Update main.py

main.py

@@ -69,23 +69,7 @@ async def chat(request: ChatCompletionRequest):
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
 
-    async def format_response(chat_chunks: Generator) -> Any:
-        for chat_chunk in chat_chunks:
-            response = {
-                'choices': [
-                    {
-                        'message': {
-                            'role': 'system',
-                            'content': llm.detokenize(chat_chunk)
-                        },
-                        'finish_reason': 'stop' if llm.detokenize(chat_chunk) == "[DONE]" else 'unknown'
-                    }
-                ]
-            }
-            yield f"data: {json.dumps(response)}\n\n"
-        yield "event: done\ndata: {}\n\n"
-
-    return StreamingResponse(format_response(chat_chunks), media_type="text/event-stream")
+    return StreamingResponse(generate_response(chat_chunks, llm), media_type="text/event-stream")
 
 async def stream_response(tokens: Any) -> None:
     try: