[Trainer] Fix .to call on 4bit models (#24444)

* fix `.to` call on 4bit models

* better check

src/transformers/trainer.py

@@ -493,7 +493,8 @@ def __init__(
         self.eval_dataset = eval_dataset
         self.tokenizer = tokenizer
 
-        if self.place_model_on_device and not getattr(model, "is_loaded_in_8bit", False):
+        # Quantized models doesn't support `.to` operation.
+        if self.place_model_on_device and not getattr(model, "is_quantized", False):
             self._move_model_to_device(model, args.device)
 
         # Force n_gpu to 1 to avoid DataParallel as MP will manage the GPUs