diff --git a/backend/operations_gguf.py b/backend/operations_gguf.py
index 5a01d1ca..abd52a53 100644
--- a/backend/operations_gguf.py
+++ b/backend/operations_gguf.py
@@ -2,9 +2,46 @@ import gguf
 import torch
 
 
-# def functional_quantize_gguf(weight):
-#     gguf_cls = weight.gguf_cls
-#     gguf_cls.en
+quants_mapping = {
+    gguf.GGMLQuantizationType.Q4_0: gguf.Q4_0,
+    gguf.GGMLQuantizationType.Q5_0: gguf.Q5_0,
+    gguf.GGMLQuantizationType.Q8_0: gguf.Q8_0,
+}
+
+
+class ParameterGGUF(torch.nn.Parameter):
+    def __init__(self, tensor=None, requires_grad=False, no_init=False):
+        super().__init__()
+        self.is_gguf = True
+
+        if no_init:
+            return
+
+        self.gguf_type = tensor.tensor_type
+        self.gguf_real_shape = torch.Size(reversed(list(tensor.shape)))
+        self.gguf_cls = quants_mapping.get(self.gguf_type, None)
+
+    @property
+    def shape(self):
+        return self.gguf_real_shape
+
+    def __new__(cls, tensor=None, requires_grad=False, no_init=False):
+        return super().__new__(cls, torch.tensor(tensor.data), requires_grad=requires_grad)
+
+    def to(self, *args, **kwargs):
+        new = ParameterGGUF(self.data.to(*args, **kwargs), no_init=True)
+        new.gguf_type = self.gguf_type
+        new.gguf_real_shape = self.gguf_real_shape
+        new.gguf_cls = self.gguf_cls
+        return new
+
+    @classmethod
+    def make(cls, data, gguf_type, gguf_cls, gguf_real_shape):
+        new = ParameterGGUF(data, no_init=True)
+        new.gguf_type = gguf_type
+        new.gguf_real_shape = gguf_real_shape
+        new.gguf_cls = gguf_cls
+        return new
 
 
 def functional_linear_gguf(x, weight, bias=None):
diff --git a/backend/utils.py b/backend/utils.py
index 7860eb08..d01863e0 100644
--- a/backend/utils.py
+++ b/backend/utils.py
@@ -4,48 +4,7 @@ import os
 import json
 import safetensors.torch
 import backend.misc.checkpoint_pickle
-
-
-quants_mapping = {
-    gguf.GGMLQuantizationType.Q4_0: gguf.Q4_0,
-    gguf.GGMLQuantizationType.Q5_0: gguf.Q5_0,
-    gguf.GGMLQuantizationType.Q8_0: gguf.Q8_0,
-}
-
-
-class ParameterGGUF(torch.nn.Parameter):
-    def __init__(self, tensor=None, requires_grad=False, no_init=False):
-        super().__init__()
-        self.is_gguf = True
-
-        if no_init:
-            return
-
-        self.gguf_type = tensor.tensor_type
-        self.gguf_real_shape = torch.Size(reversed(list(tensor.shape)))
-        self.gguf_cls = quants_mapping.get(self.gguf_type, None)
-
-    @property
-    def shape(self):
-        return self.gguf_real_shape
-
-    def __new__(cls, tensor=None, requires_grad=False, no_init=False):
-        return super().__new__(cls, torch.tensor(tensor.data), requires_grad=requires_grad)
-
-    def to(self, *args, **kwargs):
-        new = ParameterGGUF(self.data.to(*args, **kwargs), no_init=True)
-        new.gguf_type = self.gguf_type
-        new.gguf_real_shape = self.gguf_real_shape
-        new.gguf_cls = self.gguf_cls
-        return new
-
-    @classmethod
-    def make(cls, data, gguf_type, gguf_cls, gguf_real_shape):
-        new = ParameterGGUF(data, no_init=True)
-        new.gguf_type = gguf_type
-        new.gguf_real_shape = gguf_real_shape
-        new.gguf_cls = gguf_cls
-        return new
+from backend.operations_gguf import ParameterGGUF
 
 
 def read_arbitrary_config(directory):