feat(model parallelism): moving the labels to the same device as the logits...

feat(model parallelism): moving the labels to the same device as the logits for gpt2 and bart (#22591)

src/transformers/models/bart/modeling_bart.py

@@ -1398,6 +1398,7 @@ class BartForConditionalGeneration(BartPretrainedModel):
 
         masked_lm_loss = None
         if labels is not None:
+            labels = labels.to(lm_logits.device)
             loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
 
@@ -1553,6 +1554,7 @@ class BartForSequenceClassification(BartPretrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             if self.config.problem_type is None:
                 if self.config.num_labels == 1:
                     self.config.problem_type = "regression"
@@ -1896,6 +1898,7 @@ class BartForCausalLM(BartPretrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -2581,6 +2581,7 @@ class BigBirdPegasusForConditionalGeneration(BigBirdPegasusPreTrainedModel):
 
         masked_lm_loss = None
         if labels is not None:
+            labels = labels.to(lm_logits.device)
             loss_fct = CrossEntropyLoss()
             masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
 
@@ -2735,6 +2736,7 @@ class BigBirdPegasusForSequenceClassification(BigBirdPegasusPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             if self.config.problem_type is None:
                 if self.config.num_labels == 1:
                     self.config.problem_type = "regression"

src/transformers/models/blenderbot/modeling_blenderbot.py

@@ -1596,6 +1596,7 @@ class BlenderbotForCausalLM(BlenderbotPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 

src/transformers/models/blenderbot_small/modeling_blenderbot_small.py

@@ -1563,6 +1563,7 @@ class BlenderbotSmallForCausalLM(BlenderbotSmallPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 

src/transformers/models/gpt2/modeling_gpt2.py

@@ -1098,6 +1098,8 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
 
         loss = None
         if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(lm_logits.device)
             # Shift so that tokens < n predict n
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
@@ -1318,6 +1320,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             mc_loss = loss_fct(mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1))
         lm_loss = None
         if labels is not None:
+            labels = labels.to(lm_logits.device)
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
             loss_fct = CrossEntropyLoss()
@@ -1569,6 +1572,7 @@ class GPT2ForTokenClassification(GPT2PreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
 

src/transformers/models/marian/modeling_marian.py

@@ -1715,6 +1715,7 @@ class MarianForCausalLM(MarianPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 

src/transformers/models/mbart/modeling_mbart.py

@@ -1528,6 +1528,7 @@ class MBartForSequenceClassification(MBartPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             if self.config.problem_type is None:
                 if self.config.num_labels == 1:
                     self.config.problem_type = "regression"
@@ -1866,6 +1867,7 @@ class MBartForCausalLM(MBartPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 

src/transformers/models/pegasus/modeling_pegasus.py

@@ -1694,6 +1694,7 @@ class PegasusForCausalLM(PegasusPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 

src/transformers/models/plbart/modeling_plbart.py

@@ -1499,6 +1499,7 @@ class PLBartForSequenceClassification(PLBartPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             if self.config.problem_type is None:
                 if self.config.num_labels == 1:
                     self.config.problem_type = "regression"
@@ -1713,6 +1714,7 @@ class PLBartForCausalLM(PLBartPreTrainedModel):
 
         loss = None
         if labels is not None:
+            labels = labels.to(logits.device)
             loss_fct = CrossEntropyLoss()
             loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))