Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@deepmd/pt_expt/train/training.py`:
- Around line 384-385: The current logic sets resume_model = init_model or
restart_model or finetune_model so finetune can incorrectly pick init/restart
checkpoints; change this by (a) validating inputs up front in the function that
defines init_model/restart_model/finetune_model and raise an error if more than
one of those is set, OR (b) keep the existing resume_model variable but in the
finetune branch explicitly load weights from finetune_model (not resume_model)
and use that checkpoint to populate descriptor/fitting weights and
_extra_state["model_params"]; update both the initial resume/resuming block
(resume_model/resuming) and the finetune-specific code region (~lines 487-527)
to follow the chosen approach so finetune never inherits init/restart weights.
- Around line 991-1002: The log attempts to convert CUDA tensors returned by
_model.get_out_bias() to numpy via np.asarray which raises RuntimeError on CUDA;
replace the np.asarray(...) calls with to_numpy_array(...) from
deepmd.dpmodel.common when building the log message after calling
_model.change_out_bias (and similarly anywhere else you call np.asarray on
_model.get_out_bias()), so call to_numpy_array(old_bias).reshape(-1) and
to_numpy_array(new_bias).reshape(-1) (slicing by len(model_type_map) as before)
to ensure device-safe conversion for logging.

In `@deepmd/pt_expt/utils/finetune.py`:
- Around line 35-40: The code currently falls back to returning only
{"type_map": ...} when serialize_from_file(finetune_model) lacks "model_params",
which silently allows change_model_params=True to proceed with incomplete
config; modify the logic in the finetune model-loading blocks (where
serialize_from_file, finetune_model is used and again in the block around lines
79-92) to detect when change_model_params is True and "model_params" is missing,
and immediately raise a clear error (including mention of using
--use-pretrain-script or that legacy .pte lacks model_params.json) instead of
returning the minimal dict; ensure the error path prevents calling
get_finetune_rule_single with incomplete input.

In `@source/tests/consistent/model/test_ener.py`:
- Around line 1333-1423: The test wrongly sets dp_std_orig =
to_numpy_array(dp_model.get_out_bias()) instead of snapshotting the original
out_std and then never asserts remapping for old types; change the snapshot to
dp_std_orig = to_numpy_array(dp_model.atomic_model.out_std) (and similarly
ensure any other std snapshots use atomic_model.out_std), seed a non-trivial
out_std on dp_model before change_type_map, then add assertions that the
remapped old entries land at indices 3 and 0 (compare dp_std_new[:, 3, :] to
dp_std_orig[:, 0, :] for "O" and dp_std_new[:, 0, :] to dp_std_orig[:, 1, :] for
"H"), keep cross-backend equality checks (pt_model, pt_expt_model) and remove or
use any now-unused locals; run ruff check . and ruff format . before committing.

In `@source/tests/pt_expt/test_finetune.py`:
- Around line 564-580: The loop comparing ft_state and pre_state must, when
random_fitting is True, assert that fitting tensors are not all identical:
locate the loop over ft_state and the variables ft_state, pre_state and
random_fitting; gather keys containing ".fitting" present in both ft_state and
pre_state and assert that at least one of those tensors differs (e.g., by
checking torch.any(ft_state[k] != pre_state[k]) for at least one k), failing the
test if all fitting tensors are equal.

Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@source/tests/pt_expt/test_finetune.py`:
- Around line 122-132: The helper _subsample_data currently does a redundant
local import "import shutil as _shutil"; remove that line and use the module
already imported at the top-level (shutil) when calling copytree in
_subsample_data so the function uses shutil.copytree(dst_dir, ...) instead of
the locally imported _shutil; update references in _subsample_data to shutil to
avoid the duplicate import.