Allow W8A8Linear to accept dtype during initialization instead of hardcoding torch.float16

Author: 43758726

lmdeploy/pytorch/nn/linear/w8a8.py

@@ -25,7 +25,7 @@ def __init__(self,
                  all_reduce: bool = True,
                  quant_dtype: torch.dtype | None = torch.int8,
                  layer_type: str = 'attn'):
-        super().__init__(dtype=torch.float16,
+        super().__init__(dtype=dtype,
                          device=device,
                          colwise=colwise,
                          is_tp=is_tp,