Have you faced the bp incorrect in training

[jerrywind]

Hi all,
I recently take some efforts on training paddleOCR-VL-0.9B , and I found a wired thing ,the gradiants passed from a transform layer to the previous one will lost part of gradiants (Maybe due to flash mask tech) with transformers family.

How to resolve it ? Do you have any idea ?

[jzhang533]

No, never met such issue.

[jerrywind]

Thanks.
However the training performances are not the same. Could you explain a bit why -100 needed ? I'm not very clear about that part.
BR