MAX Serve Multi-Modal Structured Output

In this recipe, we will cover:

• How to run a multimodal vision model using Llama 3.2 Vision and MAX Serve
• Implementing structured output parsing with Pydantic models
• Converting image analysis into strongly-typed JSON

We'll walk through building a solution that showcases:

• MAX Serve capabilities with multimodal models
• Structured output parsing with type safety
• Simple deployment using magic CLI

Let's get started.

Requirements

Please make sure your system meets our system requirements.

To proceed, ensure you have the magic CLI installed:

curl -ssL https://magic.modular.com/ | bash

or update it via:

magic self-update

You'll need:

• Access to meta-llama/Llama-3.2-11B-Vision-Instruct model
• A valid Hugging Face token for accessing Llama 3.2 Vision

Set up your environment variables:

export HUGGING_FACE_HUB_TOKEN=

GPU requirements

Structured output with MAX Serve requires GPU access. For running the app on GPU, ensure your system meets these GPU requirements:

• Supported GPUs: NVIDIA A100 (optimized), A10G, L4, or L40
• NVIDIA Drivers: Installation guide here
• NVIDIA Container Toolkit: Installation guide here

Quick start

1. Download the code for this recipe using git:

   git clone https://github.com/modular/max-recipes.git
   cd max-recipes/max-serve-multimodal-structured-output
   

2. Running the vision model:

   magic run app
   

   This will start MAX Serve with Llama 3.2 Vision and run the example code that extracts player information from a basketball image.

   

   The output will look like:

   {
   "players": [
       {
       "name": "Klay Thompson",
       "number": 11
       },
       {
       "name": "Stephen Curry",
       "number": 30
       },
       {
       "name": "Kevin Durant",
       "number": 35
       }
   ]
   }
   

3. And once done, to clean up the resources run:

   magic run clean
   

Features of Llama 3.2 Vision structured output

• Multimodal processing: Handles both image and text inputs seamlessly
• OpenAI-compatible API: Works with standard OpenAI client libraries
• Type-safe output: Uses Pydantic models to ensure structured and validated JSON output

Structured output with Pydantic

The core of our structured output system uses Pydantic models to define the expected data structure:

from pydantic import BaseModel

class Player(BaseModel):
    name: str = Field(description="Player name on jersey")
    number: int = Field(description="Player number on jersey")

class Players(BaseModel):
    players: List[Player] = Field(description="List of players visible in the image")

How it works:

• Type validation: Pydantic ensures all fields match their expected types
• Field descriptions: Help guide the model in extracting the right information
• Nested structures: Support complex data hierarchies for detailed information extraction

Vision model integration

The application uses the OpenAI client to communicate with MAX Serve:

from openai import OpenAI

client = OpenAI(api_key="local", base_url="http://localhost:8000/v1")

completion = client.beta.chat.completions.parse(
    model="meta-llama/Llama-3.2-11B-Vision-Instruct",
    messages=[
        {"role": "system", "content": "Extract player information from the image."},
        {"role": "user", "content": [
            {
                "type": "text",
                "text": "Please provide a list of players visible in this photo with their jersey numbers."
            },
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://ei.marketwatch.com/Multimedia/2019/02/15/Photos/ZH/MW-HE047_nbajer_20190215102153_ZH.jpg"
                }
            }
        ]},
    ],
    response_format=Players,
)

Key components:

• Multimodal input: Combines text instructions with image data
• Structured parsing: Uses the Players model to format the response
• Inference: Runs on MAX Serve with OpenAI-compatible API

MAX Serve Options

To enable structured output in MAX Serve, simply include --enable-structured-output.

To see other options, make sure to check out the help:

max-pipelines serve --help

Conclusion

In this recipe, we've built a system for extracting structured data from images using Llama 3.2 Vision and MAX Serve. We've explored:

• Basic setup: Using MAX Serve with GPU support
• Structured output: Implementing type-safe data extraction with Pydantic
• Multimodal processing: Handling both image and text inputs
• Local deployment: Running the model with MAX Serve

This implementation provides a foundation for building more complex vision-based applications with structured output.

Next steps

• Deploy Llama 3.2 Vision on GPU with MAX Serve to AWS, GCP or Azure or on Kubernetes
• Explore MAX's documentation for additional features
• Join our Modular Forum and Discord community to share your experiences and get support

We're excited to see what you'll build with Llama 3.2 Vision and MAX! Share your projects and experiences with us using #ModularAI on social media.