Fixes bug that appears when using QA bert and distilation. (#12026)

* Fixing bug that appears when using distilation (and potentially other uses).
During backward pass Pytorch complains with:
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation
This happens because the QA model code modifies the start_positions and end_positions input tensors, using clamp_ function: as a consequence the teacher and the student both modifies the inputs, and backward pass fails.

* Fixing all models QA clamp_ bug.

src/transformers/models/roformer/modeling_roformer.py

@@ -1554,8 +1554,8 @@ class RoFormerForQuestionAnswering(RoFormerPreTrainedModel):
                 end_positions = end_positions.squeeze(-1)
             # sometimes the start/end positions are outside our model inputs, we ignore these terms
             ignored_index = start_logits.size(1)
-            start_positions.clamp_(0, ignored_index)
-            end_positions.clamp_(0, ignored_index)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
 
             loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)

src/transformers/models/squeezebert/modeling_squeezebert.py

@@ -1080,8 +1080,8 @@ class SqueezeBertForQuestionAnswering(SqueezeBertPreTrainedModel):
                 end_positions = end_positions.squeeze(-1)
             # sometimes the start/end positions are outside our model inputs, we ignore these terms
             ignored_index = start_logits.size(1)
-            start_positions.clamp_(0, ignored_index)
-            end_positions.clamp_(0, ignored_index)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
 
             loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)

src/transformers/models/xlm/modeling_xlm.py

@@ -953,8 +953,8 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
                 end_positions = end_positions.squeeze(-1)
             # sometimes the start/end positions are outside our model inputs, we ignore these terms
             ignored_index = start_logits.size(1)
-            start_positions.clamp_(0, ignored_index)
-            end_positions.clamp_(0, ignored_index)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
 
             loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)

src/transformers/models/xlnet/modeling_xlnet.py

@@ -1874,8 +1874,8 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
                 end_positions = end_positions.squeeze(-1)
             # sometimes the start/end positions are outside our model inputs, we ignore these terms
             ignored_index = start_logits.size(1)
-            start_positions.clamp_(0, ignored_index)
-            end_positions.clamp_(0, ignored_index)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
 
             loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)

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

@@ -1516,8 +1516,8 @@ class {{cookiecutter.camelcase_modelname}}ForQuestionAnswering({{cookiecutter.ca
                 end_positions = end_positions.squeeze(-1)
             # sometimes the start/end positions are outside our model inputs, we ignore these terms
             ignored_index = start_logits.size(1)
-            start_positions.clamp_(0, ignored_index)
-            end_positions.clamp_(0, ignored_index)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
 
             loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)
@@ -3066,8 +3066,8 @@ class {{cookiecutter.camelcase_modelname}}ForQuestionAnswering({{cookiecutter.ca
                 end_positions = end_positions.squeeze(-1)
             # sometimes the start/end positions are outside our model inputs, we ignore these terms
             ignored_index = start_logits.size(1)
-            start_positions.clamp_(0, ignored_index)
-            end_positions.clamp_(0, ignored_index)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
 
             loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)