Fix race condition in LoRA streaming requests

[Lokiiiiii]

The Bug

When using vLLM async service with LoRA adapters, streaming requests were not receiving the lora_request parameter. This was due to an async generator timing issue in the original implementation.

Original Implementation (Buggy)

if processed_request.lora_request:
    original_add_request = self.vllm_engine.add_request
    
    async def add_request_with_lora(*args, **kwargs):
        kwargs['lora_request'] = processed_request.lora_request
        return await original_add_request(*args, **kwargs)
    
    self.vllm_engine.add_request = add_request_with_lora
    try:
        response = await processed_request.inference_invoker(...)
    finally:
        self.vllm_engine.add_request = original_add_request

Problem: For streaming requests, inference_invoker returns an AsyncGenerator that hasn't started executing yet. The wrapper is removed in the finally block before the generator actually runs and calls add_request.

The Fix

The fix passes adapter directly as a parameter to the OpenAI Request object and then send it to inference invoker instead of wrapping the method:

processed_request.vllm_request.model = adapter_name
response = await processed_request.inference_invoker(
            processed_request.vllm_request

This ensures the lora_request is properly passed to vLLM's OpenAI serving methods for both streaming and non-streaming requests.

Testing

Manual testing confirms adapter is getting utilized

wait for tests

[Lokiiiiii]

Manual testing confirms adapter is getting utilized