Fix Gradient Checkpointing for Deberta & Deberta-V2 using PEFT / Adapters (#35898)

Replace In-Place Operations for Deberta and Deberta-V2

Author: lenglaender

src/transformers/models/deberta/modeling_deberta.py

@@ -418,10 +418,10 @@ def forward(self, input_ids=None, token_type_ids=None, position_ids=None, mask=N
 
         embeddings = inputs_embeds
         if self.position_biased_input:
-            embeddings += position_embeddings
+            embeddings = embeddings + position_embeddings
         if self.token_type_embeddings is not None:
             token_type_embeddings = self.token_type_embeddings(token_type_ids)
-            embeddings += token_type_embeddings
+            embeddings = embeddings + token_type_embeddings
 
         if self.embed_proj is not None:
             embeddings = self.embed_proj(embeddings)

src/transformers/models/deberta_v2/modeling_deberta_v2.py

@@ -551,10 +551,10 @@ def forward(self, input_ids=None, token_type_ids=None, position_ids=None, mask=N
 
         embeddings = inputs_embeds
         if self.position_biased_input:
-            embeddings += position_embeddings
+            embeddings = embeddings + position_embeddings
         if self.token_type_embeddings is not None:
             token_type_embeddings = self.token_type_embeddings(token_type_ids)
-            embeddings += token_type_embeddings
+            embeddings = embeddings + token_type_embeddings
 
         if self.embed_proj is not None:
             embeddings = self.embed_proj(embeddings)