LLM之基于llama-index部署当地embedding与GLM-4模子并开端搭建RAG（其他大 ...

前言
日常没空，留着以后写

llama-index简介

官网：https://docs.llamaindex.ai/en/stable/
简介也没空，以后再写
注：先阐明，随着官方的变动，代码也大概变动，各人运行不起来，可以进官网查查资料
加载当地embedding模子

假如没有找到 llama_index.embeddings.huggingface
那么：pip install llama_index-embeddings-huggingface
还不可进入官网，输入huggingface举行搜索

1. from llama_index.embeddings.huggingface import HuggingFaceEmbedding
2. from llama_index.core import Settings
4. Settings.embed_model = HuggingFaceEmbedding(
5.     model_name=f"{embed_model_path}",device='cuda'
7. )

复制代码

 加载当地LLM模子

还是那句话，假如以下代码不可，进官网搜索Custom LLM Model

1. from llama_index.core.llms import (
2.     CustomLLM,
3.     CompletionResponse,
4.     CompletionResponseGen,
5.     LLMMetadata,
6. )
7. from llama_index.core.llms.callbacks import llm_completion_callback
8. from transformers import AutoTokenizer, AutoModelForCausalLM
10. class GLMCustomLLM(CustomLLM):
11.     context_window: int = 8192  # 上下文窗口大小
12.     num_output: int = 8000  # 输出的token数量
13.     model_name: str = "glm-4-9b-chat"  # 模型名称
14.     tokenizer: object = None  # 分词器
15.     model: object = None  # 模型
16.     dummy_response: str = "My response"
18.     def __init__(self, pretrained_model_name_or_path):
19.         super().__init__()
21.         # GPU方式加载模型
22.         self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, device_map="cuda", trust_remote_code=True)
23.         self.model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, device_map="cuda", trust_remote_code=True).eval()
25.         # CPU方式加载模型
26.         # self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, device_map="cpu", trust_remote_code=True)
27.         # self.model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, device_map="cpu", trust_remote_code=True)
28.         self.model = self.model.float()
30.     @property
31.     def metadata(self) -> LLMMetadata:
32.         """Get LLM metadata."""
33.         # 得到LLM的元数据
34.         return LLMMetadata(
35.             context_window=self.context_window,
36.             num_output=self.num_output,
37.             model_name=self.model_name,
38.         )
40.     # @llm_completion_callback()
41.     # def complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
42.     #     return CompletionResponse(text=self.dummy_response)
43.     #
44.     # @llm_completion_callback()
45.     # def stream_complete(
46.     #     self, prompt: str, **kwargs: Any
47.     # ) -> CompletionResponseGen:
48.     #     response = ""
49.     #     for token in self.dummy_response:
50.     #         response += token
51.     #         yield CompletionResponse(text=response, delta=token)
53.     @llm_completion_callback()  # 回调函数
54.     def complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
55.         # 完成函数
56.         print("完成函数")
58.         inputs = self.tokenizer.encode(prompt, return_tensors='pt').cuda()  # GPU方式
59.         # inputs = self.tokenizer.encode(prompt, return_tensors='pt')  # CPU方式
60.         outputs = self.model.generate(inputs, max_length=self.num_output)
61.         response = self.tokenizer.decode(outputs[0])
62.         return CompletionResponse(text=response)
64.     @llm_completion_callback()
65.     def stream_complete(
66.         self, prompt: str, **kwargs: Any
67.     ) -> CompletionResponseGen:
68.         # 流式完成函数
69.         print("流式完成函数")
71.         inputs = self.tokenizer.encode(prompt, return_tensors='pt').cuda()  # GPU方式
72.         # inputs = self.tokenizer.encode(prompt, return_tensors='pt')  # CPU方式
73.         outputs = self.model.generate(inputs, max_length=self.num_output)
74.         response = self.tokenizer.decode(outputs[0])
75.         for token in response:
76.             yield CompletionResponse(text=token, delta=token)

复制代码

基于当地模子搭建简易RAG

1. from typing import Any
3. from llama_index.core.llms import (
4.     CustomLLM,
5.     CompletionResponse,
6.     CompletionResponseGen,
7.     LLMMetadata,
8. )
9. from llama_index.core.llms.callbacks import llm_completion_callback
10. from transformers import AutoTokenizer, AutoModelForCausalLM
11. from llama_index.core import Settings,VectorStoreIndex,SimpleDirectoryReader
12. from llama_index.embeddings.huggingface import HuggingFaceEmbedding
15. class GLMCustomLLM(CustomLLM):
16.     context_window: int = 8192  # 上下文窗口大小
17.     num_output: int = 8000  # 输出的token数量
18.     model_name: str = "glm-4-9b-chat"  # 模型名称
19.     tokenizer: object = None  # 分词器
20.     model: object = None  # 模型
21.     dummy_response: str = "My response"
23.     def __init__(self, pretrained_model_name_or_path):
24.         super().__init__()
26.         # GPU方式加载模型
27.         self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, device_map="cuda", trust_remote_code=True)
28.         self.model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, device_map="cuda", trust_remote_code=True).eval()
30.         # CPU方式加载模型
31.         # self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, device_map="cpu", trust_remote_code=True)
32.         # self.model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, device_map="cpu", trust_remote_code=True)
33.         self.model = self.model.float()
35.     @property
36.     def metadata(self) -> LLMMetadata:
37.         """Get LLM metadata."""
38.         # 得到LLM的元数据
39.         return LLMMetadata(
40.             context_window=self.context_window,
41.             num_output=self.num_output,
42.             model_name=self.model_name,
43.         )
45.     # @llm_completion_callback()
46.     # def complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
47.     #     return CompletionResponse(text=self.dummy_response)
48.     #
49.     # @llm_completion_callback()
50.     # def stream_complete(
51.     #     self, prompt: str, **kwargs: Any
52.     # ) -> CompletionResponseGen:
53.     #     response = ""
54.     #     for token in self.dummy_response:
55.     #         response += token
56.     #         yield CompletionResponse(text=response, delta=token)
58.     @llm_completion_callback()  # 回调函数
59.     def complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
60.         # 完成函数
61.         print("完成函数")
63.         inputs = self.tokenizer.encode(prompt, return_tensors='pt').cuda()  # GPU方式
64.         # inputs = self.tokenizer.encode(prompt, return_tensors='pt')  # CPU方式
65.         outputs = self.model.generate(inputs, max_length=self.num_output)
66.         response = self.tokenizer.decode(outputs[0])
67.         return CompletionResponse(text=response)
69.     @llm_completion_callback()
70.     def stream_complete(
71.         self, prompt: str, **kwargs: Any
72.     ) -> CompletionResponseGen:
73.         # 流式完成函数
74.         print("流式完成函数")
76.         inputs = self.tokenizer.encode(prompt, return_tensors='pt').cuda()  # GPU方式
77.         # inputs = self.tokenizer.encode(prompt, return_tensors='pt')  # CPU方式
78.         outputs = self.model.generate(inputs, max_length=self.num_output)
79.         response = self.tokenizer.decode(outputs[0])
80.         for token in response:
81.             yield CompletionResponse(text=token, delta=token)
84. if __name__ == "__main__":
87.     # 定义你的LLM
88.     pretrained_model_name_or_path = r'/home/nlp/model/LLM/THUDM/glm-4-9b-chat'
89.     embed_model_path = '/home/nlp/model/Embedding/BAAI/bge-m3'
91.     Settings.embed_model = HuggingFaceEmbedding(
92.         model_name=f"{embed_model_path}",device='cuda'
94.     )
96.     Settings.llm = GLMCustomLLM(pretrained_model_name_or_path)
98.     documents = SimpleDirectoryReader(input_dir="home/xxxx/input").load_data()
99.     index = VectorStoreIndex.from_documents(
100.         documents,
101.     )
104.     # 查询和打印结果
105.     query_engine = index.as_query_engine()
106.     response = query_engine.query("萧炎的表妹是谁?")
108.     print(response)

复制代码

ollama 

1. from llama_index.core import VectorStoreIndex, SimpleDirectoryReader, Settings
2. from llama_index.embeddings.huggingface import HuggingFaceEmbedding
3. from llama_index.llms.ollama import Ollama
5. documents = SimpleDirectoryReader("data").load_data()
7. # bge-base embedding model
8. Settings.embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-base-en-v1.5")
10. # ollama
11. Settings.llm = Ollama(model="llama3", request_timeout=360.0)
13. index = VectorStoreIndex.from_documents(
14.     documents,
15. )

复制代码

欢迎各人点赞或收藏
各人的点赞或收藏可以鼓励作者加快更新哟~
作者也新建了一个学习交换群，欢迎各人加入一起学习成长
通过链接获取二维码
链接：https://pan.baidu.com/s/1ZZZZ-ANJvMFHlanl-tbtew?pwd=9el1 

参加链接：
LlamaIndex中的CustomLLM（当地加载模子）
llamaIndex 基于GPU加载当地embedding模子
 
官网文档
官网_starter_example_loca
官网_usage_custom

免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。