It seems that "-1" should not be used.
I checked the "start" num and compared to Pi05-pytorch code,if “-1” is deleted, the start num is equal to the start number of position_ids in Pi05-pytorch。
This problem causes the result that the error increases with steps increasing.
The problem may originate from that the pytorch model uses "-1"
It seems that "-1" should not be used.
I checked the "start" num and compared to Pi05-pytorch code,if “-1” is deleted, the start num is equal to the start number of position_ids in Pi05-pytorch。
This problem causes the result that the error increases with steps increasing.
The problem may originate from that the pytorch model uses "-1"