Code clean up for fp8 quantization (#16982)

python/sglang/srt/layers/linear.py

@@ -362,6 +362,7 @@ class ColumnParallelLinear(LinearBase):
 
     def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
         output_dim = getattr(param, "output_dim", None)
+        param_data = param.data
 
         # Special case for GGUF
         is_gguf_weight = getattr(param, "is_gguf_weight", False)
@@ -373,11 +374,9 @@ class ColumnParallelLinear(LinearBase):
         if is_gguf_weight and isinstance(param, UninitializedParameter):
             param.materialize(loaded_weight.shape, dtype=loaded_weight.dtype)
 
-        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
-
-        param_data = param.data
         # bitsandbytes loads the weights of the specific portion
         # no need to narrow here
+        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
         if output_dim is not None and not use_bitsandbytes_4bit:
             shard_size = param_data.shape[output_dim]
             start_idx = self.tp_rank * shard_size