Fix azure deployed gpt-image image generation by not sending response_format

[Takalele]

The image_generation view injects response_format=b64_json into every request body, but gpt-image models (Azure and OpenAI) reject that parameter - they only return b64_json. Requests to a gpt-image deployment therefore failed with
400 "Unknown parameter: 'response_format'".

Drop response_format for gpt-image* relay models before the request is sent. dall-e deployments keep the parameter and are unaffected.

Example router config entry for an Azure AI Foundry gpt-image deployment (sanitized - api_base and api_key are placeholders):

• model_name: cloud-gpt-image-2
  litellm_params:
  model: azure/gpt-image-2-example
  api_base: https://example.cognitiveservices.azure.com/openai/v1
  api_key: <api_key>
  api_version: 2025-04-01-preview
  model_info: id: cloud-gpt-image-2
  mode: image_generation aliases:
  • gpt-image-2

[meffmadd]

I would ideally like to avoid hardcoding specific models in the endpoints... could we maybe (ab)use the model_info config option for this?

[Takalele]

Agreed, i added a dedicated supports_response_format option to each deployment's model_info (following LiteLLM's supports_* naming), read by a small helper in gateway/config.py alongside the existing aliases / request_limit_multiplier handling.

• supports_response_format: false → the endpoint omits response_format for that model
• absent or true (the default) → current behavior, so existing configs (dall-e etc.) are unaffected
  The hardcoded model_relay.startswith("gpt-image") check is gone — the endpoint just asks the router config, e.g. for the Azure deployment:

- model_name: cloud-gpt-image-2
  litellm_params:
    model: azure/gpt-image-2-example
    ...
  model_info:
    id: cloud-gpt-image-2
    mode: image_generation
    supports_response_format: false

Verified with unit tests for the helper plus integration tests that assert the actual upstream request body (no response_format for flagged models, b64_json for unflagged ones), and end-to-end against a live Azure gpt-image deployment (400 without the flag → 200 with it). Happy to adjust the option name if you'd prefer something else.

[meffmadd]

LGTM!