fix hooked linear init

gptqmodel/nn_modules/hooked_linear.py

@@ -2,8 +2,12 @@
 
 
 class HookedLinear(torch.nn.Linear):
-    def __init__(self, in_features: int, out_features: int, bias: bool = True, device=None, dtype=None) -> None:
-        super().__init__(in_features, out_features, bias, device, dtype)
+    def __init__(self, in_features: int, out_features: int) -> None:
+        # avoid calling super().__init__() as it would allocate memory based on in/out features
+        torch.nn.Module.__init__(self)
+        self.in_features = in_features
+        self.out_features = out_features
+
         self.forward_hook = None
 
     @staticmethod
@@ -16,15 +20,13 @@ def replace_linear_with_hooked_linear(module):
 
     @staticmethod
     def from_linear(linear: torch.nn.Linear):
-        custom_linear = HookedLinear(linear.in_features, linear.out_features, bias=linear.bias is not None,
-                                     device=linear.weight.device, dtype=linear.weight.dtype)
+        custom_linear = HookedLinear(linear.in_features, linear.out_features)
         custom_linear.weight = linear.weight
-        if linear.bias is not None:
-            custom_linear.bias = linear.bias
+        custom_linear.bias = linear.bias
         return custom_linear
 
     def forward(self, input: torch.Tensor) -> torch.Tensor:
         output = super().forward(input)
         if self.forward_hook:
             self.forward_hook(self, input, output)
-        return output
+        return output