Drop inplace operation for loss computation with gradient accumulation (

#35416)

Fix inplace loss computation

src/transformers/trainer.py

@@ -3700,7 +3700,7 @@ def training_step(
         else:
             # Finally we need to normalize the loss for reporting
             if num_items_in_batch is None:
-                loss /= self.args.gradient_accumulation_steps
+                loss = loss / self.args.gradient_accumulation_steps
 
             self.accelerator.backward(loss, **kwargs)