SD3.5 timestep shift

I am currently training the SD3.5 model (LORA) based on SimpleTuner framework. The size of my image is 512*512. During the training process, I enabled the flux_schedule_auto_shift. I found that this would affect the sigma calculation of the noise scheduler. However, I have found that during the validation generation phase of the training process, the original noise scheduler was used, which means that the shift was not adaptively changed according to the image resolution. This results in a difference between the sigma used during training and the sigma used during validation (the step() function of the noise scheduler in the sampling process requires sigma)? I suspect this will lead to different results in the generation process. Furthermore, for the SD3.5 model, the image sizes during training and inference are both 512*512. So, what is the recommended value for the shift parameter? I'm not sure. I sincerely hope to receive your assistance or advice. Thank you!