Add model.capabilities configuration for client detection

[mostlygeek]

This issue is a follows up discussion #713 on automatically adding capabilities metadata to v1/models responses. I'm opening this up to gather some community feedback before working on it.

LLM clients are using metadata from v1/models to determine the capabilities of a model. Within llama-swap a new models.capabilities configuration would contain metadata about what the local inference server supports.

models:
  model_id: 
    capabilities:
      text: 
        completions: true. # /chat/completions (openai)
        responses: false # /responses (openai)
        messages: false # /messages (anthropic)
        fim: true # (llama.cpp FIM)
      audio: 
        speech: true # text to speech
        transcriptions: true # ASR
      image:
        vision: true

Design:

• model.capabilities is a map of key and value
  • values can be a boolean, string, int, map, etc
• valid keys are defined in config-schema.json
• backwards compatibility is to be maintained. Default (empty map) generates no capabilities metadata.
• v1/models endpoint will use capabilities to generate metadata for clients
  • this is no published standard for clients so support the popular projects first (tbd)
• The v1 of capabilities to introduce a programming pattern that makes adding new client support easier in the future.

[github-actions]

This issue is stale because it has been open for 2 weeks with no activity.