Define AI Models in Docker Compose applications

Table of contents
Requires: Docker Compose 2.38.0 and later

Compose lets you define AI models as core components of your application, so you can declare model dependencies alongside services and run the application on any platform that supports the Compose Specification.

Prerequisites

  • Docker Compose v2.38 or later
  • A platform that supports Compose models such as Docker Model Runner (DMR) or compatible cloud providers. If you are using DMR, see the requirements.

What are Compose models?

Compose models are a standardized way to define AI model dependencies in your application. By using the models top-level element in your Compose file, you can:

  • Declare which AI models your application needs
  • Specify model configurations and requirements
  • Make your application portable across different platforms
  • Let the platform handle model provisioning and lifecycle management

Basic model definition

To define models in your Compose application, use the models top-level element:

services:
  chat-app:
    image: my-chat-app
    models:
      - llm

models:
  llm:
    model: ai/smollm2

This example defines:

  • A service called chat-app that uses a model named llm
  • A model definition for llm that references the ai/smollm2 model image

Model configuration options

Models support various configuration options:

models:
  llm:
    model: ai/smollm2
    context_size: 1024
    runtime_flags:
      - "--a-flag"
      - "--another-flag=42"

Common configuration options include:

  • model (required): The OCI artifact identifier for the model. This is what Compose pulls and runs via the model runner.

  • context_size: Defines the maximum token context size for the model.

    Note

    Each model has its own maximum context size. When increasing the context length, consider your hardware constraints. In general, try to keep context size as small as feasible for your specific needs.

  • runtime_flags: A list of raw command-line flags passed to the inference engine when the model is started. For example, if you use llama.cpp, you can pass any of the available parameters.

  • Platform-specific options may also be available via extension attributes x-*

Tip

See more example in the Common runtime configurations section.

Service model binding

Services can reference models in two ways: short syntax and long syntax.

Short syntax

The short syntax is the simplest way to bind a model to a service:

services:
  app:
    image: my-app
    models:
      - llm
      - embedding-model

models:
  llm:
    model: ai/smollm2
  embedding-model:
    model: ai/all-minilm

With short syntax, the platform automatically generates environment variables based on the model name:

  • LLM_URL - URL to access the LLM model
  • LLM_MODEL - Model identifier for the LLM model
  • EMBEDDING_MODEL_URL - URL to access the embedding-model
  • EMBEDDING_MODEL_MODEL - Model identifier for the embedding-model

Long syntax

The long syntax allows you to customize environment variable names:

services:
  app:
    image: my-app
    models:
      llm:
        endpoint_var: AI_MODEL_URL
        model_var: AI_MODEL_NAME
      embedding-model:
        endpoint_var: EMBEDDING_URL
        model_var: EMBEDDING_NAME

models:
  llm:
    model: ai/smollm2
  embedding-model:
    model: ai/all-minilm

With this configuration, your service receives:

  • AI_MODEL_URL and AI_MODEL_NAME for the LLM model
  • EMBEDDING_URL and EMBEDDING_NAME for the embedding model

Platform portability

One of the key benefits of using Compose models is portability across different platforms that support the Compose specification.

Docker Model Runner

When Docker Model Runner is enabled:

services:
  chat-app:
    image: my-chat-app
    models:
      llm:
        endpoint_var: AI_MODEL_URL
        model_var: AI_MODEL_NAME

models:
  llm:
    model: ai/smollm2
    context_size: 4096
    runtime_flags:
      - "--no-prefill-assistant"

Docker Model Runner will:

  • Pull and run the specified model locally
  • Provide endpoint URLs for accessing the model
  • Inject environment variables into the service

Cloud providers

The same Compose file can run on cloud providers that support Compose models:

services:
  chat-app:
    image: my-chat-app
    models:
      - llm

models:
  llm:
    model: ai/smollm2
    # Cloud-specific configurations
    x-cloud-options:
      - "cloud.instance-type=gpu-small"
      - "cloud.region=us-west-2"

Cloud providers might:

  • Use managed AI services instead of running models locally
  • Apply cloud-specific optimizations and scaling
  • Provide additional monitoring and logging capabilities
  • Handle model versioning and updates automatically

Common runtime configurations

Below are some example configurations for various use cases.

Development

services:
  app:
    image: app
    models:
      dev_model:
        endpoint_var: DEV_URL
        model_var: DEV_MODEL

models:
  dev_model:
    model: ai/model
    context_size: 4096
    runtime_flags:
      - "--verbose"                       # Set verbosity level to infinity
      - "--verbose-prompt"                # Print a verbose prompt before generation
      - "--log-prefix"                    # Enable prefix in log messages
      - "--log-timestamps"                # Enable timestamps in log messages
      - "--log-colors"                    # Enable colored logging

Conservative with disabled reasoning

services:
  app:
    image: app
    models:
      conservative_model:
        endpoint_var: CONSERVATIVE_URL
        model_var: CONSERVATIVE_MODEL

models:
  conservative_model:
    model: ai/model
    context_size: 4096
    runtime_flags:
      - "--temp"                # Temperature
      - "0.1"
      - "--top-k"               # Top-k sampling
      - "1"
      - "--reasoning-budget"    # Disable reasoning
      - "0"

Creative with high randomness

services:
  app:
    image: app
    models:
      creative_model:
        endpoint_var: CREATIVE_URL
        model_var: CREATIVE_MODEL

models:
  creative_model:
    model: ai/model
    context_size: 4096
    runtime_flags:
      - "--temp"                # Temperature
      - "1"
      - "--top-p"               # Top-p sampling
      - "0.9"

Highly deterministic

services:
  app:
    image: app
    models:
      deterministic_model:
        endpoint_var: DET_URL
        model_var: DET_MODEL

models:
  deterministic_model:
    model: ai/model
    context_size: 4096
    runtime_flags:
      - "--temp"                # Temperature
      - "0"
      - "--top-k"               # Top-k sampling
      - "1"

Concurrent processing

services:
  app:
    image: app
    models:
      concurrent_model:
        endpoint_var: CONCURRENT_URL
        model_var: CONCURRENT_MODEL

models:
  concurrent_model:
    model: ai/model
    context_size: 2048
    runtime_flags:
      - "--threads"             # Number of threads to use during generation
      - "8"
      - "--mlock"               # Lock memory to prevent swapping

Rich vocabulary model

services:
  app:
    image: app
    models:
      rich_vocab_model:
        endpoint_var: RICH_VOCAB_URL
        model_var: RICH_VOCAB_MODEL

models:
  rich_vocab_model:
    model: ai/model
    context_size: 4096
    runtime_flags:
      - "--temp"                # Temperature
      - "0.1"
      - "--top-p"               # Top-p sampling
      - "0.9"

Alternative configuration with provider services

Important

This approach is deprecated. Use the models top-level element instead.

You can also use the provider service type, which allows you to declare platform capabilities required by your application. For AI models, you can use the model type to declare model dependencies.

To define a model provider:

services:
  chat:
    image: my-chat-app
    depends_on:
      - ai_runner

  ai_runner:
    provider:
      type: model
      options:
        model: ai/smollm2
        context-size: 1024
        runtime-flags: "--no-prefill-assistant"

Reference