It seems you train LMM due to "lmm lora" in the inference code, different from your paper. How much does training lmm influence?