Adding `safetensors` variant of this model

all layer trained for every step.AdaptiveLayerLoss(model=model,

initial commit