Fix for the issue of device-id getting hardcoded for token_type_ids during Tracing [WIP] (#11252)

* registering a buffer for token_type_ids, to pass the error of device-id getting hardcoded when tracing

* sytle format

* adding persistent flag to the resgitered buffers that prevent from adding them to the state_dict and addresses the Backward compatibility issue

* adding the try catch to the fix as persistent flag is only available from PT >1.6

* adding version check

* added the condition to only use the token_type_ids buffer when its autogenerated not passed by user

* adding comments and making the conidtion where token_type_ids are None to use the registered buffer

* taking out position-embeddding from the if block

* adding comments

* handling the case if buffer for position_ids was not registered

* reverted the changes on position_ids, fix the issue with size of token_type_ids buffer, moved the modification for generated token_type_ids to Bertmodel, instead of Embeddings

* reverting the token_type_ids in case of None to the previous version

* reverting changes on position_ids adding back the if block

* changes added by running make fix-copies

* changes added by running make fix-copies and added the import version as it was getting used

* changes added by running make fix-copies

* changes added by running make fix-copies

* fixing the import format

* fixing the import format

* modified to use temp tensor for trimed and expanded token_type_ids buffer

* changes made by fix-copies after temp tensor modifications

* changes made by fix-copies after temp tensor modifications

* changes made by fix-copies after temp tensor modifications

* clean up

* clean up

* clean up

* clean up

* Nit

* Nit

* Nit

* modified according to support device conversion on traced models

* modified according to support device conversion on traced models

* modified according to support device conversion on traced models

* modified according to support device conversion on traced models

* changes based on latest in master

* Adapt templates

* Add version import

Co-authored-by: Ubuntu <ubuntu@ip-172-31-32-81.us-west-2.compute.internal>
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py

@@ -22,6 +22,7 @@ import os
 
 import torch
 import torch.utils.checkpoint
+from packaging import version
 from torch import nn
 from torch.nn import CrossEntropyLoss, MSELoss
 
@@ -156,6 +157,12 @@ class {{cookiecutter.camelcase_modelname}}Embeddings(nn.Module):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if version.parse(torch.__version__) > version.parse("1.6.0"):
+            self.register_buffer(
+                "token_type_ids",
+                torch.zeros(self.position_ids.size(), dtype=torch.long, device=self.position_ids.device),
+                persistent=False,
+            )
 
     def forward(
         self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
@@ -170,9 +177,17 @@ class {{cookiecutter.camelcase_modelname}}Embeddings(nn.Module):
         if position_ids is None:
             position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
 
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
         if token_type_ids is None:
-            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
-
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+                
         if inputs_embeds is None:
             inputs_embeds = self.word_embeddings(input_ids)
         token_type_embeddings = self.token_type_embeddings(token_type_ids)
@@ -846,8 +861,14 @@ class {{cookiecutter.camelcase_modelname}}Model({{cookiecutter.camelcase_modelna
 
         if attention_mask is None:
             attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
         if token_type_ids is None:
-            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
 
         # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
         # ourselves in which case we just need to make it broadcastable to all heads.