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/albert/modeling_albert.py

@@ -1230,8 +1230,8 @@ class AlbertForQuestionAnswering(AlbertPreTrainedModel):
                 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/bart/modeling_bart.py

@@ -1578,8 +1578,8 @@ class BartForQuestionAnswering(BartPretrainedModel):
                 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/bert/modeling_bert.py

@@ -1813,8 +1813,8 @@ class BertForQuestionAnswering(BertPreTrainedModel):
                 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/big_bird/modeling_big_bird.py

@@ -2995,8 +2995,8 @@ class BigBirdForQuestionAnswering(BigBirdPreTrainedModel):
                 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/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -2783,8 +2783,8 @@ class BigBirdPegasusForQuestionAnswering(BigBirdPegasusPreTrainedModel):
                 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/convbert/modeling_convbert.py

@@ -1305,8 +1305,8 @@ class ConvBertForQuestionAnswering(ConvBertPreTrainedModel):
                 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/deberta/modeling_deberta.py

@@ -1376,8 +1376,8 @@ class DebertaForQuestionAnswering(DebertaPreTrainedModel):
                 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/deberta_v2/modeling_deberta_v2.py

@@ -1500,8 +1500,8 @@ class DebertaV2ForQuestionAnswering(DebertaV2PreTrainedModel):
                 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/distilbert/modeling_distilbert.py

@@ -740,8 +740,8 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
                 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 = nn.CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)

src/transformers/models/electra/modeling_electra.py

@@ -1330,8 +1330,8 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
                 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/funnel/modeling_funnel.py

@@ -1561,8 +1561,8 @@ class FunnelForQuestionAnswering(FunnelPreTrainedModel):
                 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/ibert/modeling_ibert.py

@@ -1331,8 +1331,8 @@ class IBertForQuestionAnswering(IBertPreTrainedModel):
                 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/led/modeling_led.py

@@ -2607,8 +2607,8 @@ class LEDForQuestionAnswering(LEDPreTrainedModel):
                 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/longformer/modeling_longformer.py

@@ -2029,8 +2029,8 @@ class LongformerForQuestionAnswering(LongformerPreTrainedModel):
                 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/mbart/modeling_mbart.py

@@ -1585,8 +1585,8 @@ class MBartForQuestionAnswering(MBartPreTrainedModel):
                 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/megatron_bert/modeling_megatron_bert.py

@@ -1806,8 +1806,8 @@ class MegatronBertForQuestionAnswering(MegatronBertPreTrainedModel):
                 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/mobilebert/modeling_mobilebert.py

@@ -1383,8 +1383,8 @@ class MobileBertForQuestionAnswering(MobileBertPreTrainedModel):
                 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/mpnet/modeling_mpnet.py

@@ -1035,8 +1035,8 @@ class MPNetForQuestionAnswering(MPNetPreTrainedModel):
                 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/reformer/modeling_reformer.py

@@ -2567,8 +2567,8 @@ class ReformerForQuestionAnswering(ReformerPreTrainedModel):
                 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/roberta/modeling_roberta.py

@@ -1484,8 +1484,8 @@ class RobertaForQuestionAnswering(RobertaPreTrainedModel):
                 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)