ModelOption.THINKING fails sometimes

[HendrikStrobelt]

OpenAIBackend by default assumes open ai semantic. e.g. THINKING = medium. However, a user may use OpenAIBackend with litellm proxy and so on. In my case, I am using litellm proxy to access granite-4 family on rits cluster. AFAIK, granite does not set thinking effort as medium instead it uses "thinking=True"

        thinking = model_opts.get(ModelOption.THINKING, None)
        if type(thinking) is bool and thinking:
            # OpenAI uses strings for its reasoning levels.
            thinking = "medium"

following throws an error TypeError: openai.resources.chat.completions.completions.Completions.create() got multiple values for keyword argument 'reasoning_effort'

                    backend = OpenAIBackend(
                        base_url="http://0.0.0.0:4000/",
                        model_id=model,
                        api_key="API KEY",
                        header=True,
                        model_options={
                            ModelOption.MAX_NEW_TOKENS: 4096,
                            ModelOption.SYSTEM_PROMPT: "You are an expert Site Reliability Engineer.",
                            ModelOption.THINKING: True
                            }
...```

credit: Saurabh Jha

[nrfulton]

this has been solved by thinking-effort model option management code.