vLLM is a high-performance inference engine for large language models that provides significant performance improvements for local inference. It’s designed to maximize throughput and memory efficiency for serving LLMs.

​
Prerequisites

  1. Install vLLM: 
    pip install vllm
  2. Start vLLM server: 
    # For testing with a small model
vllm serve microsoft/DialoGPT-medium --port 8000

# For production with a larger model (requires GPU)
vllm serve Qwen/Qwen2.5-32B-Instruct --port 8000

​
Usage

import os
from mem0 import Memory

os.environ["OPENAI_API_KEY"] = "your-api-key"  # used for embedding model

config = {
    "llm": {
        "provider": "vllm",
        "config": {
            "model": "Qwen/Qwen2.5-32B-Instruct",
            "vllm_base_url": "http://localhost:8000/v1",
            "temperature": 0.1,
            "max_tokens": 2000,
        }
    }
}

m = Memory.from_config(config)
messages = [
    {"role": "user", "content": "I'm planning to watch a movie tonight. Any recommendations?"},
    {"role": "assistant", "content": "How about thriller movies? They can be quite engaging."},
    {"role": "user", "content": "I'm not a big fan of thrillers, but I love sci-fi movies."},
    {"role": "assistant", "content": "Got it! I'll avoid thrillers and suggest sci-fi movies instead."}
]
m.add(messages, user_id="alice", metadata={"category": "movies"})

​
Configuration Parameters

ParameterDescriptionDefaultEnvironment Variable
modelModel name running on vLLM server"Qwen/Qwen2.5-32B-Instruct"-
vllm_base_urlvLLM server URL"http://localhost:8000/v1"VLLM_BASE_URL
api_keyAPI key (dummy for local)"vllm-api-key"VLLM_API_KEY
temperatureSampling temperature0.1-
max_tokensMaximum tokens to generate2000-

​
Environment Variables

You can set these environment variables instead of specifying them in config: 
export VLLM_BASE_URL="http://localhost:8000/v1"
export VLLM_API_KEY="your-vllm-api-key"
export OPENAI_API_KEY="your-openai-api-key"  # for embeddings

​
Benefits

  • High Performance: 2-24x faster inference than standard implementations
  • Memory Efficient: Optimized memory usage with PagedAttention
  • Local Deployment: Keep your data private and reduce API costs
  • Easy Integration: Drop-in replacement for other LLM providers
  • Flexible: Works with any model supported by vLLM

​
Troubleshooting

  1. Server not responding: Make sure vLLM server is running 
    curl http://localhost:8000/health
  2. 404 errors: Ensure correct base URL format 
    "vllm_base_url": "http://localhost:8000/v1"  # Note the /v1
  3. Model not found: Check model name matches server
  4. Out of memory: Try smaller models or reduce max_model_len 
    vllm serve Qwen/Qwen2.5-32B-Instruct --max-model-len 4096

​
Config

All available parameters for the vllm config are present in Master List of All Params in Config.