fix: load_best_model_at_end error when load_in_8bit is True

    Ref: huggingface/peft#394
    Loading a quantized checkpoint into non-quantized Linear8bitLt is not supported.
    call module.cuda() before module.load_state_dict()

src/transformers/trainer.py

@@ -2235,6 +2235,8 @@ def _load_best_model(self):
                     # If the model is on the GPU, it still works!
                     # workaround for FSDP bug https://github.com/pytorch/pytorch/issues/82963
                     # which takes *args instead of **kwargs
+                    if model._is_int8_training_enabled or model.is_8bit_serializable:
+                        model.cuda()
                     load_result = model.load_state_dict(state_dict, False)
                 if not is_sagemaker_mp_enabled():
                     self._issue_warnings_after_load(load_result)