
注意:请根据自己的需求配置其中的模型路径、数据集路径、batchsize、学习率等参数;
zhaoying1's avatar
zhaoying1 committed
119
120


zhaoying1's avatar
zhaoying1 committed
121
122
### 推理测评
```
zhaoying1's avatar
zhaoying1 committed
123
124
    cd ptuning
    bash evaluate_ft.sh
zhaoying1's avatar
zhaoying1 committed
125
```
zhaoying1's avatar
zhaoying1 committed
126
### Results
zhaoying1's avatar
zhaoying1 committed
127
128
129
130
131
132
- 训练Loss
<div align="center">
<img src="./ptuning/media/6B_ds_ft_bs32_accum1_4cards_zero3_5e-5.jpg" width="400" height="300">
</div>

- 推理测试结果:
zhaoying1's avatar
zhaoying1 committed
133
134
135
136
137
138
139
140
141
142
143
144
145
146

| Checkpoint | Training Loss |BLEU-4 | Rouge-1 |  Rouge-2 | Rouge-l |
| :------: | :------: |:------: | :------: |:------: | :------: |
| 3000 steps |  2.3398 | 7.6501 | 29.2229 | 6.466 | 23.8506 |


<!-- ## 评估结果
|               | Finetune | P-tuning v2 | LoRA |
| ------------- | ----------- | ----- | ------------- |
| BLEU-4        | 8.01    | 8.10 | 7.62 |
| Rouge-1       | 31.23  | 31.12 | 30.60 |
| Rouge-2       | 7.36    | 7.11 | 6.96 |
| Rouge-l       | 25.08  | 24.97 | 24.80 |
| Training Loss | 3.00 | 3.57 | 3.32 | -->
zhaoying1's avatar
zhaoying1 committed
147
148


zhaoying1's avatar
zhaoying1 committed
149
150
151
152
153
154
155
156
157
158
159
160
161
## LoRA 微调训练
### 单机多卡训练
```
    cd ptuning
    bash lora_train.sh
```
### LoRA推理
```
    python infer_lora.py
```



zhaoying1's avatar
zhaoying1 committed
162
## 推理
zhaoying1's avatar
zhaoying1 committed
163
164
165
运行如下命令:

    python cli_demo.py
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
166

zhaoying1's avatar
zhaoying1 committed
167
168
程序会在命令行中进行交互式的对话,在命令行中输入指示并回车即可生成回复,输入 clear 可以清空对话历史,输入 stop 终止程序。

yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
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
## 重新 pretrain

由于当前的 [GLM-130B](https://github.com/THUDM/GLM-130B#news) 与 ChatGLM 的模型结构非常类似,所以对于有训练 GLM-130B 的用户来说,可以通过修改 ChatGLM 的 config.json 使用堆参数的方式将参数量达到130B。该项目为了满足用户对 ChatGLM 重新 pretrain 的需求,继续添加了 simple-pretrain 目录,旨在提供一种改动最小的 pretrain 示例。pretrain步骤如下:

1. 将 simple-pretrain/ptuning 下的文件移到本 ptuning 目录下,替换相关文件

2. 将 modeling_chatglm.py 移到[ ChatGLM 模型](https://huggingface.co/THUDM/chatglm-6b)所在目录替换原始 modeling_chatglm.py 

3. 在本 ptuning 目录下:

   ```
   bash ds_pretrain.sh
   ```

说明:convert.py 可以将原始的txt数据转换成 chatglm 可用的 json 形式的数据集格式。该示例使用指环王1书籍作为预训练数据集。

### 实验设置

```
LR=1e-5

MASTER_PORT=$(shuf -n 1 -i 10000-65535)

HIP_VISIBLE_DEVICES=0,1,2,3 deepspeed --num_gpus=4 --master_port $MASTER_PORT main.py \
    --deepspeed deepspeed.json \
    --do_train \
    --train_file The-Lord-of-the-Rings-1.json \
    --prompt_column prompt \
    --response_column response \
    --overwrite_cache \
    --model_name_or_path THUDM/chatglm-6b \
    --output_dir ./output/pretrain \
    --overwrite_output_dir \
    --max_source_length 8 \
    --max_target_length 128 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --predict_with_generate \
    --max_steps 2000 \
    --logging_steps 5 \
    --save_steps 1000 \
    --learning_rate $LR \
    --fp16
```

### 训练loss收敛情况

由于该示例预训练数据集较小,loss会降的至较低水平到0.1左右。
zhaoying1's avatar
zhaoying1 committed
218
219
220
<div align="center">
<img src="./ptuning/media/pretrain.jpeg" width="400" height="300">
</div>
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
221
222
223
224


## 强化学习(RLHF)微调方案

yuguo960516yuguo's avatar
readme  
yuguo960516yuguo committed
225
目前在 DCU 上 ChatGLM 使用强化学习微调有两种方案可以走通:
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
226

yuguo960516yuguo's avatar
readme  
yuguo960516yuguo committed
227
- 使用 Lora,只更新低秩适应层,可以直接参考项目:https://github.com/hiyouga/ChatGLM-Efficient-Tuning/blob/main/examples/covid_doctor.md
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
228
229
- 使用 DeepSpeed-Chat 方案全参微调,目前已经适配完成,欢迎尝试:https://github.com/yuguo-Jack/ChatGLM-6B-in-DeepSpeed-Chat

zhaoying1's avatar
zhaoying1 committed
230
231
232
233
234
235
236
237
238
239
240

## 应用场景

### 算法类别

`自然语言处理`

### 热点应用行业

`nlp,智能聊天助手,科研`

yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
241
242
## 源码仓库及问题反馈

zhaoying1's avatar
zhaoying1 committed
243
- https://developer.hpccube.com/codes/modelzoo/chatglm
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
244

zhaoying1's avatar
zhaoying1 committed
245
246
## 参考

zhaoying1's avatar
zhaoying1 committed
247
- [THUDM/ChatGLM-6B](https://github.com/THUDM/ChatGLM-6B/tree/main)
zhaoying1's avatar
zhaoying1 committed
248