Ubuntu20.04系统安装,使用系统盘安装
1、系统安装
Ubuntu20.04系统安装,使用系统盘安装
查看ubuntu系统版本
lsb_release -a
:显示发行版名称、版本号及代号
(base) root@ai-System-Product-Name:/media/ai/wh/clash-for-linux-master# lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description: Ubuntu 20.04.6 LTS
Release: 20.04
Codename: focal
查看Ubuntu系统内核版本
uname -r
:直接显示内核版本号
(base) root@ai-System-Product-Name:/media/ai/wh/clash-for-linux-master# uname -r
5.15.0-139-generic
2、NVIDIA驱动安装
Ubuntu操作系统安装Nvidia GPU驱动-物理机-用户指南-实例-GPU驱动安装说明 - 天翼云
Ubuntu操作系统安装Nvidia GPU驱动
验证驱动是否正常安装
nvidia-smi
安装cuna工具包(一般情况选择带有系统版本的CUDA就能直接安装完)
CUDA Toolkit Archive | NVIDIA Developer
检查 CUDA Toolkit 是否安装成功
nvcc --version
操作系统及软件相关 | 安装 | 检查命令 |
系统驱动 | uname -r | |
C/C++编译运行环境 | 检查命令 gcc --version 、 g++ --version 、 cmake --version | |
Pytorch | Python -c "import torch" | |
Langchain | Python -c "import langchain" | |
Stable - diffusion | 通过查看代码仓库有无文件判断 | |
Lora微调框架 | 通过查看代码仓库有无文件判断 | |
Dockers容器组件 | 检查命令 sudo systemctl status docker 、 Docker --version | |
Python | python --version |
2.1 Pytorch安装
PyTorch官网下载链接:Start Locally | PyTorch
离线“:https://download.pytorch.org/whl/torch/
首先更新系统并安装必要的工具和依赖:
sudo apt update && sudo apt upgrade -y
sudo apt install python3-pip python3-dev build-essential libssl-dev libffi-dev -y
确保pip是最新版本:
pip3 install --upgrade pip
直接安装
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
2.2 安装Miniconda的步骤
Miniconda - Anaconda
下载Miniconda安装脚本
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh
运行安装脚本
bash Miniconda3-latest-Linux-x86_64.sh
按照提示进行操作:
-
按Enter查看许可协议
-
输入
yes
接受许可条款 -
选择安装位置(默认在
~/miniconda3
) -
是否初始化Miniconda(建议选择
yes
)
激活conda
source ~/.bashrc
验证安装
conda –version
配置环境(可选)
conda config --set auto_activate_base false
添加清华镜像源(国内推荐)
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
conda config --set show_channel_urls yes
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
conda config --set show_channel_urls yes
创建和管理环境
conda create -n xx_conda python=3.9conda activate xx_conda
conda activate xx_conda
卸载 Miniconda(如需)
rm -rf ~/miniconda3
清理 ~/.bashrc
中的 Conda 初始化代码(删除相关行)
模型下载
pip install modelscope
modelscope download --model Valdemardi/DeepSeek-R1-Distill-Qwen-32B-AWQ
3、编译开发组件安装
apt update 安装基本开发组件
组件名称 | 安装命令 | 检查命令 |
GUN调试工具(gdb) | sudo apt install gdb | gdb --version |
内存调试分析工具(valgrind) | sudo apt install valgrind | valgrind --version |
格式化C/C++代码的工具(clang - format) | sudo apt install clang - format | clang - format --version |
C/C++代码格式化和美化工具(astyle) | sudo apt install astyle | astyle --version |
版本控制工具组件(git) | sudo apt install git | git --version |
ssh设备访问工具 | sudo apt install openssh - client (客户端 ), sudo apt install openssh - server (服务端 ) | ssh -V |
网络工具组件(scp是ssh附带工具 ) | scp -V | |
下载网络内容的非交互工具(wget) | sudo apt install wget | wget --version |
容器镜像仓库工具组件(docker) | sudo apt-get update && sudo apt-get install docker - ce docker - ce - cli containerd.io | docker --version |
系统分析与监控工具组件(性能分析 = perf) | sudo apt install linux - tools - common linux - tools - generic | perf --version |
系统性能分析(sysstat) | sudo apt install sysstat | sar -V |
增强的交互式系统监控工具(htop) | sudo apt install htop | htop --version |
网络带宽工具组件(nload) | sudo apt install nload | nload --version |
文本编辑工具(vim) | sudo apt install vim | vim --version |
命令文本编辑器工具(nano) | sudo apt install nano | nano --version |
开发库和依赖工具组件(build - essential/tree ) | sudo apt install build - essential tree | tree --version |
树状图列出目录内容工具(tree) | tree --version | |
基本的网络诊断工具(ping) | 一般系统自带,若缺失sudo apt install iputils - ping | ping -V |
ifconfig | 一般系统自带,若缺失sudo apt install net - tools | ifconfig |
4、AI生态框架安装
框架名称 | |
PyTorch | |
LangChain | Introduction | 🦜️🔗 LangChain |
Stable Diffusion | Stable Diffusion API Docs | Stable Diffusion API Documentation |
SAM | https://github.com/facebookresearch/segment-anything GitCode - 全球开发者的开源社区,开源代码托管平台 |
Lora微调组件 |
4.1 LangChain-ChatChat 本地部署文档
一、系统要求
-
操作系统:Ubuntu 20.04+
-
Python:推荐 3.9 或 3.10
-
内存:建议 ≥ 8GB
-
GPU(可选):用于部署大语言模型(如 Qwen、ChatGLM)
-
Conda(推荐)或 venv 用于隔离环境
二、环境准备
1. 安装基础依赖
sudo apt update
sudo apt install -y git wget curl build-essential
2. 安装 Anaconda 或 Miniconda(推荐)
从官网下载安装:https://docs.conda.io/en/latest/miniconda.html
3. 创建并激活环境
conda create -n chatchat python=3.10 -y conda activate chatchat
三、克隆项目
git clone https://github.com/chatchat-space/langchain-chatchat.git
cd langchain-chatchat
四、安装依赖
pip install -r requirements.txt
安装较慢时可使用国内镜像:
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
五、下载语言模型
这里为了方便调用选择小模型
手动下载
从 Hugging Face 或模型官方仓库下载模型文件
六、配置启动参数
编辑 .env
文件(或创建)设置默认参数:
LLM_MODEL=Qwen3-0.6B-Base
VECTOR_STORE_TYPE=faiss
EMBEDDING_MODEL=text2vec-base-chinese
七、运行服务
项目使用 FastAPI + Web UI + 后端服务,推荐使用一键脚本:
# 启动服务(Web UI + FastAPI 后台)
bash start.sh
等待约 30~60 秒后,打开浏览器访问:
http://localhost:7860
八、测试功能
-
上传 PDF 文档、TXT、DOCX
-
输入自然语言提问
-
支持多模型切换
-
支持上下文记忆、插件功能、对话历史
4.2 SAM 部署流程文档
一、系统要求
-
操作系统:Ubuntu 20.04
-
Python:推荐使用 Python 3.9+
-
显卡驱动:NVIDIA GPU + CUDA 11.7 或以上
-
内存:≥8GB
-
显卡显存:≥6GB(建议使用 12GB 以上显存)
二、环境准备
1. 安装系统依赖
sudo apt update
sudo apt install -y git wget unzip build-essential libgl1-mesa-glx
2. 创建 Python 虚拟环境
# 安装 Python 虚拟环境工具(可选)
sudo apt install python3.9-venv# 创建并激活虚拟环境
python3.9 -m venv sam_env
source sam_env/bin/activate
三、安装 PyTorch
请根据你的 CUDA 版本从 PyTorch 获取对应安装指令。以下为示例:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
四、克隆 SAM 仓库
git clone https://github.com/facebookresearch/segment-anything.git
cd segment-anything
pip install -e .
五、下载 SAM 模型权重
可以从 Meta 官方提供的地址下载权重模型(需要梯子):
# 创建模型目录
mkdir -p ~/models/sam# 示例下载 ViT-H SAM 模型
wget https://dl.fbaipublicfiles.com/s egment_anything/sam_vit_h_4b8939.pth -P ~/models/sam
测试部署(命令行接口)
创建一个测试脚本 run_sam.py
:
import torch
from segment_anything import sam_model_registry, SamPredictor
import cv2# 加载模型
sam = sam_model_registry["vit_h"](checkpoint="/home/yourname/models/sam/sam_vit_h_4b8939.pth")
sam.to("cuda")predictor = SamPredictor(sam)# 加载图像
image = cv2.imread("test.jpg")
predictor.set_image(image)# 示例点
input_point = [[500, 375]]
input_label = [1]masks, scores, logits = predictor.predict(point_coords=input_point,point_labels=input_label,multimask_output=True,
)# 保存掩码图
import matplotlib.pyplot as plt
import numpy as npfor i, mask in enumerate(masks):plt.imsave(f"mask_{i}.png", mask, cmap="gray")
运行:
python run_sam.py
七、(可选)部署为 Web 服务
你可以使用 gradio
快速部署一个 Web UI:
pip install gradio opencv-python
# 创建 web_app.py
import gradio as gr
import cv2
from segment_anything import sam_model_registry, SamPredictorsam = sam_model_registry["vit_h"](checkpoint="/home/yourname/models/sam/sam_vit_h_4b8939.pth")
sam.to("cuda")
predictor = SamPredictor(sam)def segment(image, x, y):predictor.set_image(image)masks, _, _ = predictor.predict(point_coords=[[x, y]], point_labels=[1], multimask_output=False)return masks[0]gr.Interface(fn=segment,inputs=["image", gr.Number(label="X"), gr.Number(label="Y")],outputs="image").launch()
运行:
python web_app.py
八、常见问题
问题 | 解决方法 |
---|---|
RuntimeError: CUDA out of memory | 使用较小的模型(如 ViT-B)、减小图像尺寸 |
cv2.error: OpenCV | 确保图像路径正确,格式为 RGB |
下载权重失败 | 使用代理或科学上网工具 |
九、总结
部署 SAM 的关键在于:
-
安装兼容的 CUDA + PyTorch 环境;
-
下载官方模型权重;
-
使用
segment_anything
提供的 API 接口进行加载与预测; -
可选:集成 Web 服务或 API 接口用于后续调用。
4.3 Stable Diffusion 部署文档
使用 CompVis 官方仓库(或 Stability AI 的扩展版):
git clone https://github.com/CompVis/stable-diffusion.git
cd stable-diffusion
五、安装依赖
pip install -r requirements.txt
如果你看到 transformers
、diffusers
缺失,可以补装:
pip install diffusers transformers accelerate
六、下载模型权重
下载文件
https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt
七、运行图像生成(命令行)
python scripts/txt2img.py \
--prompt "a fantasy castle in the sky" \
--plms \
--ckpt models/ldm/stable-diffusion-v1/model.ckpt \
--config configs/stable-diffusion/v1-inference.yaml \
--outdir outputs/ \
--n_samples 1 \
--n_iter 1 \
--H 512 --W 512 \
--ddim_steps 50
八、输出结果查看
图片会保存在 outputs/
文件夹中。
九、测试示例
python scripts/txt2img.py \
--prompt "a cyberpunk cityscape at night, neon lights, rain" \
--ckpt models/ldm/stable-diffusion-v1/model.ckpt \
--config configs/stable-diffusion/v1-inference.yaml \
--outdir outputs/
十、常见问题解决
问题 | 解决方案 |
---|---|
CUDA out of memory | 尝试减小图像尺寸(如 384x384)、使用 --n_samples 1 |
No module named xformers | 可选模块,pip install xformers (需编译) |
权重文件太大下载失败 | 可手动从浏览器下载 .ckpt 文件,放入指定目录 |
十一、可选扩展
-
使用
diffusers
+transformers
替代原生结构,部署更轻量 -
部署 Gradio Web UI 示例(需要我可补充)
-
Docker 化部署
-
使用自动 Web UI(如 A1111 或 ComfyUI)
总结
Stable Diffusion 原生部署流程为:
-
安装环境(Python + PyTorch)
-
克隆项目并安装依赖
-
下载模型权重
-
使用
txt2img.py
脚本生成图像
4.4 Lora微调框架安装
5、AI加速框架安装
加速组件 | |
vLLM | vLLM - vLLM |
vllm安装
6、KVM虚拟化和Docker容器安装
容器化组件 | |
Docker | https://docs.docker.com/ |
docker-compose | |
Harbor镜像仓库 | Harbor docs | Harbor 2.4 Documentation |