perf: Optimize VLLM OpenAI integration vLLM model validation result parsing

Author: fit2cloud-chenyw

backend/apps/ai_model/model_factory.py

@@ -66,13 +66,14 @@ def _init_llm(self) -> VLLMOpenAI:
             openai_api_key=self.config.api_key or 'Empty',
             openai_api_base=self.config.api_base_url,
             model_name=self.config.model_name,
+            streaming=True,
             **self.config.additional_params,
         )
 class OpenAILLM(BaseLLM):
     def _init_llm(self) -> BaseChatModel:
         return BaseChatOpenAI(
             model=self.config.model_name,
-            api_key=self.config.api_key,
+            api_key=self.config.api_key or 'Empty',
             base_url=self.config.api_base_url,
             stream_usage=True,
             **self.config.additional_params,

backend/apps/system/api/aimodel.py

@@ -28,7 +28,10 @@ async def generate():
             )
             llm_instance = LLMFactory.create_llm(config)
             async for chunk in llm_instance.llm.astream("1+1=?"):
-                if chunk and chunk.content:
+                SQLBotLogUtil.info(chunk)
+                if chunk and isinstance(chunk, str):
+                    yield json.dumps({"content": chunk}) + "\n"
+                if chunk and isinstance(chunk, dict) and chunk.content:
                     yield json.dumps({"content": chunk.content}) + "\n"
 
         except Exception as e: