possible linear module fix for shared embeddings

Author: galqiwi

bitsandbytes/nn/modules.py

@@ -992,12 +992,10 @@ def _load_from_state_dict(
     def init_8bit_state(self):
         self.state.CB = self.weight.CB
         self.state.SCB = self.weight.SCB
-        self.weight.CB = None
-        self.weight.SCB = None
 
     def forward(self, x: torch.Tensor):
         self.state.is_training = self.training
-        if self.weight.CB is not None:
+        if self.state.CB is None:
             self.init_8bit_state()
 
         # weights are cast automatically as Int8Params, but the bias has to be cast manually
@@ -1069,13 +1067,11 @@ def __init__(
     def init_8bit_state(self):
         self.state.CB = self.weight.CB
         self.state.SCB = self.weight.SCB
-        self.weight.CB = None
-        self.weight.SCB = None
 
     def forward(self, x):
         self.state.is_training = self.training
 
-        if self.weight.CB is not None:
+        if self.state.CB is None:
             self.init_8bit_state()
 
         out = bnb.matmul_mixed(x.half(), self.weight.half(), bias=None, state=self.state) + self.bias