[Model] Refactor modernbert with the attention interface

src/transformers/models/modernbert/configuration_modernbert.py

@@ -110,9 +110,6 @@ class ModernBertConfig(PreTrainedConfig):
             the model will be compiled if 1) `triton` is installed, 2) the model is not on MPS, 3) the model is not
             shared between devices, and 4) the model is not resized after initialization. If `True`, then the model may
             be faster in some scenarios.
-        repad_logits_with_grad (`bool`, *optional*, defaults to `False`):
-            When True, ModernBertForMaskedLM keeps track of the logits' gradient when repadding for output. This only
-            applies when using Flash Attention 2 with passed labels. Otherwise output logits always have a gradient.
 
     Examples:
 
@@ -168,7 +165,6 @@ def __init__(
         sparse_prediction: Optional[bool] = False,
         sparse_pred_ignore_index: Optional[int] = -100,
         reference_compile: Optional[bool] = None,
-        repad_logits_with_grad: Optional[bool] = False,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -197,7 +193,6 @@ def __init__(
         self.sparse_prediction = sparse_prediction
         self.sparse_pred_ignore_index = sparse_pred_ignore_index
         self.reference_compile = reference_compile
-        self.repad_logits_with_grad = repad_logits_with_grad
 
         if self.classifier_pooling not in ["cls", "mean"]:
             raise ValueError(