Convert observer input to float data type if it is not (#1529) (#1564)

Author: Xia-Weiwen

intel_extension_for_pytorch/quantization/_quantization_state.py

@@ -231,6 +231,10 @@ def _maybe_observe(arg, tensor_info):
             # TODO: do not run this twice on input and output
             if str(tensor_id) in self.tensor_id_to_observer:
                 observer = self.tensor_id_to_observer[str(tensor_id)]
+                if isinstance(arg, torch.Tensor) and arg.dtype != torch.float32:
+                    dtype = arg.dtype
+                    out = observer(arg.float())
+                    return out.to(dtype)
                 return observer(arg)
             else:
                 return arg
@@ -290,7 +294,7 @@ def _observer_output(output, tensor_info):
             tensor_id = tensor_info.id
             if str(tensor_id) in self.tensor_id_to_observer:
                 obs = self.tensor_id_to_observer[str(tensor_id)]
-                obs(output)
+                obs(output.float())
         if isinstance(outputs, torch.Tensor):
             tensor_info = seen_q_op_info.output_tensor_infos[0]
             _observer_output(outputs, tensor_info)