Skip to main content

vLLM Omni

Inferoa treats multimodal capabilities as endpoint-backed tools. A single main chat model can delegate image, video, audio, or speech work to configured vLLM-Omni-compatible endpoints.

Capabilities​

Supported endpoint keys are:

  • vision
  • image_generation
  • image_edit
  • video_understanding
  • video_generation
  • audio_understanding
  • audio_generation
  • speech

The exact capabilities depend on the served model profile and routes exposed by the endpoint. A route being present is not enough for acceptance; a compatible profile must complete a real tool invocation.

Example Configuration​

omni:
  enabled: true
  endpoints:
    vision:
      base_url: http://localhost:8091/v1
      model: qwen2.5-omni-7b
    image_generation:
      base_url: http://localhost:8092/v1
      model: image-model
    video_generation:
      base_url: http://localhost:8093/v1
      model: video-model

Use /setup for interactive configuration so secrets are stored in the vault.

Artifact Handling​

Generated media is stored as managed resources rather than pasted into the prompt. This keeps prompts smaller and preserves stable references in the session log.

Validation​

Use the runtime validation commands when working against real Omni endpoints:

npm run validate:omni-real
npm run validate:omni-e2e-runtime

Use /doctor run for user-facing endpoint health. Strict release acceptance is available through the debug runner when validating a release environment.