Matching output with shape in T5 model

I’m trying to fine-tune T5 (ForConditionalGeneration) model with local files (added 3d attention mask in forward function of T5ForConditionalGeneration), and got error when the model evaluates,

File “/home/paulc/customized_t5/models/custom_amr_to_sg/t5_model/”, line 568, in forward
scores += position_bias
RuntimeError: output with shape [32, 12, 1, 2] doesn’t match the broadcast shape [32, 12, 2, 2]

It does’t give an error while training though…
I tried with the solution from the link which is similar to T5,
but it still gives me the same error…

Might evaluate function be the problem?

Thank you in advance!