In `@deepmd/dpmodel/utils/serialization.py`:
- Around line 299-301: The loop handling data_dict currently deep-copies the
"@variables" payload (variables = deepcopy(vv)), which unnecessarily duplicates
large numpy arrays; change this to a shallow copy instead so we don't duplicate
underlying array buffers—e.g., assign a shallow copy appropriate to the
container type (for numpy arrays use vv.copy() or for mapping types use
dict(vv)) or simply vv if already an independent object; update the assignment
for the "@variables" branch (the variables variable assigned from
data_dict["@variables"]) to use a shallow copy strategy rather than deepcopy to
avoid doubling memory during deserialization.

In `@deepmd/dpmodel/utils/serialization.py`:
- Around line 348-358: The current merge logic in the loop over self.children
(using children_buff, child_repr, last_repr) collapses distinct dict keys when
their stringified child subtrees match; restrict merging to only list-type
parents by checking the parent node type before merging (e.g., only perform the
"if child_repr == last_repr: ... merge ..." sequence when the current node
represents a list node), or add a guard that inspects the key type/format (the
kk variable) to ensure keys are positional indices before combining; update the
loop around children_buff, kk, vv and the merge branch so dict children are
never concatenated into "a, b" while preserving the existing merge behavior for
true list nodes.
- Around line 181-201: The format_big_number function can produce results like
"1000.0K" when rounding crosses a unit boundary; update format_big_number to
detect when the rounded quotient equals or exceeds 1000.0 and promote to the
next suffix (e.g., convert 1000.0K → 1.0M, 1000.0M → 1.0B). Concretely, compute
the quotient for each branch (q = x / 1_000, 1_000_000, 1_000_000_000),
round/format it to one decimal, and if q >= 1000.0, fall through to the next
unit branch so the final string uses the higher suffix; keep the same
one-decimal formatting and ensure the else branch returns str(x).

In `@deepmd/dpmodel/utils/serialization.py`:
- Around line 348-378: The closure format_value currently captures the loop
variable ii from the outer for loop (line with self.children.items()), causing
wrong indentation; rename that outer ii to _ (unused) and change format_value to
accept ii as a parameter (e.g., def format_value(vv: str, ii: int) -> str) and
update the generator that builds buff to call format_value(vv, ii) using the ii
from enumerate(children_buff); keep format_list_keys and children_buff logic
unchanged (ListNode handling remains the same).