feat: start a recipe with llama-stack backend

[feloy]

Signed-off-by: Philippe Martin phmartin@redhat.com

What does this PR do?

This PR is a first step towards being able to run a recipe using llama stack as backend.

This PR covers:

• AI Lab supports recipes with backend="llama-stack"
• for such recipe, no model is asked in the "Start Recipe" form
• when AI Lab deploys a recipe with llama-stack backend, it does not deploy any inference server
• the MODEL_ENDPOINT passed to the recipe is the endpoint of Llama Stack
• all models served by Llama stack are accessible to the app. It is the responsability of the app to choose with which model to work

Screenshot / video of UI

llama-stack-recipe-2.mp4

What issues does this PR fix or reference?

Fixes #2625

How to test this PR?

1. Add this recipe to your ~/.local/share/containers/podman-desktop/extensions-storage/redhat.ai-lab/user-catalog.json file:

{
  "version": "1.0",
  "recipes": [
    {
      "id": "chatbot-llama-stack",
      "description": "This recipe provides a blueprint for developers to create their own AI-powered chat applications using Streamlit and llama-stack.",
      "name": "ChatBot using Llama Stack",
      "repository": "https://github.com/feloy/chatbot-llama-stack-recipe",
      "ref": "main",
      "icon": "natural-language-processing",
      "categories": ["natural-language-processing"],
      "basedir": "/",
      "readme": "# Chat Application using Llama Stack\n",
      "backend": "llama-stack",
      "languages": ["python"],
      "frameworks": ["streamlit", "llama-stack"]
    }
  ],
  "models": [],
  "categories": []
}

2. The app will work with the first model listed by llama stack: be sure to have only models compatible with chat served by Llama stack
3. Go to the Recipes catalog and start the "ChatBot using Llama Stack" recipe
4. Go to the running application details and open the app url

[axel7083]

suggestion: use a single if model && !recipeComponents.inferenceServer

[axel7083]

I don't really like that, use something more understandable like <none> or none

[feloy]

Sure, me neither:

Note: I'm not very happy with the way the applicationManager functions accept an optional model, depending on if we want to use llama-stack or not. Let's first discuss the high-level architecture of this solution, I can change these implementation details when we agree on the architecture

[axel7083]

Sorry I went to code too fast 🙏

Let's first discuss the high-level architecture of this solution,

What depends on the modelId? We use it for operation like hasApplicationPod & removeApplication. Maybe we should only use the recipeId as key?

[feloy]

I'm not very comfortable changing the labels, as the model label is also used for the tasks (the user can start the same recipe with different models, and we want to differentiate the tasks for them).

I have pushed a new commit changing how the parameters are passed: having explicit options inferModel / startLlamaStack, which enforce the other parameters (the model parameter for the moment). I also changed this arbitrary 0 model Id to <none>

@axel7083 WDYT?

[gastoner]

Please add "llama-stack" here https://github.com/containers/podman-desktop-extension-ai-lab/blob/main/packages/frontend/src/lib/RecipeCardTags.ts#L25 so it works with other tags (I think this should be only change to make it work)

[axel7083]

why not including the labels in the options? feel weird to move everything inside ApplicationOptions but the labels?

[feloy]

First reason is because labels is not part of the parameters of requestPullApplication, and I would like to keep the same options structure everywhere