Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,493 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
pipeline_tag: text-generation
|
3 |
+
license: other
|
4 |
+
---
|
5 |
+
# XiXiLM-14b
|
6 |
+
|
7 |
+
<div align="center">
|
8 |
+
|
9 |
+
<img src="https://github.com/AI4Bread/GouMang/blob/main/assets/goumang_logoallnew.png?raw=true" width="600"/>
|
10 |
+
<div> </div>
|
11 |
+
<div align="center">
|
12 |
+
<!-- <b><font size="5">XiXiLM</font></b> -->
|
13 |
+
<sup>
|
14 |
+
<a href="http://www.ai4bread.com">
|
15 |
+
</a>
|
16 |
+
</sup>
|
17 |
+
<div> </div>
|
18 |
+
</div>
|
19 |
+
|
20 |
+
|
21 |
+
[💻Github Repo](https://github.com/AI4Bread/GouMang) • [🤔Reporting Issues](https://github.com/AI4Bread/GouMang/issues) • [📜Technical Report](https://github.com/AI4Bread)
|
22 |
+
|
23 |
+
</div>
|
24 |
+
|
25 |
+
<p align="center">
|
26 |
+
👋 join us on <a href="https://github.com/AI4Bread/GouMang" target="_blank">Github</a>
|
27 |
+
</p>
|
28 |
+
|
29 |
+
|
30 |
+
|
31 |
+
## Introduction
|
32 |
+
|
33 |
+
XiXiLM-14b(GouMang LLM-14b) has open-sourced a 14 billion parameter base model and a chat model tailored for agricultural scenarios. The model has the following characteristics:
|
34 |
+
|
35 |
+
1. **High Professionalism**: XiXiLM focuses on the agricultural field, providing professional and accurate answers especially in areas such as tuber crop cultivation, pest and disease control, and soil management.
|
36 |
+
|
37 |
+
2. **Academic Support**: The model is based on the latest agricultural research findings, capable of providing academic-level answers to help researchers and agricultural practitioners gain a deeper understanding of agricultural issues.
|
38 |
+
|
39 |
+
3. **Multilingual Support**: Supports both Chinese and English languages, making it convenient for users both domestically and internationally.
|
40 |
+
|
41 |
+
4. **Free Commercial Use**: The model weights are fully open, supporting not only academic research but also allowing **free** commercial usage. Users can use the model in commercial projects for free, lowering the usage threshold.
|
42 |
+
|
43 |
+
5. **Efficient Training**: Employs advanced training algorithms and techniques, enabling the model to respond quickly to user inquiries and provide efficient Q&A services.
|
44 |
+
|
45 |
+
6. **Continuous Optimization**: The model will be continuously optimized based on user feedback and the latest research findings, constantly improving the quality and coverage of its answers.
|
46 |
+
|
47 |
+
## XiXiLM-Qwen-14B
|
48 |
+
|
49 |
+
|
50 |
+
**Limitations:** Although we have made efforts to ensure the safety of the model during the training process and to
|
51 |
+
encourage the model to generate text that complies with ethical and legal requirements, the model may still produce unexpected
|
52 |
+
outputs due to its size and probabilistic generation paradigm. For example, the generated responses may contain biases, discrimination,
|
53 |
+
or other harmful content. Please do not propagate such content. We are not responsible for any consequences resulting from the
|
54 |
+
dissemination of harmful information.
|
55 |
+
|
56 |
+
### Import from Transformers
|
57 |
+
|
58 |
+
To load the XiXiLM model using Transformers, use the following code:
|
59 |
+
|
60 |
+
```python
|
61 |
+
import torch
|
62 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
63 |
+
tokenizer = AutoTokenizer.from_pretrained("AI4Bread/XiXi_Qwen_base_14b", trust_remote_code=True)
|
64 |
+
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and cause OOM Error.
|
65 |
+
model = AutoModelForCausalLM.from_pretrained("AI4Bread/XiXi_Qwen_base_14b", torch_dtype=torch.float16, trust_remote_code=True).cuda()
|
66 |
+
model = model.eval()
|
67 |
+
response, history = model.chat(tokenizer, "你好", history=[])
|
68 |
+
print(response)
|
69 |
+
# Hello! How can I help you today?
|
70 |
+
response, history = model.chat(tokenizer, "马铃薯育种有什么注意事项?需要注意什么呢?", history=history)
|
71 |
+
print(response)
|
72 |
+
```
|
73 |
+
|
74 |
+
The responses can be streamed using `stream_chat`:
|
75 |
+
|
76 |
+
```python
|
77 |
+
import torch
|
78 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
79 |
+
|
80 |
+
model_path = "AI4Bread/XiXi_Qwen_base_14b"
|
81 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16, trust_remote_code=True).cuda()
|
82 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
83 |
+
|
84 |
+
model = model.eval()
|
85 |
+
length = 0
|
86 |
+
for response, history in model.stream_chat(tokenizer, "Hello", history=[]):
|
87 |
+
print(response[length:], flush=True, end="")
|
88 |
+
length = len(response)
|
89 |
+
```
|
90 |
+
|
91 |
+
|
92 |
+
## Deployment
|
93 |
+
|
94 |
+
### LMDeploy
|
95 |
+
|
96 |
+
LMDeploy is a toolkit for compressing, deploying, and serving LLM, developed by the MMRazor and MMDeploy teams.
|
97 |
+
|
98 |
+
```bash
|
99 |
+
pip install lmdeploy
|
100 |
+
```
|
101 |
+
|
102 |
+
Or you can launch an OpenAI compatible server with the following command:
|
103 |
+
|
104 |
+
```bash
|
105 |
+
lmdeploy serve api_server internlm/internlm2-chat-7b --model-name internlm2-chat-7b --server-port 23333
|
106 |
+
```
|
107 |
+
|
108 |
+
Then you can send a chat request to the server:
|
109 |
+
|
110 |
+
```bash
|
111 |
+
curl http://localhost:23333/v1/chat/completions \
|
112 |
+
-H "Content-Type: application/json" \
|
113 |
+
-d '{
|
114 |
+
"model": "internlm2-chat-7b",
|
115 |
+
"messages": [
|
116 |
+
{"role": "system", "content": "你是一个专业的农业专家"},
|
117 |
+
{"role": "user", "content": "马铃薯种植的时候有哪些注意事项?"}
|
118 |
+
]
|
119 |
+
}'
|
120 |
+
```
|
121 |
+
|
122 |
+
The output be like:
|
123 |
+
|
124 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/658a3c4cbbb04840e3ce7e2c/NPdRr5Y5l5E0m0URCVZ1f.png)
|
125 |
+
|
126 |
+
|
127 |
+
|
128 |
+
Find more details in the [LMDeploy documentation](https://lmdeploy.readthedocs.io/en/latest/)
|
129 |
+
|
130 |
+
### vLLM
|
131 |
+
|
132 |
+
Launch OpenAI compatible server with `vLLM>=0.3.2`:
|
133 |
+
|
134 |
+
```bash
|
135 |
+
pip install vllm
|
136 |
+
```
|
137 |
+
|
138 |
+
```bash
|
139 |
+
python -m vllm.entrypoints.openai.api_server --model internlm/internlm2-chat-7b --served-model-name internlm2-chat-7b --trust-remote-code
|
140 |
+
```
|
141 |
+
|
142 |
+
Then you can send a chat request to the server:
|
143 |
+
|
144 |
+
```bash
|
145 |
+
curl http://localhost:8000/v1/chat/completions \
|
146 |
+
-H "Content-Type: application/json" \
|
147 |
+
-d '{
|
148 |
+
"model": "internlm2-chat-7b",
|
149 |
+
"messages": [
|
150 |
+
{"role": "system", "content": "You are a professional agriculture expert."},
|
151 |
+
{"role": "user", "content": "Introduce potato farming to me."}
|
152 |
+
]
|
153 |
+
}'
|
154 |
+
```
|
155 |
+
|
156 |
+
Find more details in the [vLLM documentation](https://docs.vllm.ai/en/latest/index.html)
|
157 |
+
|
158 |
+
## Used local trained model
|
159 |
+
|
160 |
+
### First: Convert lmdeploy TurboMind
|
161 |
+
|
162 |
+
Here, we will use our pre-trained model file and execute the conversion in the user's root directory, as shown below.
|
163 |
+
|
164 |
+
```bash
|
165 |
+
# Converting Model to TurboMind (FastTransformer Format)
|
166 |
+
lmdeploy convert internlm2-chat-7b /root/autodl-tmp/agri_intern/XiXiLM --tokenizer-path ./GouMang/tokenizer.json
|
167 |
+
```
|
168 |
+
|
169 |
+
After execution, a workspace folder will be generated in the current directory.
|
170 |
+
This folder contains the necessary files for TurboMind and Triton "Model Inference." as shown below:
|
171 |
+
|
172 |
+
|
173 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/658a3c4cbbb04840e3ce7e2c/CqdwhshIL8xxjog_WD_St.png)
|
174 |
+
|
175 |
+
|
176 |
+
### Second: Chat Locally
|
177 |
+
|
178 |
+
```bash
|
179 |
+
lmdeploy chat turbomind ./workspace
|
180 |
+
```
|
181 |
+
|
182 |
+
### Third(Optional): TurboMind Inference + API Service
|
183 |
+
|
184 |
+
In the previous section, we tried starting the Client directly using the command line. Now, we will attempt to use lmdeploy for service deployment.
|
185 |
+
|
186 |
+
The "Model Inference/Service" currently offers two service deployment methods: TurboMind and TritonServer. In this case, the Server is either TurboMind or TritonServer, and the API Server can provide external API services. We recommend using TurboMind.
|
187 |
+
|
188 |
+
First, start the service with the following command:
|
189 |
+
|
190 |
+
|
191 |
+
```bash
|
192 |
+
# ApiServer+Turbomind api_server => AsyncEngine => TurboMind
|
193 |
+
lmdeploy serve api_server ./workspace \
|
194 |
+
--server-name 0.0.0.0 \
|
195 |
+
--server-port 23333 \
|
196 |
+
--tp 1
|
197 |
+
```
|
198 |
+
|
199 |
+
In the above parameters, `server_name` and `server_port` indicate the service address and port, respectively. The `tp` parameter, as mentioned earlier, stands for Tensor Parallelism.
|
200 |
+
|
201 |
+
After this, users can start the Web Service as described in [TurboMind Service as the Backend](#--turbomind-service-as-the-backend).
|
202 |
+
|
203 |
+
## Web Service Startup Method 1:
|
204 |
+
|
205 |
+
### Starting the Service with Gradio
|
206 |
+
|
207 |
+
This section demonstrates using Gradio as a front-end demo.
|
208 |
+
|
209 |
+
> Since Gradio requires local access to display the interface,
|
210 |
+
> you also need to forward the data to your local machine via SSH. The command is as follows:
|
211 |
+
>
|
212 |
+
> ssh -CNg -L 6006:127.0.0.1:6006 root@ssh.intern-ai.org.cn -p <your ssh port>
|
213 |
+
|
214 |
+
#### --TurboMind Service as the Backend
|
215 |
+
|
216 |
+
The API Server is started the same way as in the previous section. Here, we directly start Gradio as the front-end.
|
217 |
+
|
218 |
+
```bash
|
219 |
+
# Gradio+ApiServer. The Server must be started first, and Gradio acts as the Client
|
220 |
+
lmdeploy serve gradio http://0.0.0.0:23333 --server-port 6006
|
221 |
+
```
|
222 |
+
|
223 |
+
#### --Other way(Recommended!!!)
|
224 |
+
|
225 |
+
Of course, Gradio can also connect directly with TurboMind, as shown below
|
226 |
+
|
227 |
+
```bash
|
228 |
+
# Gradio+Turbomind(local)
|
229 |
+
lmdeploy serve gradio ./workspace
|
230 |
+
```
|
231 |
+
|
232 |
+
You can start Gradio directly. In this case, there is no API Server, and TurboMind communicates directly with Gradio.
|
233 |
+
|
234 |
+
## Web Service Startup Method 2:
|
235 |
+
|
236 |
+
### Starting the Service with Streamlit
|
237 |
+
|
238 |
+
```bash
|
239 |
+
pip install streamlit==1.24.0
|
240 |
+
```
|
241 |
+
|
242 |
+
Download the [GouMang](https://huggingface.co/AI4Bread/GouMang) project model (please Star if you like it)
|
243 |
+
|
244 |
+
```bash
|
245 |
+
git clone https://github.com/AI4Bread/GouMang.git
|
246 |
+
cd GouMang
|
247 |
+
```
|
248 |
+
|
249 |
+
|
250 |
+
Replace the model path in `web_demo.py` with the path where the downloaded parameters of `GouMang` are stored
|
251 |
+
|
252 |
+
Run the `web_demo.py` file in the directory, and after entering the following command, [**check this tutorial 5.2 for local port configuration**](https://github.com/InternLM/tutorial/blob/main/helloworld/hello_world.md#52-%E9%85%8D%E7%BD%AE%E6%9C%AC%E5%9C%B0%E7%AB%AF%E5%8F%A3),to map the port to your local machine. Enter `http://127.0.0.1:6006` in your local browser.
|
253 |
+
|
254 |
+
```
|
255 |
+
streamlit run web_demo.py --server.address 127.0.0.1 --server.port 6006
|
256 |
+
```
|
257 |
+
|
258 |
+
Note: The model will load only after you open the `http://127.0.0.1:6006` page in your browser.
|
259 |
+
Once the model is loaded, you can start conversing with GouMang like this.
|
260 |
+
|
261 |
+
|
262 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/658a3c4cbbb04840e3ce7e2c/VcuSpAKrRGY1HP1mwLGI6.png)
|
263 |
+
|
264 |
+
|
265 |
+
## Open Source License
|
266 |
+
|
267 |
+
The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow **free** commercial usage. To apply for a commercial license, please fill in the <a href="https://wj.qq.com/s2/14897739/e871/" target="_blank">申���表(中文)</a>. For other questions or collaborations, please contact <laiyifu@xjtu.edu.cn>.
|
268 |
+
|
269 |
+
## Citation
|
270 |
+
|
271 |
+
|
272 |
+
|
273 |
+
## 简介
|
274 |
+
|
275 |
+
XiXiLM-14b ,即西西大模型(又名:句芒大模型),开源了面向农业问答的大模型。模型具有以下特点:
|
276 |
+
|
277 |
+
1. **专业性强**:XiXiLM 专注于农业领域,特别是薯类作物的种植、病虫害防治、土壤管理等方面,提供专业、精准的解答。
|
278 |
+
|
279 |
+
2. **学术化支持**:模型基于最新的农业研究成果,能够提供学术化的回答,帮助研究人员和农业从业者深入理解农业问题。
|
280 |
+
|
281 |
+
3. **多语言支持**:支持中文和英文两种语言,方便国内外用户使用。
|
282 |
+
|
283 |
+
4. **免费商业使用**:模型权重完全开放,不仅支持学术研究,还允许**申请**商业使用。用户可以在商业项目中免费使用该模型,降低了使用门槛。
|
284 |
+
|
285 |
+
5. **高效训练**:采用先进的训练算法和技术,使得模型能够快速响应用户提问,提供高效的问答服务。
|
286 |
+
|
287 |
+
6. **持续优化**:模型会根据用户反馈和最新研究成果进行持续优化,不断提升问答质量和覆盖面。
|
288 |
+
|
289 |
+
|
290 |
+
## XiXiLM-14B
|
291 |
+
|
292 |
+
|
293 |
+
**局限性:** 尽管在训练过程中我们非常注重模型的安全性,尽力促使模型输出符合伦理和法律要求的文本,但受限于模型大小以及概率生成范式,模型可能会产生各种不符合预期的输出,例如回复内容包含偏见、歧视等有害内容,请勿传播这些内容。由于传播不良信息导致的任何后果,本项目不承担责任。
|
294 |
+
|
295 |
+
### 通过 Transformers 加载
|
296 |
+
|
297 |
+
通过以下的代码加载 XiXiLM-14b Chat 模型
|
298 |
+
|
299 |
+
```python
|
300 |
+
import torch
|
301 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
302 |
+
tokenizer = AutoTokenizer.from_pretrained("AI4Bread/XiXiLM_14b", trust_remote_code=True)
|
303 |
+
# Set `torch_dtype=torch.float16` to load model in float16, otherwise it will be loaded as float32 and cause OOM Error.
|
304 |
+
model = AutoModelForCausalLM.from_pretrained("AI4Bread/XiXiLM_14b", torch_dtype=torch.float16, trust_remote_code=True).cuda()
|
305 |
+
model = model.eval()
|
306 |
+
response, history = model.chat(tokenizer, "你好", history=[])
|
307 |
+
print(response)
|
308 |
+
# Hello! How can I help you today?
|
309 |
+
response, history = model.chat(tokenizer, "马铃薯育种有什么注意事项?需要注意什么呢?", history=history)
|
310 |
+
print(response)
|
311 |
+
```
|
312 |
+
|
313 |
+
如果想进行流式生成,则可以使用 `stream_chat` 接口:
|
314 |
+
|
315 |
+
```python
|
316 |
+
import torch
|
317 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
318 |
+
|
319 |
+
model_path = "AI4Bread/XiXi_Qwen_base_14b"
|
320 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16, trust_remote_code=True).cuda()
|
321 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
322 |
+
|
323 |
+
model = model.eval()
|
324 |
+
length = 0
|
325 |
+
for response, history in model.stream_chat(tokenizer, "马铃薯育种有什么注意事项?需要注意什么呢?", history=[]):
|
326 |
+
print(response[length:], flush=True, end="")
|
327 |
+
length = len(response)
|
328 |
+
```
|
329 |
+
|
330 |
+
## 部署
|
331 |
+
|
332 |
+
### LMDeploy
|
333 |
+
|
334 |
+
LMDeploy 由 MMDeploy 和 MMRazor 团队联合开发,是涵盖了 LLM 任务的全套轻量化、部署和服务解决方案。
|
335 |
+
|
336 |
+
```bash
|
337 |
+
pip install lmdeploy
|
338 |
+
```
|
339 |
+
|
340 |
+
你可以使用以下命令启动兼容 OpenAI API 的服务:
|
341 |
+
|
342 |
+
```bash
|
343 |
+
lmdeploy serve api_server internlm/internlm2-chat-7b --server-port 23333
|
344 |
+
```
|
345 |
+
|
346 |
+
然后你可以向服务端发起一个聊天请求:
|
347 |
+
|
348 |
+
```bash
|
349 |
+
curl http://localhost:23333/v1/chat/completions \
|
350 |
+
-H "Content-Type: application/json" \
|
351 |
+
-d '{
|
352 |
+
"model": "internlm2-chat-7b",
|
353 |
+
"messages": [
|
354 |
+
{"role": "system", "content": "你是一个专业的农业专家"},
|
355 |
+
{"role": "user", "content": "马铃薯种植的时候有哪些注意事项?"}
|
356 |
+
]
|
357 |
+
}'
|
358 |
+
```
|
359 |
+
|
360 |
+
更多信息请查看 [LMDeploy 文档](https://lmdeploy.readthedocs.io/en/latest/)
|
361 |
+
|
362 |
+
### vLLM
|
363 |
+
|
364 |
+
使用`vLLM>=0.3.2`启动兼容 OpenAI API 的服务:
|
365 |
+
|
366 |
+
```bash
|
367 |
+
pip install vllm
|
368 |
+
```
|
369 |
+
|
370 |
+
```bash
|
371 |
+
python -m vllm.entrypoints.openai.api_server --model internlm/internlm2-chat-7b --trust-remote-code
|
372 |
+
```
|
373 |
+
|
374 |
+
然后你可以向服务端发起一个聊天请求:
|
375 |
+
|
376 |
+
```bash
|
377 |
+
curl http://localhost:8000/v1/chat/completions \
|
378 |
+
-H "Content-Type: application/json" \
|
379 |
+
-d '{
|
380 |
+
"model": "internlm2-chat-7b",
|
381 |
+
"messages": [
|
382 |
+
{"role": "system", "content": "你是一个专业的农业专家."},
|
383 |
+
{"role": "user", "content": "请给我介绍一下马铃薯育种."}
|
384 |
+
]
|
385 |
+
}'
|
386 |
+
```
|
387 |
+
|
388 |
+
更多信息请查看 [vLLM 文档](https://docs.vllm.ai/en/latest/index.html)
|
389 |
+
|
390 |
+
## 使用本地训练模型
|
391 |
+
|
392 |
+
### 第一步:转换为 lmdeploy TurboMind 格式
|
393 |
+
|
394 |
+
这里,我们将使用预训练的模型文件,并在用户的根目录下执行转换,如下所示。
|
395 |
+
|
396 |
+
```bash
|
397 |
+
# 将模型转换为 TurboMind (FastTransformer 格式)
|
398 |
+
lmdeploy convert internlm2-chat-7b /root/autodl-tmp/agri_intern/XiXiLM --tokenizer-path ./GouMang/tokenizer.json
|
399 |
+
```
|
400 |
+
|
401 |
+
执行完毕后,当前目录下将生成一个 workspace 文件夹。
|
402 |
+
这个文件夹包含 TurboMind 和 Triton “模型推���”所需的文件,如下所示:
|
403 |
+
|
404 |
+
|
405 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/658a3c4cbbb04840e3ce7e2c/CqdwhshIL8xxjog_WD_St.png)
|
406 |
+
|
407 |
+
|
408 |
+
### 第二步:本地聊天
|
409 |
+
|
410 |
+
```bash
|
411 |
+
lmdeploy chat turbomind ./workspace
|
412 |
+
```
|
413 |
+
|
414 |
+
### 第三步(可选):TurboMind 推理 + API 服务
|
415 |
+
|
416 |
+
在前一部分中,我们尝试通过命令行直接启动客户端。现在,我们将尝试使用 lmdeploy 进行服务部署。
|
417 |
+
|
418 |
+
“模型推理/服务”目前提供两种服务部署方式:TurboMind 和 TritonServer。在这种情况下,服务器可以是 TurboMind 或 TritonServer,而 API 服务器可以提供外部 API 服务。我们推荐使用 TurboMind。
|
419 |
+
|
420 |
+
首先,使用以下命令启动服务:
|
421 |
+
|
422 |
+
```bash
|
423 |
+
# ApiServer+Turbomind api_server => AsyncEngine => TurboMind
|
424 |
+
lmdeploy serve api_server ./workspace \
|
425 |
+
--server-name 0.0.0.0 \
|
426 |
+
--server-port 23333 \
|
427 |
+
--tp 1
|
428 |
+
```
|
429 |
+
|
430 |
+
在上述参数中,server_name 和 server_port 分别表示服务地址和端口。tp 参数如前所述代表 Tensor 并行性。
|
431 |
+
|
432 |
+
之后,用户可以按照[TurboMind Service as the Backend](#--turbomind-service-as-the-backend) 中描述的启动 Web 服务。
|
433 |
+
|
434 |
+
|
435 |
+
|
436 |
+
## 网页服务启动方式1:
|
437 |
+
|
438 |
+
### Gradio 方式启动服务
|
439 |
+
|
440 |
+
这一部分主要是将 Gradio 作为前端 Demo 演示。在上一节的基础上,我们不执行后面的 `api_client` 或 `triton_client`,而是执行 `gradio`。
|
441 |
+
请参考[LMDeploy](#lmdeploy)部分获取详细信息。
|
442 |
+
|
443 |
+
> 由于 Gradio 需要本地访问展示界面,因此也需要通过 ssh 将数据转发到本地。命令如下:
|
444 |
+
>
|
445 |
+
> ssh -CNg -L 6006:127.0.0.1:6006 root@ssh.intern-ai.org.cn -p <你的 ssh 端口号>
|
446 |
+
|
447 |
+
#### --TurboMind 服务作为后端
|
448 |
+
|
449 |
+
直接启动作为前端的 Gradio。
|
450 |
+
|
451 |
+
```bash
|
452 |
+
# Gradio+ApiServer。必须先开启 Server,此时 Gradio 为 Client
|
453 |
+
lmdeploy serve gradio http://0.0.0.0:23333 --server-port 6006
|
454 |
+
```
|
455 |
+
|
456 |
+
#### --其他方式(推荐!!!)
|
457 |
+
|
458 |
+
当然,Gradio 也可以直接和 TurboMind 连接,如下所示。
|
459 |
+
|
460 |
+
```bash
|
461 |
+
# Gradio+Turbomind(local)
|
462 |
+
lmdeploy serve gradio ./workspace
|
463 |
+
```
|
464 |
+
|
465 |
+
可以直接启动 Gradio,此时没有 API Server,TurboMind 直接与 Gradio 通信。
|
466 |
+
|
467 |
+
## 网页服务启动方式2:
|
468 |
+
|
469 |
+
### Streamlit 方式启动服务:
|
470 |
+
|
471 |
+
下载 [GouMang](https://huggingface.co/AI4Bread/GouMang) 项目模型(如果喜欢请给个 Star)
|
472 |
+
|
473 |
+
```bash
|
474 |
+
git clone https://github.com/AI4Bread/GouMang.git
|
475 |
+
cd GouMang
|
476 |
+
```
|
477 |
+
|
478 |
+
将 `web_demo.py` 中的模型路径替换为下载的 `GouMang` 参数存储路径
|
479 |
+
|
480 |
+
在目录中运行 `web_demo.py` 文件,并在输入以下命令后,[**查看本教程 5.2 以配置本地端口**](https://github.com/InternLM/tutorial/blob/main/helloworld/hello_world.md#52-%E9%85%8D%E7%BD%AE%E6%9C%AC%E5%9C%B0%E7%AB%AF%E5%8F%A3),将端口映射到本地。在本地浏览器中输入 `http://127.0.0.1:6006`。
|
481 |
+
|
482 |
+
```
|
483 |
+
streamlit run /root/personal_assistant/code/InternLM/web_demo.py --server.address 127.0.0.1 --server.port 6006
|
484 |
+
```
|
485 |
+
|
486 |
+
注意:只有在浏览器中打开 `http://127.0.0.1:6006` 页面后,模型才会加载。
|
487 |
+
模型加载完成后,您就可以开始与 西西(句芒) 进行对话了。
|
488 |
+
|
489 |
+
## 开源许可证
|
490 |
+
|
491 |
+
本仓库的代码依照 Apache-2.0 协议开源。模型权重对学术研究完全开放,也可申请免费的商业使用授权(<a href="https://wj.qq.com/s2/14897739/e871/" target="_blank">申请表(中文)</a>)。其他问题与合作请联系 <laiyifu@xjtu.edu.cn>。
|
492 |
+
|
493 |
+
## 引用
|