FIX / OPT: Fix OPT multi-GPU training for `OPTForQuestionAnswering` (#31092)

Update modeling_opt.py

src/transformers/models/opt/modeling_opt.py

@@ -1427,8 +1427,8 @@ class OPTForQuestionAnswering(OPTPreTrainedModel):
                 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 = start_positions.clamp(0, ignored_index)
-            end_positions = end_positions.clamp(0, ignored_index)
+            start_positions = start_positions.clamp(0, ignored_index).to(logits.device)
+            end_positions = end_positions.clamp(0, ignored_index).to(logits.device)
 
             loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
             start_loss = loss_fct(start_logits, start_positions)