In LLMs, is it true that long prompts require more encoding time than decoding?

Yes, in LLMs, it is true that long prompts require more encoding time than decoding. This is because the encoder is designed to learn embeddings for various predictive modeling tasks[1], while the decoder is designed to generate new texts[1] based on the encoded information. Therefore, the process of encoding long inputs can be more time-consuming than decoding them.