README.md 10.2 KB
Newer Older
zhuwenwen's avatar
zhuwenwen committed
1
2
3
4
<!--
 * @Author: zhuww
 * @email: zhuww@sugon.com
 * @Date: 2024-06-13 14:38:07
zhuwenwen's avatar
zhuwenwen committed
5
 * @LastEditTime: 2024-09-30 09:16:01
zhuwenwen's avatar
zhuwenwen committed
6
-->
laibao's avatar
laibao committed
7

zhuwenwen's avatar
zhuwenwen committed
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
## 论文

`GLM: General Language Model Pretraining with Autoregressive Blank Infilling`

- [https://arxiv.org/abs/2103.10360](https://arxiv.org/abs/2103.10360)

## 模型结构

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。ChatGLM2-6B 是开源中英双语对话模型 ChatGLM-6B 的第二代版本,ChatGLM3 是智谱AI和清华大学 KEG 实验室联合发布的新一代对话预训练模型。ChatGLM3-6B 是 ChatGLM3 系列中的开源模型,在保留了前两代模型对话流畅、部署门槛低等众多优秀特性的基础上,ChatGLM3-6B 具有更强大的基础模型、更完整的功能支持、更全面的开源序列。

<div align="center">
<img src="docs/transformers.jpg" width="300" height="400">
</div>

以下是ChatGLM系列模型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 | 最大序列长度 |
| ----------- | ---------- | ---- | ---- | -------- | -------- | ------------ |
| ChatGLM2-6B | 4096       | 28   | 32   | 65024    | RoPE     | 8192         |
| ChatGLM3-6B | 4096       | 28   | 32   | 65024    | RoPE     | 8192         |
laibao's avatar
laibao committed
28
| glm-4-9b    | 4096       | 40   | 32   | 151552   | RoPE     | 131072       |
zhuwenwen's avatar
zhuwenwen committed
29
30
31
32
33
34
35
36
37
38
39
40

## 算法原理

ChatGLM系列模型基于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。

<div align="center">
<img src="docs/GLM.png" width="550" height="200">
</div>

## 环境配置

### Docker(方法一)
laibao's avatar
laibao committed
41

zhuwenwen's avatar
zhuwenwen committed
42
43
44
提供[光源](https://www.sourcefind.cn/#/image/dcu/custom)拉取推理的docker镜像:

```
laibao's avatar
laibao committed
45
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.3.0-py3.10-dtk24.04.3-ubuntu20.04-vllm0.6
zhuwenwen's avatar
zhuwenwen committed
46
47
48
# <Image ID>用上面拉取docker镜像的ID替换
# <Host Path>主机端路径
# <Container Path>容器映射路径
zhuwenwen's avatar
zhuwenwen committed
49
# 若要在主机端和容器端映射端口需要删除--network host参数
zhuwenwen's avatar
zhuwenwen committed
50
51
docker run -it --name chatglm_vllm --privileged --shm-size=64G  --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal -v <Host Path>:<Container Path> <Image ID> /bin/bash
```
laibao's avatar
laibao committed
52

zhuwenwen's avatar
zhuwenwen committed
53
`Tips:若在K100/Z100L上使用,使用定制镜像docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:vllm0.5.0-dtk24.04.1-ubuntu20.04-py310-zk-v1,K100/Z100L不支持awq量化`
zhuwenwen's avatar
zhuwenwen committed
54
55

### Dockerfile(方法二)
laibao's avatar
laibao committed
56

zhuwenwen's avatar
zhuwenwen committed
57
58
59
60
```
# <Host Path>主机端路径
# <Container Path>容器映射路径
docker build -t chatglm:latest .
zhuwenwen's avatar
zhuwenwen committed
61
docker run -it --name chatglm_vllm --privileged --shm-size=64G  --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal:ro -v <Host Path>:<Container Path> llama:latest /bin/bash
zhuwenwen's avatar
zhuwenwen committed
62
63
64
```

### Anaconda(方法三)
laibao's avatar
laibao committed
65

zhuwenwen's avatar
zhuwenwen committed
66
67
68
```
conda create -n chatglm_vllm python=3.10
```
laibao's avatar
laibao committed
69

zhuwenwen's avatar
zhuwenwen committed
70
关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.hpccube.com/tool/)开发者社区下载安装。
laibao's avatar
laibao committed
71
72
73
74
75
76
77

* DTK驱动:dtk24.04.3
* Pytorch: 2.3.0
* triton: 2.1.0
* lmslim: 0.1.2
* flash_attn: 2.6.1
* vllm: 0.6.2
zhuwenwen's avatar
zhuwenwen committed
78
79
* python: python3.10

zhuwenwen's avatar
zhuwenwen committed
80
`Tips:需先安装相关依赖,最后安装vllm包`
zhuwenwen's avatar
zhuwenwen committed
81
82

## 数据集
laibao's avatar
laibao committed
83

zhuwenwen's avatar
zhuwenwen committed
84
85
86
87
88
89


## 推理

### 模型下载

laibao's avatar
laibao committed
90
91
92
93
94
| 基座模型                                                             | 长文本模型                                                                     |
| -------------------------------------------------------------------- | ------------------------------------------------------------------------------ |
| [chatglm2-6b](http://113.200.138.88:18080/aimodels/chatglm2-6b)         | [chatglm2-6b-32k](http://113.200.138.88:18080/aimodels/thudm/chatglm2-6b-32k.git) |
| [chatglm3-6b](http://113.200.138.88:18080/aimodels/chatglm3-6b)         | [chatglm3-6b-32k](http://113.200.138.88:18080/aimodels/chatglm3-6b-32k)           |
| [glm-4-9b-chat](http://113.200.138.88:18080/aimodels/glm-4-9b-chat.git) |                                                                                |
zhuwenwen's avatar
zhuwenwen committed
95
96

### 离线批量推理
laibao's avatar
laibao committed
97

zhuwenwen's avatar
zhuwenwen committed
98
```bash
zhuwenwen's avatar
zhuwenwen committed
99
python examples/offline_inference.py
zhuwenwen's avatar
zhuwenwen committed
100
```
laibao's avatar
laibao committed
101

zhuwenwen's avatar
zhuwenwen committed
102
103
104
105
其中,`prompts`为提示词;`temperature`为控制采样随机性的值,值越小模型生成越确定,值变高模型生成更随机,0表示贪婪采样,默认为1;`max_tokens=16`为生成长度,默认为1;
`model`为模型路径;`tensor_parallel_size=1`为使用卡数,默认为1;`dtype="float16"`为推理数据类型,如果模型权重是bfloat16,需要修改为float16推理,`quantization="gptq"`为使用gptq量化进行推理,需下载以上GPTQ模型。

### 离线批量推理性能测试
laibao's avatar
laibao committed
106

zhuwenwen's avatar
zhuwenwen committed
107
1、指定输入输出
laibao's avatar
laibao committed
108

zhuwenwen's avatar
zhuwenwen committed
109
```bash
zhuwenwen's avatar
zhuwenwen committed
110
python benchmarks/benchmark_throughput.py --num-prompts 1 --input-len 32 --output-len 128 --model THUDM/glm-4-9b-chat -tp 1 --trust-remote-code --enforce-eager --dtype float16
zhuwenwen's avatar
zhuwenwen committed
111
```
laibao's avatar
laibao committed
112
113

其中 `--num-prompts`是batch数,`--input-len`是输入seqlen,`--output-len`是输出token长度,`--model`为模型路径,`-tp`为使用卡数,`dtype="float16"`为推理数据类型,如果模型权重是bfloat16,需要修改为float16推理。若指定 `--output-len  1`即为首字延迟。`-q gptq`为使用gptq量化模型进行推理。
zhuwenwen's avatar
zhuwenwen committed
114
115
116

2、使用数据集
下载数据集:
laibao's avatar
laibao committed
117

zhuwenwen's avatar
zhuwenwen committed
118
```bash
dcuai's avatar
dcuai committed
119
wget http://113.200.138.88:18080/aidatasets/vllm_data/-/raw/main/ShareGPT_V3_unfiltered_cleaned_split.json
zhuwenwen's avatar
zhuwenwen committed
120
121
122
```

```bash
zhuwenwen's avatar
zhuwenwen committed
123
python benchmarks/benchmark_throughput.py --num-prompts 1 --model THUDM/glm-4-9b-chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json -tp 1 --trust-remote-code --enforce-eager --dtype float16
zhuwenwen's avatar
zhuwenwen committed
124
125
```

laibao's avatar
laibao committed
126
127
128
其中 `--num-prompts`是batch数,`--model`为模型路径,`--dataset`为使用的数据集,`-tp`为使用卡数,`dtype="float16"`为推理数据类型,如果模型权重是bfloat16,需要修改为float16推理。`-q gptq`为使用gptq量化模型进行推理。

### OpenAI api服务推理性能测试
zhuwenwen's avatar
zhuwenwen committed
129
130

1、启动服务端:
laibao's avatar
laibao committed
131

zhuwenwen's avatar
zhuwenwen committed
132
```bash
zhuwenwen's avatar
zhuwenwen committed
133
python -m vllm.entrypoints.openai.api_server  --model THUDM/glm-4-9b-chat  --dtype float16 --enforce-eager -tp 1 
zhuwenwen's avatar
zhuwenwen committed
134
135
136
```

2、启动客户端:
laibao's avatar
laibao committed
137

zhuwenwen's avatar
zhuwenwen committed
138
```bash
zhuwenwen's avatar
zhuwenwen committed
139
python benchmarks/benchmark_serving.py --model THUDM/glm-4-9b-chat --dataset ShareGPT_V3_unfiltered_cleaned_split.json  --num-prompts 1 --trust-remote-code
zhuwenwen's avatar
zhuwenwen committed
140
141
```

laibao's avatar
laibao committed
142
参数同使用数据集,离线批量推理性能测试,具体参考[benchmarks/benchmark_serving.py](benchmarks/benchmark_serving.py)
zhuwenwen's avatar
zhuwenwen committed
143
144

### OpenAI兼容服务
laibao's avatar
laibao committed
145

zhuwenwen's avatar
zhuwenwen committed
146
启动服务:
laibao's avatar
laibao committed
147

zhuwenwen's avatar
zhuwenwen committed
148
```bash
laibao's avatar
laibao committed
149
vllm serve THUDM/glm-4-9b-chat --enforce-eager --dtype float16 --trust-remote-code --chat-template template_chatglm2.jinja --port 8000
zhuwenwen's avatar
zhuwenwen committed
150
```
laibao's avatar
laibao committed
151
152

这里serve之后 为加载模型路径,`--dtype`为数据类型:float16,默认情况使用tokenizer中的预定义聊天模板,`--chat-template`可以添加新模板覆盖默认模板,`-q gptq`为使用gptq量化模型进行推理。
zhuwenwen's avatar
zhuwenwen committed
153
154

列出模型型号:
laibao's avatar
laibao committed
155

zhuwenwen's avatar
zhuwenwen committed
156
157
158
159
160
```bash
curl http://localhost:8000/v1/models
```

### OpenAI Completions API和vllm结合使用
laibao's avatar
laibao committed
161

zhuwenwen's avatar
zhuwenwen committed
162
163
164
165
166
167
168
169
170
171
172
```bash
curl http://localhost:8000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "THUDM/glm-4-9b-chat",
        "prompt": "晚上睡不着怎么办",
        "max_tokens": 7,
        "temperature": 0
    }'
```

laibao's avatar
laibao committed
173
或者使用[examples/openai_completion_client.py](examples/openai_completion_client.py)
zhuwenwen's avatar
zhuwenwen committed
174
175

### OpenAI Chat API和vllm结合使用
laibao's avatar
laibao committed
176

zhuwenwen's avatar
zhuwenwen committed
177
178
179
180
181
182
183
184
185
186
187
```bash
curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "THUDM/glm-4-9b-chat",
        "messages": [
            {"role": "system", "content": "晚上睡不着怎么办"},
            {"role": "user", "content": "晚上睡不着怎么办"}
        ]
    }'
```
laibao's avatar
laibao committed
188

zhuwenwen's avatar
zhuwenwen committed
189
或者使用[examples/openai_chatcompletion_client.py](examples/openai_chatcompletion_client.py)
laibao's avatar
laibao committed
190

laibao's avatar
laibao committed
191
### **gradio和vllm结合使用**
zhuwenwen's avatar
zhuwenwen committed
192

laibao's avatar
laibao committed
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
1.安装gradio

```
pip install gradio
```

2.安装必要文件

    2.1 启动gradio服务,根据提示操作

```
python  gradio_openai_chatbot_webserver.py --model "THUDM/glm-4-9b-chat" --model-url http://localhost:8000/v1 --temp 0.8 --stop-token-ids ""
```

    2.2 更改文件权限

打开提示下载文件目录,输入以下命令给予权限

```
chmod +x frpc_linux_amd64_v0.*
```
laibao's avatar
laibao committed
214

laibao's avatar
laibao committed
215
216
217
218
    2.3端口映射

```
ssh -L 8000:计算节点IP:8000 -L 8001:计算节点IP:8001 用户名@登录节点 -p 登录节点端口
laibao's avatar
laibao committed
219
```
laibao's avatar
laibao committed
220
221
222
223

3.启动OpenAI兼容服务

```
laibao's avatar
laibao committed
224
vllm serve THUDM/glm-4-9b-chat  --enforce-eager --dtype float16 --trust-remote-code --chat-template template_chatglm2.jinja --port 8000
laibao's avatar
laibao committed
225
226
227
228
229
```

4.启动gradio服务

```
laibao's avatar
laibao committed
230
python  gradio_openai_chatbot_webserver.py --model "THUDM/glm-4-9b-chat" --model-url http://localhost:8000/v1 --temp 0.8 --stop-token-ids --host "0.0.0.0" --port 8001"
laibao's avatar
laibao committed
231
232
233
234
235
```

5.使用对话服务

在浏览器中输入本地 URL,可以使用 Gradio 提供的对话服务。
zhuwenwen's avatar
zhuwenwen committed
236
237

## result
laibao's avatar
laibao committed
238

zhuwenwen's avatar
zhuwenwen committed
239
使用的加速卡:1张 DCU-K100_AI-64G
laibao's avatar
laibao committed
240

zhuwenwen's avatar
zhuwenwen committed
241
242
243
244
245
```
Prompt: '晚上睡不着怎么办', Generated text: '?\n晚上睡不着可以尝试以下方法来改善睡眠质量:\n\n1. **调整作息时间**:尽量每天同一时间上床睡觉和起床,建立规律的生物钟。\n\n2. **放松身心**:睡前进行深呼吸、冥想或瑜伽等放松活动,有助于减轻压力和焦虑。\n\n3. **避免咖啡因和酒精**:晚上避免摄入咖啡因和酒精,因为它们可能会干扰睡眠。\n\n'
```

### 精度
laibao's avatar
laibao committed
246

zhuwenwen's avatar
zhuwenwen committed
247
248
249
250
251


## 应用场景

### 算法类别
laibao's avatar
laibao committed
252

zhuwenwen's avatar
zhuwenwen committed
253
254
255
对话问答

### 热点应用行业
laibao's avatar
laibao committed
256

zhuwenwen's avatar
zhuwenwen committed
257
258
259
医疗,金融,科研,教育

## 源码仓库及问题反馈
laibao's avatar
laibao committed
260

zhuwenwen's avatar
zhuwenwen committed
261
262
263
* [https://developer.hpccube.com/codes/modelzoo/llama_vllm](https://developer.hpccube.com/codes/modelzoo/chatglm_vllm)

## 参考资料
laibao's avatar
laibao committed
264

zhuwenwen's avatar
zhuwenwen committed
265
* [https://github.com/vllm-project/vllm](https://github.com/vllm-project/vllm)
laibao's avatar
laibao committed
266
* [https://github.com/THUDM/ChatGLM3](https://github.com/THUDM/ChatGLM3)