When I finetune CogVideoX-5B-I2V in SFT, I find it use class free guidance for text condition in inference phase(default 5), but there is no text condition dropout during training phase in finetune code, may I ask is it a normal behaviour?