README.md 10.8 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

dcuai's avatar
dcuai committed
8
9
# ChatGLM

zhuwenwen's avatar
zhuwenwen committed
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
## 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
30
| glm-4-9b    | 4096       | 40   | 32   | 151552   | RoPE     | 131072       |
zhuwenwen's avatar
zhuwenwen committed
31
32
33
34
35
36
37
38
39
40
41
42

## 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
43

zhuwenwen's avatar
zhuwenwen committed
44
45
46
提供[光源](https://www.sourcefind.cn/#/image/dcu/custom)拉取推理的docker镜像:

```
47
docker pull image.sourcefind.cn:5000/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.1-rc5-rocblas104381-0915-das1.6-py3.10-20250916-rc2
zhuwenwen's avatar
zhuwenwen committed
48
49
50
# <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
51
# 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
52
53
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
54

zhuwenwen's avatar
zhuwenwen committed
55
`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
56
57

### Dockerfile(方法二)
laibao's avatar
laibao committed
58

zhuwenwen's avatar
zhuwenwen committed
59
60
61
62
```
# <Host Path>主机端路径
# <Container Path>容器映射路径
docker build -t chatglm:latest .
63
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> chatglm:latest /bin/bash
zhuwenwen's avatar
zhuwenwen committed
64
65
66
```

### Anaconda(方法三)
laibao's avatar
laibao committed
67

zhuwenwen's avatar
zhuwenwen committed
68
69
70
```
conda create -n chatglm_vllm python=3.10
```
laibao's avatar
laibao committed
71

72
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
73

74
75
76
77
* DTK驱动:dtk25.04.01
* Pytorch: 2.4.0
* triton: 3.0.0
* lmslim: 0.2.1
laibao's avatar
laibao committed
78
* flash_attn: 2.6.1
79
* flash_mla: 1.0.0
80
* vllm: 0.9.2
zhuwenwen's avatar
zhuwenwen committed
81
82
* python: python3.10

83
`Tips:需先安装相关依赖,最后安装vllm包`
84

85
86
87
88
89
90
91
92
93
94
95
环境变量:
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export VLLM_NUMA_BIND=1
export VLLM_RANK0_NUMA=0
export VLLM_RANK1_NUMA=1
export VLLM_RANK2_NUMA=2
export VLLM_RANK3_NUMA=3
export VLLM_RANK4_NUMA=4
export VLLM_RANK5_NUMA=5
export VLLM_RANK6_NUMA=6
export VLLM_RANK7_NUMA=7
zhuwenwen's avatar
zhuwenwen committed
96
97

## 数据集
laibao's avatar
laibao committed
98

zhuwenwen's avatar
zhuwenwen committed
99
100
101
102
103
104


## 推理

### 模型下载

105
106
107
108
109
| 基座模型                                                 | 长文本模型                                                             |
| -------------------------------------------------------- | ---------------------------------------------------------------------- |
| [chatglm2-6b](https://huggingface.co/THUDM/chatglm2-6b)     | [chatglm2-6b-32k](https://huggingface.co/THUDM/chatglm2-6b-32k)           |
| [chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b)     | [chatglm3-6b-32k](https://huggingface.co/THUDM/chatglm3-6b-32k)           |
| [glm-4-9b-chat](https://huggingface.co/THUDM/glm-4-9b-chat) | [glm-4-9b-chat-1m](https://modelscope.cn/models/ZhipuAI/glm-4-9b-chat-1m) |
zhuwenwen's avatar
zhuwenwen committed
110
111

### 离线批量推理
laibao's avatar
laibao committed
112

zhuwenwen's avatar
zhuwenwen committed
113
```bash
114
 python examples/offline_inference/basic/basic.py
zhuwenwen's avatar
zhuwenwen committed
115
```
laibao's avatar
laibao committed
116

zhuwenwen's avatar
zhuwenwen committed
117
118
119
120
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
121

zhuwenwen's avatar
zhuwenwen committed
122
1、指定输入输出
laibao's avatar
laibao committed
123

zhuwenwen's avatar
zhuwenwen committed
124
```bash
125
 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
126
```
laibao's avatar
laibao committed
127
128

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。
panhb's avatar
panhb committed
129
glm-4-9b-chat-1m模型默认的model_max_length为1024000,官方vllm也尚不支持该长度,模型启动时必须添加--max_model_len(包括后面的启动命令), 经测试,500000左右也可以正常进行推理。
zhuwenwen's avatar
zhuwenwen committed
130
131
132

2、使用数据集
下载数据集:
133
[sharegpt_v3_unfiltered_cleaned_split](https://huggingface.co/datasets/learnanything/sharegpt_v3_unfiltered_cleaned_split)
laibao's avatar
laibao committed
134

zhuwenwen's avatar
zhuwenwen committed
135
```bash
136
 python benchmarks/benchmark_throughput.py --num-prompts 1 --model THUDM/glm-4-9b-chat --dataset-name sharegpt --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json -tp 1 --trust-remote-code --enforce-eager --dtype float16
zhuwenwen's avatar
zhuwenwen committed
137
138
139
```

```bash
zhuwenwen's avatar
zhuwenwen committed
140
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
141
142
```

laibao's avatar
laibao committed
143
144
145
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
146
147

1、启动服务端:
laibao's avatar
laibao committed
148

zhuwenwen's avatar
zhuwenwen committed
149
```bash
150
 vllm serve --model THUDM/glm-4-9b-chat --enforce-eager --dtype float16 --trust-remote-code --tensor-parallel-size 1
zhuwenwen's avatar
zhuwenwen committed
151
152
153
```

2、启动客户端:
laibao's avatar
laibao committed
154

155
156
```
python benchmarks/benchmark_serving.py --model THUDM/glm-4-9b-chat --dataset-name sharegpt --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json  --num-prompts 1 --trust-remote-code
zhuwenwen's avatar
zhuwenwen committed
157
```
158

159
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
160
161

### OpenAI兼容服务
laibao's avatar
laibao committed
162

zhuwenwen's avatar
zhuwenwen committed
163
启动服务:
laibao's avatar
laibao committed
164

zhuwenwen's avatar
zhuwenwen committed
165
```bash
166
 vllm serve THUDM/glm-4-9b-chat --enforce-eager --dtype float16 --trust-remote-code
zhuwenwen's avatar
zhuwenwen committed
167
```
laibao's avatar
laibao committed
168
169

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
170
171

列出模型型号:
laibao's avatar
laibao committed
172

zhuwenwen's avatar
zhuwenwen committed
173
174
175
176
177
```bash
curl http://localhost:8000/v1/models
```

### OpenAI Completions API和vllm结合使用
laibao's avatar
laibao committed
178

zhuwenwen's avatar
zhuwenwen committed
179
180
181
182
183
184
185
186
187
188
189
```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
190
或者使用[examples/openai_completion_client.py](examples/openai_completion_client.py)
zhuwenwen's avatar
zhuwenwen committed
191
192

### OpenAI Chat API和vllm结合使用
laibao's avatar
laibao committed
193

zhuwenwen's avatar
zhuwenwen committed
194
195
```bash
curl http://localhost:8000/v1/chat/completions \
196
197
198
199
200
201
202
203
204
205
206
207
  -X POST \
  -H "Content-Type: application/json" \
  -d '{
    "model": "THUDM/glm-4-9b-chat",
    "max_tokens": 128,
    "messages": [
      {
        "role": "user",
        "content": "晚上睡不着怎么办"
      }
    ]
  }'
zhuwenwen's avatar
zhuwenwen committed
208
```
laibao's avatar
laibao committed
209

210
或者使用[examples/online_serving/openai_chat_completion_client.py](examples/online_serving/openai_chat_completion_client.py)
laibao's avatar
laibao committed
211

laibao's avatar
laibao committed
212
### **gradio和vllm结合使用**
zhuwenwen's avatar
zhuwenwen committed
213

laibao's avatar
laibao committed
214
215
216
217
218
219
220
221
222
223
224
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

```
225
python examples/online_serving/gradio_openai_chatbot_webserver.py --model "THUDM/glm-4-9b-chat" --model-url http://localhost:8000/v1 --temp 0.8 --stop-token-ids ""
laibao's avatar
laibao committed
226
227
228
229
230
231
232
233
234
```

    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
235

laibao's avatar
laibao committed
236
237
238
239
    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
240
```
laibao's avatar
laibao committed
241
242
243
244

3.启动OpenAI兼容服务

```
245
 vllm serve THUDM/glm-4-9b-chat --enforce-eager --dtype float16 --trust-remote-code --host "0.0.0.0"
laibao's avatar
laibao committed
246
247
248
249
250
```

4.启动gradio服务

```
laibao's avatar
laibao committed
251
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
252
253
254
255
256
```

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
257
258

## result
laibao's avatar
laibao committed
259

zhuwenwen's avatar
zhuwenwen committed
260
使用的加速卡:1张 DCU-K100_AI-64G
laibao's avatar
laibao committed
261

zhuwenwen's avatar
zhuwenwen committed
262
263
264
265
266
```
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
267

zhuwenwen's avatar
zhuwenwen committed
268
269
270
271
272


## 应用场景

### 算法类别
laibao's avatar
laibao committed
273

zhuwenwen's avatar
zhuwenwen committed
274
275
276
对话问答

### 热点应用行业
laibao's avatar
laibao committed
277

zhuwenwen's avatar
zhuwenwen committed
278
279
280
医疗,金融,科研,教育

## 源码仓库及问题反馈
laibao's avatar
laibao committed
281

chenzk's avatar
chenzk committed
282
* [https://developer.sourcefind.cn/codes/modelzoo/llama_vllm](https://developer.sourcefind.cn/codes/modelzoo/chatglm_vllm)
zhuwenwen's avatar
zhuwenwen committed
283
284

## 参考资料
laibao's avatar
laibao committed
285

zhuwenwen's avatar
zhuwenwen committed
286
* [https://github.com/vllm-project/vllm](https://github.com/vllm-project/vllm)
laibao's avatar
laibao committed
287
* [https://github.com/THUDM/ChatGLM3](https://github.com/THUDM/ChatGLM3)