According to the docmuents, it says:
“Enables the gradients for the input embeddings. This is useful for fine-tuning adapter weights while keeping the model weights fixed.”
But why it is useful in this situation? Does it mean it will enable the embedding_layer’s weights been updated during training?