fix #4742

src/llamafactory/chat/vllm_engine.py

@@ -135,7 +135,7 @@ class VllmEngine(BaseEngine):
             if is_vllm_version_greater_than_0_5_1():
                 multi_modal_data = {"image": pixel_values}
             elif is_vllm_version_greater_than_0_5():
-                multi_modal_data = ImagePixelData(image=pixel_values)   
+                multi_modal_data = ImagePixelData(image=pixel_values)
             else:  # TODO: remove vllm 0.4.3 support
                 multi_modal_data = MultiModalData(type=MultiModalData.Type.IMAGE, data=pixel_values)
         else:

src/llamafactory/train/dpo/trainer.py

@@ -54,6 +54,7 @@ class CustomDPOTrainer(DPOTrainer):
                 disable_dropout_in_model(ref_model)
 
         self.finetuning_args = finetuning_args
+        self.f_divergence_type = "reverse_kl"
         self.reference_free = False
         self.use_dpo_data_collator = True  # hack to avoid warning
         self.generate_during_eval = False  # disable at evaluation