fix autocast for older pytorch

src/transformers/trainer.py

@@ -1857,8 +1857,12 @@ class Trainer:
             return loss_mb.reduce_mean().detach().to(self.args.device)
 
         if self.use_amp:
+            if version.parse(torch.__version__) >= version.parse("1.10"):
                 with autocast(dtype=self.amp_dtype):
                     loss = self.compute_loss(model, inputs)
+            else:
+                with autocast():
+                    loss = self.compute_loss(model, inputs)
         else:
             loss = self.compute_loss(model, inputs)
 
@@ -2501,8 +2505,12 @@ class Trainer:
             else:
                 if has_labels:
                     if self.use_amp:
+                        if version.parse(torch.__version__) >= version.parse("1.10"):
                             with autocast(dtype=self.amp_dtype):
                                 loss, outputs = self.compute_loss(model, inputs, return_outputs=True)
+                        else:
+                            with autocast():
+                                loss, outputs = self.compute_loss(model, inputs, return_outputs=True)
                     else:
                         loss, outputs = self.compute_loss(model, inputs, return_outputs=True)
                     loss = loss.mean().detach()
@@ -2514,8 +2522,12 @@ class Trainer:
                 else:
                     loss = None
                     if self.use_amp:
+                        if version.parse(torch.__version__) >= version.parse("1.10"):
                             with autocast(dtype=self.amp_dtype):
                                 outputs = model(**inputs)
+                        else:
+                            with autocast():
+                                outputs = model(**inputs)
                     else:
                         outputs = model(**inputs)
                     if isinstance(outputs, dict):