diff --git a/modeling/inference_models/generic_hf_torch/class.py b/modeling/inference_models/generic_hf_torch/class.py
index 9a59650e..fe2308ac 100644
--- a/modeling/inference_models/generic_hf_torch/class.py
+++ b/modeling/inference_models/generic_hf_torch/class.py
@@ -78,7 +78,6 @@ class model_backend(HFTorchInferenceModel):
         }
         
         if self.use_4_bit:
-            self.lazy_load = False
             tf_kwargs.update({
                 "quantization_config":BitsAndBytesConfig(
                     load_in_4bit=True,
diff --git a/modeling/patches.py b/modeling/patches.py
index 8ffc87a3..6e2168f2 100644
--- a/modeling/patches.py
+++ b/modeling/patches.py
@@ -181,7 +181,7 @@ class LazyloadPatches:
         is_quantized = is_quantized or load_in_8bit
 
         if is_quantized:
-            from .utils.bitsandbytes import set_module_8bit_tensor_to_device
+            from transformers.utils.bitsandbytes import set_module_quantized_tensor_to_device
 
         error_msgs = []
 
@@ -299,7 +299,7 @@ class LazyloadPatches:
                     fp16_statistics = None
 
                 if "SCB" not in param_name:
-                    set_module_8bit_tensor_to_device(
+                    set_module_quantized_tensor_to_device(
                         model,
                         param_name,
                         param_device,