标签： 一粒云

---

​​一、硬件与基础环境准备​​

​​1. 服务器配置要求​​

• ​​GPU​​：6×NVIDIA A5000（24GB显存/卡，共144GB显存）
• ​​内存​​：≥64GB RAM
• ​​存储​​：≥500GB SSD（推荐NVMe）
• ​​系统​​：Ubuntu 22.04 LTS / Debian 12

​​2. 环境初始化​​

# 安装基础工具
sudo apt update && sudo apt install -y docker.io nvidia-container-toolkit
# 配置Docker使用NVIDIA GPU
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

---

​​二、VLLM多卡部署（6卡优化）​​

​​1. 安装vLLM​​

# 创建虚拟环境
conda create -n vllm python=3.10 -y && conda activate vllm
# 安装vLLM（推荐0.5.4+）
pip install vllm -i https://pypi.tuna.tsinghua.edu.cn/simple

​​2. 启动6卡推理服务​​

vllm serve --model /path/to/model \  
   --tensor-parallel-size 6 \          # 并行数=GPU数量
   --gpu-memory-utilization 0.85 \     # 显存利用率阈值（6卡建议0.8~0.9）
   --max-num-seqs 64 \                 # 高并发优化
   --enforce-eager \                   # 避免多卡兼容问题
   --port 8000 \                       # 服务端口
   --api-key "your-token"              # 访问令牌（增强安全性）

---

​​三、Dify部署与对接VLLM​​

​​1. 部署Dify服务​​

# 拉取Dify代码
git clone https://github.com/langgenius/dify.git
cd dify/docker

# 修改配置（关键步骤）
cp .env.example .env
nano .env  # 修改以下参数：

# 模型端点指向VLLM服务
MODEL_PROVIDER=vllm
VLLM_API_BASE=http://localhost:8000/v1  # VLLM的OpenAI兼容API地址
VLLM_MODEL_NAME=your-model-name         # 与vLLM启动时的模型名一致

​​2. 启动Dify​​

docker compose up -d  # 自动构建容器

---

​​四、外部应用API调用方法​​

​​1. 通过Dify调用（业务层）​​

• ​​Dify API地址​​：http://<服务器IP>:80/v1（默认端口）
• ​​认证​​：Header中添加 Authorization: Bearer {DIFY_API_KEY}
• ​​请求示例​​（生成文本）：

import requests
url = "http://<服务器IP>/v1/completion"
data = {
  "inputs": "你好，介绍一下vLLM",
  "response_mode": "blocking"
}
headers = {"Authorization": "Bearer dify-api-key"}
response = requests.post(url, json=data, headers=headers)

​​2. 直接调用VLLM（高性能场景）​​

# 使用OpenAI兼容API（Python示例）
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="your-token")
response = client.chat.completions.create(
  model="your-model-name",
  messages=[{"role": "user", "content": "解释量子力学"}]
)

---

​​五、VLLM对比Ollama的核心优势​​

​​维度​​	​​VLLM​​	​​Ollama​​
​​多卡支持​​	✅ 原生6卡张量并行（--tensor-parallel-size=6）	❌ 仅支持单卡，多卡需手动切换
​​吞吐量​​	⭐ 连续批处理技术，6卡并发提升5-10倍	⚠️ 单请求处理，并发能力弱
​​生产就绪​​	✅ 工业级部署（API密钥、监控、扩缩容）	❌ 定位开发测试，无企业级特性
​​显存管理​​	✅ PagedAttention动态分配，支持百亿模型	⚠️ 全模型加载，易OOM
​​安全性​​	✅ 内置API密钥认证	❌ 默认无认证，需Nginx反向代理

💡 ​​关键结论​​：
VLLM是​​生产级AI服务​​的首选，尤其适合高并发、低延迟场景（如API服务）；
Ollama更适合​​本地快速原型验证​​，但在多卡利用率和安全性上存在明显短板。

---

​​六、常见问题排查​​

1. ​​多卡启动失败​​： export VLLM_WORKER_MULTIPROC_METHOD=spawn # 解决多进程卡死
2. ​​显存不足​​：
   • 降低--gpu-memory-utilization至0.7
   • 添加--swap-space 16 使用主机内存扩展
3. ​​Dify连接VLLM失败​​：
   • 检查.env中VLLM_API_BASE是否含/v1路径
   • 确保vLLM启动参数含--api-key且与Dify配置一致

部署完成后，可通过 nvidia-smi 监控GPU利用率，正常运行时6卡负载应均衡（±5%差异）。

英文参考原文：Based on the information available, here’s a comparison of vLLM and Ollama, two popular frameworks for running large language models (LLMs) locally: 

vLLM 

• Focus: High-throughput, low-latency LLM inference and serving, particularly suited for production environments.
• Key Features:
  • PagedAttention: A memory management technique that optimizes GPU memory usage for faster inference speeds, especially with long sequences and large models.
  • Continuous Batching: Processes incoming requests dynamically to maximize hardware utilization.
  • High Performance: Consistently delivers superior throughput and lower latency, particularly for concurrent requests.
  • Scalability: Designed for scalability, including support for tensor parallelism and pipeline parallelism for distributed inference across multiple GPUs or nodes.
  • OpenAI-compatible API: Simplifies integration with applications.
• Hardware Requirements: Optimized for high-end, CUDA-enabled NVIDIA GPUs, although it technically supports CPU inference (less optimized).
• Ease of Use: Offers more control and optimization options but has a steeper learning curve, requiring more technical knowledge for setup. 

Ollama 

• Focus: User-friendly, local deployment and management of LLMs, prioritizing simplicity and accessibility.
• Key Features:
  • Ease of Use: Offers a streamlined workflow for downloading, running, and managing models with a simple command-line interface (CLI) and an OpenAI-compatible API.
  • Broad Hardware Compatibility: Works well on both GPUs and CPUs, making it accessible to users with consumer-grade hardware.
  • Local Deployment with Privacy: Ensures data privacy and control by keeping data processing within your local environment.
  • Adaptable: Supports various model types and offers token streaming for faster responses.
  • Growing Performance: While potentially slower than vLLM on high-end GPUs, recent updates have significantly improved its performance.
• Hardware Requirements: Designed to work reasonably well even on consumer-grade hardware.
• Ease of Use: Prioritizes simplicity, making it easy to install and run models with just a few commands. 

In Summary:

• Choose vLLM when: You need maximum performance and scalability in production environments, especially when utilizing high-end GPUs for high-throughput workloads.
• Choose Ollama when: You prioritize ease of use, broad hardware compatibility (including CPU-only setups), and local data privacy for development, prototyping, or simpler projects. 

Hybrid Approach:

It’s worth considering a hybrid approach where you use Ollama for development and prototyping and then deploy with vLLM in production for optimal performance.