Bug: Wrong model used for summarization despite config settings

[alozowski]

Description:
I encountered a bug where the model defined for summarization in the model_roles configuration is not being used. Instead, the model defined above is used, likely due to how step names are handled internally.

Configuration example:

model_roles:
  ingestion:
    - Qwen/Qwen2.5-VL-72B-Instruct
  summarization:
    - Qwen/Qwen2.5-72B-Instruct

Expected Behavior:
The summaries should be generated by the Qwen/Qwen2.5-72B-Instruct model, as specified under the summarization key.

Observed Behavior:
Instead, the summaries are being generated by Qwen/Qwen2.5-VL-72B-Instruct.

Investigation Notes:
This seems to be caused by the use of step_name="summarization_chunk" in summarization.py. It appears this step name is not matched with the "summarization" key in the config, and since no explicit model role is defined for "summarization_chunk", the first model listed (i.e., the ingestion model) is used by default. This behavior is consistent with the fallback mechanism in inference_engine.py.