fix overflow when training mDeberta in fp16 (#24116)

* Porting changes from https://github.com/microsoft/DeBERTa/ that hopefully allows for fp16 training of mdeberta

* Updates to deberta modeling from microsoft repo

* Performing some cleanup

* Undoing changes that weren't necessary

* Undoing float calls

* Minimally change the p2c block

* Fix error

* Minimally changing the c2p block

* Switch to torch sqrt

* Remove math

* Adding back the to calls to scale

* Undoing attention_scores change

* Removing commented out code

* Updating modeling_sew_d.py to satisfy utils/check_copies.py

* Missed changed

* Further reduce changes needed to get fp16 working

* Reverting changes to modeling_sew_d.py

* Make same change in TF

src/transformers/models/deberta_v2/modeling_deberta_v2.py

@@ -721,7 +721,7 @@ class DisentangledSelfAttention(nn.Module):
         if "p2c" in self.pos_att_type:
             scale_factor += 1
         scale = torch.sqrt(torch.tensor(query_layer.size(-1), dtype=torch.float) * scale_factor)
-        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2)) / scale.to(dtype=query_layer.dtype)
+        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2) / scale.to(dtype=query_layer.dtype))
         if self.relative_attention:
             rel_embeddings = self.pos_dropout(rel_embeddings)
             rel_att = self.disentangled_attention_bias(

src/transformers/models/deberta_v2/modeling_tf_deberta_v2.py

@@ -676,7 +676,7 @@ class TFDebertaV2DisentangledSelfAttention(tf.keras.layers.Layer):
         if "p2c" in self.pos_att_type:
             scale_factor += 1
         scale = tf.math.sqrt(tf.cast(shape_list(query_layer)[-1] * scale_factor, tf.float32))
-        attention_scores = tf.matmul(query_layer, tf.transpose(key_layer, [0, 2, 1])) / scale
+        attention_scores = tf.matmul(query_layer, tf.transpose(key_layer, [0, 2, 1]) / scale)
         if self.relative_attention:
             rel_embeddings = self.pos_dropout(rel_embeddings)
             rel_att = self.disentangled_att_bias(query_layer, key_layer, relative_pos, rel_embeddings, scale_factor)

src/transformers/models/sew_d/modeling_sew_d.py

@@ -789,7 +789,7 @@ class DisentangledSelfAttention(nn.Module):
         if "p2c" in self.pos_att_type:
             scale_factor += 1
         scale = torch.sqrt(torch.tensor(query_layer.size(-1), dtype=torch.float) * scale_factor)
-        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2)) / scale.to(dtype=query_layer.dtype)
+        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2) / scale.to(dtype=query_layer.dtype))
         if self.relative_attention:
             rel_embeddings = self.pos_dropout(rel_embeddings)
             rel_att = self.disentangled_attention_bias(