feat: Add support for Docker Compose models

[UMMAN2005]

Description

This PR adds support for Docker Compose's models feature at the root level, as introduced in the Docker Compose specification for AI/ML workloads.

What does this PR do?

Implements support for the models section in Docker Compose, allowing provisioners to define AI/ML models that can be referenced by services.

Changes:

• Add ComposeModels field to ProvisionOutput struct in core.go
• Apply models to compose project during provisioning in ApplyToStateAndProject
• Add models template support to template provisioner (templateprov)
• Add test coverage for models functionality

This enables provisioners to generate compose files with models like:

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

models:
  llm:
    model: ai/smollm2

Closes #367

Types of changes

• Bug fix (non-breaking change which fixes an issue)

• New feature (non-breaking change which adds functionality)

• Breaking change (fix or feature that would cause existing functionality to change)

• New chore (expected functionality to be implemented)

Checklist:

• My change requires a change to the documentation.

  • I have updated the documentation accordingly.

• I've signed off with an email address that matches the commit author.

[UMMAN2005]

@astromechza
@mevrin-ueat

Please have a look. This is my first PR in this project.😊

[mathieu-benoit]

Thanks, @UMMAN2005, sounds great!

FYI: I'm fixing the CI failing for an unrelated reason of your PR.

And the review of this PR will happen shortly.

[astromechza]

The code looks fine, can you please update 06-resource-provisioning with the new model template bit and ensure the output is as expected? Also update the README with a description of the models field.

[UMMAN2005]

Hi, @astromechza

Could you please have a look?

[mathieu-benoit]

Thanks, @UMMAN2005 for this implementation.

I just tested with this:

apiVersion: score.dev/v1b1
metadata:
  name: open-webui
containers:
  open-webui:
    image: .
resources:
  gemma3:
    type: llm-model
    params:
      model: ai/gemma3:270M-UD-IQ2_XXS

With this provisioner:

- uri: template://dmr-llm-model
  type: llm-model
  description: Generates the LLM model via the Docker Model Runner (DMR) provider.
  supported_params:
    - model
  outputs: |
    model: {{ .Init.model }}
  expected_outputs:
    - model
  init: |
    model: {{ .Params.model | default "ai/smollm2:135M-Q4_0" }}
  models: |
    {{ .Id }}:
      model: {{ .Init.model }}

By running this commands:

./score-compose init \
  --provisioners model.provisioners.yaml

./score-compose generate score.yaml \
  --image ghcr.io/open-webui/open-webui:main-slim \
  --publish 8080:open-webui:8080 \
  --output compose.yaml

It's generating this compose.yaml:

name: score-compose
services:
    open-webui-open-webui:
        annotations:
            compose.score.dev/workload-name: open-webui
        hostname: open-webui
        image: ghcr.io/open-webui/open-webui:main-slim
models:
    open-webui-gemma3:
        model: ai/gemma3:270M-UD-IQ2_XXS

But when running docker compose up -d, the local LLM models are not pulled (docker model list).

Here is what's missing in the generated compose.yaml file:

name: score-compose
services:
    open-webui-open-webui:
        annotations:
            compose.score.dev/workload-name: open-webui
        hostname: open-webui
        image: ghcr.io/open-webui/open-webui:main-slim
+       models:
+         - open-webui-gemma3
models:
    open-webui-gemma3:
        model: ai/gemma3:270M-UD-IQ2_XXS

Reference: https://docs.docker.com/ai/compose/models-and-compose/#basic-model-definition.

And then when doing docker compose up -d, we can see that the local LLM model is pulled (docker model list):

MODEL NAME              PARAMETERS  QUANTIZATION  ARCHITECTURE  MODEL ID      CREATED       CONTEXT  SIZE       
gemma3:270M-UD-IQ2_XXS  268.10 M    IQ4_NL        gemma3        bae1ee104a16  3 months ago           165.54 MiB

@UMMAN2005, could you please look at how this could be injected?

@astromechza, maybe some advice to accomplish this part?

[UMMAN2005]

Hi, @mathieu-benoit

The changes that you mentioned are integrated and tested to be working.

Please have a look.

[mathieu-benoit]

@UMMAN2005, working as expected now, thanks!

score.yaml:

apiVersion: score.dev/v1b1
metadata:
  name: open-webui
containers:
  open-webui:
    image: .
    variables:
      OPENAI_API_BASE_URL: "${resources.smollm2.url}"
      WEBUI_NAME: "Hello, DMR with Score Compose!"
    volumes:
      /app/backend/data:
        source: ${resources.data}
resources:
  data:
    type: volume
  gemma3:
    type: llm-model
    params:
      model: ai/gemma3:270M-UD-IQ2_XXS
  smollm2:
    type: llm-model
    params:
      model: ai/smollm2:135M-Q2_K
service:
  ports:
    tcp:
      port: 8080
      targetPort: 8080

model.provisioners.yaml:

- uri: template://dmr-llm-model
  type: llm-model
  description: Generates the LLM model via the Docker Model Runner (DMR) provider.
  supported_params:
    - model
    - context_size
  outputs: |
    model: {{ .Init.model }}
    url: "http://172.17.0.1:12434/engines/v1/"
  expected_outputs:
    - model
    - url
  init: |
    model: {{ .Params.model | default "ai/smollm2:135M-Q4_0" }}
  models: |
    {{ .Id }}:
      model: {{ .Init.model }}
      context_size: {{ .Params.context_size | default 2048 }}

./score-compose init   --provisioners model.provisioners.yaml

./score-compose generate score.yaml   --image ghcr.io/open-webui/open-webui:main-slim   --publish 8080:open-webui:8080   --output compose.yaml

compose.yaml generated:

name: score-compose
services:
    open-webui-open-webui:
        annotations:
            compose.score.dev/workload-name: open-webui
        environment:
            OPENAI_API_BASE_URL: http://172.17.0.1:12434/engines/v1/
            WEBUI_NAME: Hello, DMR with Score Compose!
        hostname: open-webui
        image: ghcr.io/open-webui/open-webui:main-slim
        models:
            open-webui.gemma3: {}
            open-webui.smollm2: {}
        ports:
            - target: 8080
              published: "8080"
        volumes:
            - type: volume
              source: open-webui-data-K3WUDl
              target: /app/backend/data
volumes:
    open-webui-data-K3WUDl:
        name: open-webui-data-K3WUDl
        driver: local
        labels:
            dev.score.compose.res.uid: volume.default#open-webui.data
models:
    open-webui.gemma3:
        model: ai/gemma3:270M-UD-IQ2_XXS
        context_size: 2048
    open-webui.smollm2:
        model: ai/smollm2:135M-Q2_K
        context_size: 2048

docker compose up -d

docker ps:

CONTAINER ID   IMAGE                                     COMMAND               CREATED         STATUS                   PORTS                                                     NAMES
bd92d6a541a8   ghcr.io/open-webui/open-webui:main-slim   "bash start.sh"       7 minutes ago   Up 7 minutes (healthy)   0.0.0.0:8080->8080/tcp, [::]:8080->8080/tcp               score-compose-open-webui-open-webui-1
cb731dcfbe22   docker/model-runner:latest                "/app/model-runner"   7 minutes ago   Up 7 minutes             127.0.0.1:12434->12434/tcp, 172.17.0.1:12434->12434/tcp   docker-model-runner

docker model list:

MODEL NAME              PARAMETERS  QUANTIZATION  ARCHITECTURE  MODEL ID      CREATED       CONTEXT  SIZE       
gemma3:270M-UD-IQ2_XXS  268.10 M    IQ4_NL        gemma3        bae1ee104a16  3 months ago           165.54 MiB  
smollm2:135M-Q2_K       134.52 M    Q2_K          llama         eba11bf8f361  7 months ago           82.41 MiB

And then, on localhost:8080:

🥳 🚀

[mathieu-benoit]

LGTM

@chris-stephenson or @astromechza, a second pair of eyes on the code please?

[UMMAN2005]

@chris-stephenson and @astromechza

Kind reminder!