Oh, I had similar case. First, I used qwen without any ChatML-like syntax and it continued speaking and speaking, then I used <|im_start|>/<|im_end|> to control it somehow