Add mixed precision evaluation (#8036)

* Add mixed precision evaluation

* use original flag

src/transformers/trainer.py

@@ -1466,6 +1466,10 @@ class Trainer:
         inputs = self._prepare_inputs(inputs)
 
         with torch.no_grad():
+            if self.args.fp16 and _use_native_amp:
+                with autocast():
+                    outputs = model(**inputs)
+            else:
                 outputs = model(**inputs)
             if has_labels:
                 loss = outputs[0].mean().detach()