"vscode:/vscode.git/clone" did not exist on "69301eada9a186153b47f927d14f9fad3cc56005"
README.md 3.49 KB
Newer Older
raojy's avatar
first  
raojy committed
1
2
3
# Qwen3.6
## 论文
[Qwen3.6](https://qwen.ai/blog?id=qwen3.6)
raojy's avatar
raojy committed
4

raojy's avatar
first  
raojy committed
5
## 模型简介
raojy's avatar
raojy committed
6
Qwen3.6是一款采用混合专家 (MoE) 架构并包含视觉编码器的多模态因果语言模型。此版本重点优化了智能体编程 (Agentic Coding) 的准确率,并引入了思维保留 (Thinking Preservation) 机制,特别适合长上下文和复杂的代码仓库开发任务。
raojy's avatar
first  
raojy committed
7
8
9
10
11
12
13
14
15
16

<div align=center>
    <img src="./doc/1.png"/>
</div>

## 环境依赖
| 软件 |                    版本                     |
| :------: |:-----------------------------------------:|
| DTK |                   26.04                   |
| python |                  3.10.12                  |
raojy's avatar
raojy committed
17
18
19
20
| transformers |            5.5.0               |
| vllm |      0.18.1+das.fa71803.dtk2604     |
| triton |      3.6.0+gitc73250c4.staging      |
| torch |   2.10.0+das.opt1.dtk2604.20260325.g6b060a   |
raojy's avatar
first  
raojy committed
21

raojy's avatar
raojy committed
22
当前推荐使用镜像: harbor.sourcefind.cn:5443/dcu/admin/base/custom:vllm018-ubuntu22.04-dtk26.04-qwen3.6-20260423
raojy's avatar
first  
raojy committed
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39

- 挂载地址`-v` 根据实际模型情况修改
```bash
docker run -it \
    --shm-size 200g \
    --network=host \
    --name qwen3.6 \
    --privileged \
    --device=/dev/kfd \
    --device=/dev/dri \
    --device=/dev/mkfd \
    --group-add video \
    --cap-add=SYS_PTRACE \
    --security-opt seccomp=unconfined \
    -u root \
    -v /opt/hyhal/:/opt/hyhal/:ro \
    -v /path/your_code_data/:/path/your_code_data/ \
raojy's avatar
raojy committed
40
    harbor.sourcefind.cn:5443/dcu/admin/base/custom:vllm018-ubuntu22.04-dtk26.04-qwen3.6-20260423 bash
raojy's avatar
first  
raojy committed
41
42
43
44
```
更多镜像可前往[光源](https://sourcefind.cn/#/service-list)下载使用。

关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.sourcefind.cn/tool/)开发者社区下载安装。
chenych's avatar
chenych committed
45

raojy's avatar
raojy committed
46
47
## 预训练权重
|  模型名称  | 权重大小 | 数据类型 |支持的DCU型号  | 最低卡数需求 |         下载地址          |
raojy's avatar
raojy committed
48
|:------:|:----:|:----:|:----------:|:------:|:---------------------:|
raojy's avatar
raojy committed
49
| Qwen3.6-35B-A3B | 35B | BF16 | BW1000 |   2   | [Hugging Face](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) |
raojy's avatar
raojy committed
50
| Qwen3.6-27B | 27B | BF16 | BW1000 |   1   | [Hugging Face](https://huggingface.co/Qwen/Qwen3.6-27B) |
chenych's avatar
chenych committed
51

raojy's avatar
first  
raojy committed
52
53
54
55
56
57
58
59
60
61
62
## 数据集
暂无

## 训练
暂无

## 推理
### vllm
#### 单机推理
```bash
## serve启动
raojy's avatar
raojy committed
63
vllm serve Qwen/Qwen3.6-27B \
raojy's avatar
first  
raojy committed
64
65
66
    --port 8001 \
    --trust-remote-code \
    --dtype bfloat16 \
raojy's avatar
raojy committed
67
    --tensor-parallel-size 2 \
raojy's avatar
raojy committed
68
69
70
    --gpu-memory-utilization 0.925 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder
raojy's avatar
first  
raojy committed
71
72

## client访问
raojy's avatar
raojy committed
73
74
75
76
curl http://localhost:8001/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3.6-27B",
raojy's avatar
first  
raojy committed
77
    "messages": [
raojy's avatar
raojy committed
78
      {"role": "user", "content": "帮我查询上海的天气"}
raojy's avatar
first  
raojy committed
79
    ],
raojy's avatar
raojy committed
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_weather",
          "description": "查询城市天气",
          "parameters": {
            "type": "object",
            "properties": {
              "city": {"type": "string"}
            },
            "required": ["city"]
          }
        }
      }
    ]
raojy's avatar
first  
raojy committed
96
  }'
raojy's avatar
raojy committed
97

raojy's avatar
first  
raojy committed
98
99
100
101
```

## 效果展示
<div align=center>
raojy's avatar
raojy committed
102
    <img src="./doc/4.png"/>
raojy's avatar
first  
raojy committed
103
104
105
</div>

### 精度
chenych's avatar
chenych committed
106
107
108
109
- 推理框架:vllm
- 测试数据:humaneval、gsm8k
- 使用的加速卡:bw1000

raojy's avatar
raojy committed
110
| model name| humaneval | gsm8k |
raojy's avatar
raojy committed
111
112
| :------: | :------: | :------: |
| Qwen3.6-27B | 0.8293 | 0.98 |
raojy's avatar
raojy committed
113
114
| Qwen3.6-35B-A3B | 0.878 | 0.976 |

raojy's avatar
first  
raojy committed
115
## 源码仓库及问题反馈
raojy's avatar
raojy committed
116
- https://developer.sourcefind.cn/codes/modelzoo/qwen3.6
raojy's avatar
first  
raojy committed
117
118
119

## 参考资料
- https://github.com/QwenLM/Qwen3.6