diff --git a/python/llm/src/bigdl/llm/transformers/model.py b/python/llm/src/bigdl/llm/transformers/model.py
index 2c86c83b..70a353a6 100644
--- a/python/llm/src/bigdl/llm/transformers/model.py
+++ b/python/llm/src/bigdl/llm/transformers/model.py
@@ -65,6 +65,8 @@ class _BaseAutoModelClass:
         if load_in_4bit or load_in_low_bit:
             # load int x-bit
             kwargs["low_cpu_mem_usage"] = True
+            # set default torch_dtype='auto'
+            kwargs["torch_dtype"] = kwargs.get("torch_dtype", 'auto')
             q_k = load_in_low_bit if load_in_low_bit else "sym_int4"
             model = cls.load_convert(q_k, *args, **kwargs)
         else: