delete CxB from state

bitsandbytes/nn/modules.py

@@ -260,11 +260,10 @@ class Linear8bitLt(nn.Linear):
 
         out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
 
-        # if not self.state.has_fp16_weights and self.state.CB is not None:
-            # we converted 8-bit row major to turing/ampere format in the first inference pass
-            # we no longer need the row-major weight
-            # del self.state.CB
-            # self.weight.data = self.state.CxB
+        if not self.state.has_fp16_weights and self.state.CxB is not None:
+            # In this version, we convert 8-bit row major to turing/ampere format at each inference pass
+            # Thus, we delete CxB from the state. TODO: do not store it in the state in the first place. 
+            del self.state.CxB
 
         return out