Regarding the model weights issue

Hello, I don’t quite understand the logic behind the parameter --vision_tower ${VISION_MODEL_VERSION}.
If I want to train using the full model weights you provided, is the vision tower already included in the existing weights? Does the current training code logic require loading the vision tower weights separately again?