Add `VLLMProvider` for use with `OpenAI{Chat,Responses}Model`

[DouweM]

Description

This has come up a few times now:

• openai_supports_tool_choice_required=False is still setting tool_choice to required for Responses #3270
• Incompatibility with vLLM Responses API #3245
• Unlock VideoUrl for vLLM Models #3306
• Some models ran through OpenAI-compatible APIs need JSON schema in instructions even when native structured output is used #3490

It should:

• handle gpt-oss with harmony_model_profile
• enable the openai_responses_requires_function_call_status_none model profile setting
• enable the new model profile setting from Some models ran through OpenAI-compatible APIs need JSON schema in instructions even when native structured output is used #3490

References

No response

[Anas20001]

Hi @DouweM i am interested in implementing this new method as a start. can I help with it ?

[zhcn000000]

In vllm 0.11.2, reasoning model will send both reasoning and reasoning_content field with the same throught content,and cause OpenAIChatModel print duplicate ThinkingPart from reasoning and reasoning_content