[inference] fix stop token for object detection (#6624)

* fix stop token

* update minicpm data pipeline

* fix npu qlora examples

Former-commit-id: 844919fadaa8a61dfae47020971ea80730b2346f

src/llamafactory/train/pt/trainer.py

@@ -79,6 +79,8 @@ class CustomTrainer(Trainer):
     ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
         r"""
         Fixes the loss value. See https://github.com/huggingface/transformers/pull/35438 for details.
+
+        It should be removed after https://github.com/huggingface/transformers/pull/35651 is merged.
         """
         loss = super().compute_loss(model, inputs, return_outputs, **kwargs)
         if kwargs.get("num_items_in_batch") and not getattr(self, "model_accepts_loss_kwargs", False):

src/llamafactory/train/sft/trainer.py

@@ -94,6 +94,8 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
     ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
         r"""
         Fixes the loss value. See https://github.com/huggingface/transformers/pull/35438 for details.
+
+        It should be removed after https://github.com/huggingface/transformers/pull/35651 is merged.
         """
         loss = super().compute_loss(model, inputs, return_outputs, **kwargs)
         if kwargs.get("num_items_in_batch") and not getattr(self, "model_accepts_loss_kwargs", False):