How can one load a multi-head model from 3.0.0b4 in devel version?

[QuantumMisaka]

By using these code:

from deepmd.infer.deep_pot import DeepPot
DP = DeepPot("FeCHO-dpa230-4head-80w-64-30656.pt", head="Target_FTS")

I can load my DPA-2.3.0 multitask-finetuned model by using deepmd-kit 3.0.0b4, but in the devel branch of deepmd-kit. error will occur:

In [27]: DP = DeepPot("FeCHO-dpa230-4head-80w-64-30656.pt", head="Target_FTS")
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[27], line 1
----> 1 DP = DeepPot("FeCHO-dpa230-4head-80w-64-30656.pt", head="Target_FTS")

File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/deepmd/infer/deep_eval.py:330, in DeepEval.__init__(self, model_file, auto_batch_size, neighbor_list, *args, **kwargs)
    322 def __init__(
    323     self,
    324     model_file: str,
   (...)
    328     **kwargs: Any,
    329 ) -> None:
--> 330     self.deep_eval = DeepEvalBackend(
    331         model_file,
    332         self.output_def,
    333         *args,
    334         auto_batch_size=auto_batch_size,
    335         neighbor_list=neighbor_list,
    336         **kwargs,
    337     )
    338     if self.deep_eval.get_has_spin() and hasattr(self, "output_def_mag"):
    339         self.deep_eval.output_def = self.output_def_mag

File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/deepmd/pt/infer/deep_eval.py:135, in DeepEval.__init__(self, model_file, output_def, auto_batch_size, neighbor_list, head, *args, **kwargs)
    133     model = torch.jit.script(model)
    134     self.dp = ModelWrapper(model)
--> 135     self.dp.load_state_dict(state_dict)
    136 elif str(self.model_path).endswith(".pth"):
    137     model = torch.jit.load(model_file, map_location=env.DEVICE)

File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/torch/nn/modules/module.py:2584, in Module.load_state_dict(self, state_dict, strict, assign)
   2576         error_msgs.insert(
   2577             0,
   2578             "Missing key(s) in state_dict: {}. ".format(
   2579                 ", ".join(f'"{k}"' for k in missing_keys)
   2580             ),
   2581         )
   2583 if len(error_msgs) > 0:
-> 2584     raise RuntimeError(
   2585         "Error(s) in loading state_dict for {}:\n\t{}".format(
   2586             self.__class__.__name__, "\n\t".join(error_msgs)
   2587         )
   2588     )
   2589 return _IncompatibleKeys(missing_keys, unexpected_keys)

RuntimeError: Error(s) in loading state_dict for ModelWrapper:
        Missing key(s) in state_dict: "model.Default.min_nbor_dist", "model.Default.atomic_model.descriptor.repinit.compress_info.0", "model.Default.atomic_model.descriptor.repinit.compress_data.0"

But the frozen model from deepmd-kit 3.0.0b4 can still be read-in and be used for evaluation, so the compability should be maintained, but why can't multi-head model be load ?

[iProzd]

This PR involves breaking change: #4300 fyi @njzjz @wanghan-iapcm

[njzjz]

There is no guarantees of backwards compatibility across versions for the checkpoint file.