README.md 9.01 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
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
9
<div align="center">
zhaoying1's avatar
zhaoying1 committed
10
<img src="ptuning/media/GLMblock.png" width="550" height="200">
zhaoying1's avatar
zhaoying1 committed
11
12
13
14
15
16
17
18
19
20
21
</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
22

zhaoying1's avatar
zhaoying1 committed
23
24
25
<div align="center">
<img src="ptuning/media/GLM.png" width="550" height="200">
</div>
zhaoying1's avatar
zhaoying1 committed
26
27

## 环境配置
yuguo960516yuguo's avatar
readme  
yuguo960516yuguo committed
28

zhaoying1's avatar
zhaoying1 committed
29
### Docker(方式一)
zhaoying1's avatar
zhaoying1 committed
30
31
32
33
34
35
36
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:
```
zhaoying1's avatar
zhaoying1 committed
37
docker run -dit --network=host --name=chatglm --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 image.sourcefind.cn:5000/dcu/admin/base/pytorch:1.13.1-centos7.6-dtk-23.04-py38-latest
zhaoying1's avatar
zhaoying1 committed
38
pip install transformers==4.28.0 -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
zhaoying1's avatar
zhaoying1 committed
39
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
40
```
zhaoying1's avatar
zhaoying1 committed
41
42
43
44
45
46
### 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
47

zhaoying1's avatar
zhaoying1 committed
48
49
50
51
52
### Conda(方法三)
1. 创建conda虚拟环境:
```
conda create -n chatglm python=3.8
```
zhaoying1's avatar
zhaoying1 committed
53

zhaoying1's avatar
zhaoying1 committed
54
55
56
57
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
58

zhaoying1's avatar
zhaoying1 committed
59
    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
60

zhaoying1's avatar
zhaoying1 committed
61
62
63
64
3. 其它依赖库参照requirements.txt安装:
```
pip install -r requirements.txt
```
zhaoying1's avatar
zhaoying1 committed
65

zhaoying1's avatar
zhaoying1 committed
66
67
68
69
70
71
72
73
74
## 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
75
76
77
## 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
78

zhaoying1's avatar
zhaoying1 committed
79
80
81
## 训练

### P-tuning v2 微调训练
zhaoying1's avatar
zhaoying1 committed
82
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
83

zhaoying1's avatar
zhaoying1 committed
84
#### 单机多卡训练
zhaoying1's avatar
zhaoying1 committed
85
86
87
88
89
```
    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
90

zhaoying1's avatar
zhaoying1 committed
91
#### 推理测评
zhaoying1's avatar
zhaoying1 committed
92
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
93
```
zhaoying1's avatar
zhaoying1 committed
94
    cd ptuning
zhaoying1's avatar
zhaoying1 committed
95
96
    bash evaluate_ptuning.sh
```
zhaoying1's avatar
zhaoying1 committed
97
#### Results
zhaoying1's avatar
zhaoying1 committed
98
99
100
101
102
103
- 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
104
105
106
107
108

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

zhaoying1's avatar
zhaoying1 committed
109
### Finetune全参数微调
zhaoying1's avatar
zhaoying1 committed
110

zhaoying1's avatar
zhaoying1 committed
111
#### 单机多卡训练
zhaoying1's avatar
zhaoying1 committed
112
```
zhaoying1's avatar
zhaoying1 committed
113
114
    cd ptuning
    bash ft_train.sh
zhaoying1's avatar
zhaoying1 committed
115
116
```
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
117

zhaoying1's avatar
zhaoying1 committed
118
#### 集群训练
zhaoying1's avatar
zhaoying1 committed
119
120
```
    cd ptuning/slurm_scripts
zhaoying1's avatar
zhaoying1 committed
121
    bash run_train.sh
zhaoying1's avatar
zhaoying1 committed
122
123
```
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
124
125


zhaoying1's avatar
zhaoying1 committed
126
#### 推理测评
zhaoying1's avatar
zhaoying1 committed
127
```
zhaoying1's avatar
zhaoying1 committed
128
129
    cd ptuning
    bash evaluate_ft.sh
zhaoying1's avatar
zhaoying1 committed
130
```
zhaoying1's avatar
zhaoying1 committed
131
#### Results
zhaoying1's avatar
zhaoying1 committed
132
133
134
135
136
137
- 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
138
139
140
141
142
143
144
145
146
147
148
149
150
151

| 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
152
153


zhaoying1's avatar
zhaoying1 committed
154
155
### LoRA 微调训练
#### 单机多卡训练
zhaoying1's avatar
zhaoying1 committed
156
157
158
159
```
    cd ptuning
    bash lora_train.sh
```
zhaoying1's avatar
zhaoying1 committed
160
#### LoRA推理
zhaoying1's avatar
zhaoying1 committed
161
162
163
164
165
166
```
    python infer_lora.py
```



zhaoying1's avatar
zhaoying1 committed
167
## 推理
zhaoying1's avatar
zhaoying1 committed
168
169
170
运行如下命令:

    python cli_demo.py
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
171

zhaoying1's avatar
zhaoying1 committed
172
173
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
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
218
219
220
221
222
## 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
223
224
225
<div align="center">
<img src="./ptuning/media/pretrain.jpeg" width="400" height="300">
</div>
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
226
227
228
229


## 强化学习(RLHF)微调方案

yuguo960516yuguo's avatar
readme  
yuguo960516yuguo committed
230
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
231

yuguo960516yuguo's avatar
readme  
yuguo960516yuguo committed
232
- 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
233
234
- 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
235
236
237
238
239
240
241
242
243
244
245

## 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
246
247
## 源码仓库及问题反馈

zhaoying1's avatar
zhaoying1 committed
248
- https://developer.hpccube.com/codes/modelzoo/chatglm
yuguo960516yuguo's avatar
1.0  
yuguo960516yuguo committed
249

zhaoying1's avatar
zhaoying1 committed
250
251
## 参考

zhaoying1's avatar
zhaoying1 committed
252
- [THUDM/ChatGLM-6B](https://github.com/THUDM/ChatGLM-6B/tree/main)
zhaoying1's avatar
zhaoying1 committed
253