Fix training of pipeline based peft's lora model

Fix training of pipeline based peft's lora model #927