Safetensors
clip_vision_model