diff --git a/modelscope/pipelines/multi_modal/ovis_vl_pipeline.py b/modelscope/pipelines/multi_modal/ovis_vl_pipeline.py
index eeb935d1..f19eddff 100644
--- a/modelscope/pipelines/multi_modal/ovis_vl_pipeline.py
+++ b/modelscope/pipelines/multi_modal/ovis_vl_pipeline.py
@@ -32,11 +32,13 @@ class VisionChatPipeline(VisualQuestionAnsweringPipeline):
         self._auto_collate = auto_collate
 
         # ovis
+        torch_dtype = kwargs.get('torch_dtype', torch.float16)
+        multimodal_max_length = kwargs.get('multimodal_max_length', 8192)
         self.device = 'cuda' if device == 'gpu' else device
         self.model = AutoModelForCausalLM.from_pretrained(
             model,
-            torch_dtype=torch.bfloat16,
-            multimodal_max_length=8192,
+            torch_dtype=torch_dtype,
+            multimodal_max_length=multimodal_max_length,
             trust_remote_code=True).to(self.device)
         self.text_tokenizer = self.model.get_text_tokenizer()
         self.visual_tokenizer = self.model.get_visual_tokenizer()