搭建企业内部的大语言模型系统

大纲

• 开源大语言模型
• 大语言模型管理
• 私有大语言模型服务部署方案

开源大语言模型

担心安全与隐私？可私有部署的开源大模型

• 商业大模型，不支持私有部署
• • ChatGPT
  • Claude
  • Google Gemini
  • 百度问心一言
• 开源大模型，支持私有部署
• • Mistral
  • Meta Llama
  • ChatGLM
  • 阿里通义千问

常用开源大模型列表

开源大模型分支

大语言模型管理

大语言模型管理工具

• HuggingFace 全面的大语言模型管理平台
• Ollama 在本地管理大语言模型，下载速度超快
• llama.cpp 在本地和云端的各种硬件上以最少的设置和最先进的性能实现 LLM 推理
• GPT4All 一个免费使用、本地运行、具有隐私意识的聊天机器人。无需 GPU 或互联网

Ollama 速度最快的大语言模型管理工具

Ollama 的命令

ollama pull llama2
ollama list
ollama run llama2 "Summarize this file: $(cat README.md)"

ollama serve

curl http://localhost:11434/api/generate -d '{
  "model": "llama2",
  "prompt":"Why is the sky blue?"
}'
curl http://localhost:11434/api/chat -d '{
  "model": "mistral",
  "messages": [
    { "role": "user", "content": "why is the sky blue?" }
  ]
}'

大语言模型的前端

大语言模型的应用前端

• 开源平台 ollama-chatbot、PrivateGPT、gradio
• 开源服务 hugging face TGI、langchain-serve
• 开源框架 langchain llama-index

ollama chatbot

docker run -p 3000:3000 ghcr.io/ivanfioravanti/chatbot-ollama:main
## http://localhost:3000

ollama chatbot

PrivateGPT

PrivateGPT 提供了一个 API，其中包含构建私有的、上下文感知的 AI 应用程序所需的所有构建块。该 API 遵循并扩展了 OpenAI API 标准，支持普通响应和流响应。这意味着，如果您可以在您的工具之一中使用 OpenAI API，则可以使用您自己的 PrivateGPT API，无需更改代码，并且如果您在本地模式下运行 privateGPT，则免费。

PrivateGPT 架构

• FastAPI
• LLamaIndex
• 支持本地 LLM，比如 ChatGLM llama Mistral
• 支持远程 LLM，比如 OpenAI Claud
• 支持嵌入 embeddings，比如 ollama embeddings-huggingface
• 支持向量存储，比如 Qdrant, ChromaDB and Postgres

PrivateGPT 环境准备

git clone https://github.com/imartinez/privateGPT
cd privateGPT
#不支持3.11之前的版本
python3.11 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip poetry

#虽然官网只说了要安装少部分的依赖，但是那些依赖管理不是那么完善，容易有遗漏
#所以我们的策略就是全都要。
poetry install --extras "ui llms-llama-cpp llms-openai llms-openai-like llms-ollama llms-sagemaker llms-azopenai embeddings-ollama embeddings-huggingface embeddings-openai embeddings-sagemaker embeddings-azopenai vector-stores-qdrant vector-stores-chroma vector-stores-postgres storage-nodestore-postgres"

#或者用这个安装脚本
#poetry install --extras "$(sed -n '/tool.poetry.extras/,/^$/p'  pyproject.toml | awk -F= 'NR>1{print $1}' | xargs)"

ollama 部署方式

ollama pull mistral
ollama pull nomic-embed-text
ollama serve

#官方这个依赖不够，还需要额外安装torch，所以尽量采用上面提到的全部安装的策略
poetry install --extras "ui llms-ollama embeddings-ollama vector-stores-qdrant"
PGPT_PROFILES=ollama poetry run python -m private_gpt

setting-ollama.yaml

server:
  env_name: ${APP_ENV:ollama}

llm:
  mode: ollama
  max_new_tokens: 512
  context_window: 3900
  temperature: 0.1 #The temperature of the model. Increasing the temperature will make the model answer more creatively. A value of 0.1 would be more factual. (Default: 0.1)

embedding:
  mode: ollama

ollama:
  llm_model: mistral
  embedding_model: nomic-embed-text
  api_base: http://localhost:11434
  tfs_z: 1.0 ## Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting.
  top_k: 40 ## Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)
  top_p: 0.9 ## Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)
  repeat_last_n: 64 ## Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)
  repeat_penalty: 1.2 ## Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)

vectorstore:
  database: qdrant

qdrant:
  path: local_data/private_gpt/qdrant

启动

PGPT_PROFILES=ollama poetry run python -m private_gpt

poetry run python -m private_gpt
02:36:06.928 [INFO    ] private_gpt.settings.settings_loader - Starting application with profiles=['default', 'ollama']
02:36:46.567 [INFO    ] private_gpt.components.llm.llm_component - Initializing the LLM in mode=ollama
02:36:47.405 [INFO    ] private_gpt.components.embedding.embedding_component - Initializing the embedding model in mode=ollama
02:36:47.414 [INFO    ] llama_index.core.indices.loading - Loading all indices.
02:36:47.571 [INFO    ]         private_gpt.ui.ui - Mounting the gradio UI, at path=/
02:36:47.620 [INFO    ]             uvicorn.error - Started server process [72677]
02:36:47.620 [INFO    ]             uvicorn.error - Waiting for application startup.
02:36:47.620 [INFO    ]             uvicorn.error - Application startup complete.
02:36:47.620 [INFO    ]             uvicorn.error - Uvicorn running on http://0.0.0.0:8001 (Press CTRL+C to quit)

PrivateGPT UI

local 部署模式

#todo: 需要安装llama-cpp，每个平台的安装方式都不同，参考官方文档

poetry run python scripts/setup
PGPT_PROFILES=local poetry run python -m private_gpt

setting-local.yaml

server:
  env_name: ${APP_ENV:local}

llm:
  mode: llamacpp
  ## Should be matching the selected model
  max_new_tokens: 512
  context_window: 3900
  tokenizer: mistralai/Mistral-7B-Instruct-v0.2

llamacpp:
  prompt_style: "mistral"
  llm_hf_repo_id: TheBloke/Mistral-7B-Instruct-v0.2-GGUF
  llm_hf_model_file: mistral-7b-instruct-v0.2.Q4_K_M.gguf

embedding:
  mode: huggingface

huggingface:
  embedding_hf_model_name: BAAI/bge-small-en-v1.5

vectorstore:
  database: qdrant

qdrant:
  path: local_data/private_gpt/qdrant

非私有 OpenAI-powered 部署

poetry install --extras "ui llms-openai embeddings-openai vector-stores-qdrant"
PGPT_PROFILES=openai poetry run python -m private_gpt

setting-openai.yaml

server:
  env_name: ${APP_ENV:openai}

llm:
  mode: openai

embedding:
  mode: openai

openai:
  api_key: ${OPENAI_API_KEY:}
  model: gpt-3.5-turbo

openai 风格的 API 调用

• The API is built using FastAPI and follows OpenAI's API scheme.
• The RAG pipeline is based on LlamaIndex.

curl -X POST http://localhost:8000/v1/completions \
     -H "Content-Type: application/json" \
     -d '{
  "prompt": "string",
  "stream": true

}'