diff --git a/dequant.py b/dequant.py
index 8c38d30..8ab22ad 100644
--- a/dequant.py
+++ b/dequant.py
@@ -4,9 +4,6 @@ import torch
 import numpy as np
 
 def dequantize_tensor(tensor, dtype=torch.float16):
-    if tensor is None:
-        return None
-
     data = torch.tensor(tensor.data)
     qtype = tensor.tensor_type
     oshape = tensor.tensor_shape
diff --git a/nodes.py b/nodes.py
index 3ce670c..8d63cae 100644
--- a/nodes.py
+++ b/nodes.py
@@ -1,6 +1,8 @@
 # (c) City96 || Apache-2.0 (apache.org/licenses/LICENSE-2.0)
+import torch
 import gguf
 import logging
+import numpy as np
 
 import comfy.sd
 import comfy.utils
@@ -20,7 +22,13 @@ def gguf_sd_loader(path):
     sd = {}
     dt = {}
     for tensor in reader.tensors:
-        sd[str(tensor.name)] = GGMLTensor(tensor)
+        sd[str(tensor.name)] = GGMLTensor(
+            torch.from_numpy(tensor.data), # mmap
+            tensor_type = tensor.tensor_type,
+            tensor_shape = torch.Size(
+                np.flip(list(tensor.shape))
+            )
+        )
         dt[str(tensor.tensor_type)] = dt.get(str(tensor.tensor_type), 0) + 1
 
     # sanity check debug print
diff --git a/ops.py b/ops.py
index 83aa657..95981b6 100644
--- a/ops.py
+++ b/ops.py
@@ -10,16 +10,13 @@ class GGMLTensor(torch.Tensor):
     """
     Main tensor-like class for storing quantized weights
     """
-    def __init__(self, tensor, *args, **kwargs):
+    def __init__(self, *args, tensor_type, tensor_shape, patches=[], **kwargs):
         super().__init__()
-        self.tensor_type = tensor.tensor_type
-        self.tensor_shape = torch.Size(
-            np.flip(list(tensor.shape))
-        )
+        self.tensor_type = tensor_type
+        self.tensor_shape = tensor_shape
 
-    def __new__(cls, tensor, *args, **kwargs):
-        data = torch.from_numpy(tensor.data)
-        return super().__new__(cls, data, *args, **kwargs)
+    def __new__(cls, *args, tensor_type, tensor_shape, patches=[], **kwargs):
+        return super().__new__(cls, *args, **kwargs)
 
     def to(self, *args, **kwargs):
         new = super().to(*args, **kwargs)
@@ -74,9 +71,15 @@ class GGMLLayer(torch.nn.Module):
         super()._apply(fn)
         return self
 
+    def get_weight(self, tensor, dtype):
+        if tensor is None:
+            return
+        weight = dequantize_tensor(tensor, dtype)
+        return weight
+
     def get_weights(self, dtype=torch.float16):
-        weight = dequantize_tensor(self.weight, dtype)
-        bias = dequantize_tensor(self.bias, dtype)
+        weight = self.get_weight(self.weight, dtype)
+        bias = self.get_weight(self.bias, dtype)
         return (weight, bias)
 
 class GGMLOps(comfy.ops.disable_weight_init):