My initial reaction is that we could add a class method Step._get_crds_parameters (although the naming could be confusing). Something like:

@classmethod
def _get_crds_parameters(dataset):
    ...
    return crds_parameters, observatory

The need for the class method is for get_config_from_reference (which is a classmethod and called before a Step is created) where the usage of the new _get_crds_parameters would look something like:

- with cls._datamodels_open(dataset, asn_n_members=1) as model: 
-     if isinstance(model, Sequence): 
-         crds_parameters = model._models[0].get_crds_parameters() 
-         crds_observatory = model.crds_observatory 
-     else: 
-         crds_parameters = model.get_crds_parameters() 
-         crds_observatory = model.crds_observatory 
+ crds_parameters, crds_observatory = cls._get_crds_parameters(dataset)

Things are a little trickier for _precache_references. The current code is written in a way where a container with multiple models would be iterated over to get references for all contained models. However the container is opened with asn_n_members=1, asn_exptypes=["science"] (so only the first science exposure is used for pre-fetching). This might be a bug and is worth some discussion. If we say "only the first science exposure is used for prefetching" then we could likely simplify _precache_references combining it with _precache_references_opened and _precache_references_impl adding a call to cls._get_crds_parameters. If we say this is a bug and perhaps we want to leave this up to the child class we could consider making cls._get_crds_parameters a generator and implementing _precache_references as:

for parameters, observatory in cls._get_crds_parameters(dataset):
   # handle overrides ...
   crds_refs = crds_client.get_multiple_reference_paths(...)

Another complication of _precache_references is that it logs meta.filename. This won't be loaded by spacetelescope/stdatamodels#284