Fix bug for peft compatibility

[YuliangXiu]

With this quick fix, break-a-scene could be successfully trained with peft approaches, such as LoRA and BOFT, without following errors:

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [10, 4096, 77]], which is output 0 of SoftmaxBackward0, is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!