change inference.py class StoppingCriteriaSub(StoppingCriteria) to fix the error

inference.py

@@ -176,11 +176,16 @@ class StoppingCriteriaSub(StoppingCriteria):
 
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs):
         for stop in self.stops:
+            if input_ids.shape[-1] < len(stop):
+                continue
+            # 把 stop 移动到 input_ids 的设备上
+            stop = stop.to(input_ids.device)
             if torch.all((stop == input_ids[0][-len(stop):])).item():
                 return True
-
         return False
 
+
+
 @torch.inference_mode()
 def generate_stream(
         model, tokenizer, image_processor, params, device