Dropout

In LoRA, where in the forward pass does `lora_dropout` apply its mask?

Where inside the attention sub-layer is dropout typically applied during training, and what does dropping there teach the model?

BPE-dropout is applied at which stage(s) of the LLM lifecycle?