fix(pt): finetuning property/dipole/polar/dos fitting with multi-dimensional data causes error

deepmd/pt/train/training.py

@@ -489,6 +489,26 @@ def collect_single_finetune_params(
                                 _new_state_dict[item_key] = (
                                     _random_state_dict[item_key].clone().detach()
                                 )
+                            elif _new_fitting and (
+                                (".out_bias" in item_key) or (".out_std" in item_key)
+                            ):
+                                new_key = item_key.replace(
+                                    f".{_model_key}.", f".{_model_key_from}."
+                                )
+                                if (
+                                    _random_state_dict[item_key].shape[-1]
+                                    != _origin_state_dict[new_key].shape[-1]
+                                ):
+                                    assert (
+                                        _random_state_dict[item_key].shape[:-1]
+                                        == _origin_state_dict[new_key].shape[:-1]
+                                    )
+                                    _origin_state_dict[new_key] = _origin_state_dict[
+                                        new_key
+                                    ].expand(_random_state_dict[item_key].shape)
+                                _new_state_dict[item_key] = (
+                                    _origin_state_dict[new_key].clone().detach()
+                                )
                             else:
                                 new_key = item_key.replace(
                                     f".{_model_key}.", f".{_model_key_from}."