README.md 8.57 KB
Newer Older
zhuwenwen's avatar
zhuwenwen committed
1
2
3
4
5
6
<!--
 * @Author: guanyu
 * @email: guanyu1@sugon.com
 * @Date: 2025-05-04 14:15:07
-->

zhuwenwen's avatar
zhuwenwen committed
7
# Qwen3
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
28
29
30
31
32
33
34
35

## 论文



## 模型结构

Qwen3是阿里巴巴集团Qwen团队研发的大语言模型和大型多模态模型系列。无论是语言模型还是多模态模型,均在大规模多语言和多模态数据上进行预训练,并通过高质量数据进行后期微调以贴近人类偏好。Qwen3具备自然语言理解、文本生成、视觉理解、音频理解、工具使用、角色扮演、作为AI Agent进行互动等多种能力 。

<div align=center>
    <img src="./doc/qwen3.jpg"/>
</div>

## 算法原理

Qwen3是一个decoder-only的transformer模型,使用SwiGLU激活函数、RoPE、多头注意力机制等。

<div align=center>
    <img src="./doc/qwen3.png"/>
</div>

## 环境配置

### Docker(方法一)

提供[光源](https://www.sourcefind.cn/#/image/dcu/custom)拉取推理的docker镜像:

```
36
docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:vllm0.8.5-ubuntu22.04-dtk25.04.1-rc5-das1.6-py3.10-20250724
zhuwenwen's avatar
zhuwenwen committed
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
# <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
# 若要在主机端和容器端映射端口需要删除--network host参数
docker run -it --name qwen3_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
```

### Dockerfile(方法二)

```
# <Host Path>主机端路径
# <Container Path>容器映射路径
docker build -t qwen3:latest .
docker run -it --name qwen3_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> qwen3:latest /bin/bash
```

zhuwenwen's avatar
zhuwenwen committed
53
54
55
56
57
### Anaconda(方法三)

```
conda create -n qwen3_vllm python=3.10
```
zhuwenwen's avatar
zhuwenwen committed
58

zhuwenwen's avatar
zhuwenwen committed
59
60
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。

61
* DTK驱动:dtk25.04.01
zhuwenwen's avatar
zhuwenwen committed
62
63
64
65
66
* Pytorch: 2.4.0
* triton: 3.0.0
* lmslim: 0.2.1
* flash_attn: 2.6.1
* flash_mla: 1.0.0
67
* vllm: 0.8.5
zhuwenwen's avatar
zhuwenwen committed
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
* python: python3.10

`Tips:需先安装相关依赖,最后安装vllm包` 

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
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106

## 数据集



## 推理

### 模型下载

| 基座模型                                                                         |
| -------------------------------------------------------------------------------- |
| [Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B)                  |
| [Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B)                  |
| [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B)                |
| [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)                |
| [Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B)                |
| [Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) |
| [Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B)      |
| [Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B)      |
 

### 离线批量推理

```bash
107
VLLM_USE_FLASH_ATTN_PA=1 python examples/offline_inference/basic/basic.py
zhuwenwen's avatar
zhuwenwen committed
108
109
```

zhuwenwen's avatar
zhuwenwen committed
110
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16;
111
`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。`block_size`为块大小,默认为64;
zhuwenwen's avatar
zhuwenwen committed
112
113
114
115
116
117

### 离线批量推理性能测试

1、指定输入输出

```bash
118
VLLM_USE_FLASH_ATTN_PA=1 python benchmarks/benchmark_throughput.py --num-prompts 1 --input-len 32 --output-len 128 --model /your/model/path -tp 1 --trust-remote-code --enforce-eager --dtype float16
zhuwenwen's avatar
zhuwenwen committed
119
120
```

121
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。若模型权重为 bfloat16,建议设置 `--dtype bfloat16` 或使用 `--dtype auto` 以匹配权重精度。若指定 `--output-len  1`即为首字延迟。
zhuwenwen's avatar
zhuwenwen committed
122
123
124
125
126
127

2、使用数据集
下载数据集:
[sharegpt_v3_unfiltered_cleaned_split](https://huggingface.co/datasets/learnanything/sharegpt_v3_unfiltered_cleaned_split)

```bash
128
VLLM_USE_FLASH_ATTN_PA=1 python benchmarks/benchmark_throughput.py --num-prompts 1 --model /your/model/path --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
129
130
```

131
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。若模型权重为 bfloat16,建议设置 `--dtype bfloat16` 或使用 `--dtype auto` 以匹配权重精度。
zhuwenwen's avatar
zhuwenwen committed
132
133
134
135
136
137

### OpenAI api服务推理性能测试

1.启动服务:

```bash
138
VLLM_USE_FLASH_ATTN_PA=1 vllm serve --model /your/model/path --enforce-eager --dtype float16 --trust-remote-code --tensor-parallel-size 1
zhuwenwen's avatar
zhuwenwen committed
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
```

2.启动客户端

```
python benchmarks/benchmark_serving.py --model /your/model/path --dataset-name sharegpt --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json  --num-prompts 1 --trust-remote-code
```

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](/codes/modelzoo/qwen1.5_vllm/-/blob/master/benchmarks/benchmark_serving.py)

### OpenAI兼容服务

启动服务:

```bash
154
VLLM_USE_FLASH_ATTN_PA=1 vllm serve /your/model/path --enforce-eager --dtype float16 --trust-remote-code
zhuwenwen's avatar
zhuwenwen committed
155
156
157
158
159
160
161
162
163
164
165
166
```

这里sreve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。

### OpenAI Chat API和vllm结合使用

```bash
curl http://localhost:8000/v1/chat/completions \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{
    "model": "/your/model/path",
zhuwenwen's avatar
zhuwenwen committed
167
    "max_tokens": 128,
zhuwenwen's avatar
zhuwenwen committed
168
169
170
171
172
173
174
175
176
    "messages": [
      {
        "role": "user",
        "content": "介绍重庆美食"
      }
    ]
  }'
```

177
或者使用[examples/online_serving/openai_chat_completion_client.py](examples/online_serving/openai_chat_completion_client.py)
zhuwenwen's avatar
zhuwenwen committed
178
179
180
181
182
183
184
185
186
187
188
189
190
191

### **gradio和vllm结合使用**

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

```
192
python examples/online_serving/gradio_openai_chatbot_webserver.py --model "/your/model/path" --model-url http://localhost:8000/v1 --temp 0.8 --stop-token-ids ""
zhuwenwen's avatar
zhuwenwen committed
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
```

    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.*
```

    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
```

3.启动OpenAI兼容服务

```
212
VLLM_USE_FLASH_ATTN_PA=1 vllm serve /your/model/path --enforce-eager --dtype float16 --trust-remote-code --host "0.0.0.0"
zhuwenwen's avatar
zhuwenwen committed
213
214
215
216
217
```

4.启动gradio服务

```
218
python examples/online_serving/gradio_openai_chatbot_webserver.py --model "/your/model/path" --model-url http://localhost:8000/v1 --temp 0.8 --stop-token-ids "" --host "0.0.0.0" --port 8001
zhuwenwen's avatar
zhuwenwen committed
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
```

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。

## result

使用的加速卡:1张 DCU-K100_AI-64G

```
Prompt: 'What is deep learning?', Generated text: ' Deep learning is a subset of machine learning that involves the use of neural networks to model and solve complex problems. Neural networks are a network of interconnected nodes or " neurons" that are designed to recognize patterns in data, learn from examples, and make predictions or decisions.\nThe term "deep" in deep learning refers to the use of multiple layers or hidden layers in these neural networks. Each layer processes the input data in a different way, extracting increasingly abstract features as the data passes through.'
```

### 精度



## 应用场景

### 算法类别

对话问答

### 热点应用行业

金融,科研,教育

## 源码仓库及问题反馈

zhuwenwen's avatar
zhuwenwen committed
249
* [https://developer.hpccube.com/codes/modelzoo/qwen3_vllm](https://developer.hpccube.com/codes/modelzoo/qwen3_vllm)
zhuwenwen's avatar
zhuwenwen committed
250
251
252
253

## 参考资料

* [https://github.com/vllm-project/vllm](https://github.com/vllm-project/vllm)