error of use 2 and more GPUS for training FLUX model

[NeuroDonu]

hi!
i use cluster of 4 RTX A6000 Ada for train my LORA. if i use 1 gpu - this work fine. if i use 2 and more gpu - its didnt work. (i use multigpu) screen -->

analogue if i use 4 gpus its didnt work (in 1 gpu i use 2 batch, 2 gpu - 4 batch and more)
so. i didnt und why. Has anyone encountered this? If so, how did you solve it?