Different behaviour on python-exported model and C++ API predictor

Catboost version: 1.2.2
CPU: Intel Core i5-8550U
OS: Ubuntu [20.04]

I have the trained catboost model (see multi_class.py in attachment) and saved it as
model.cbm and model.py. Besides,
I have saved the input and the answer for one row in input.1.json.
In C++ I have the following code:

auto wrapper = ModelCalcerWrapper("model.cbm");

// parse input.1.json

for (const InputRow &row : in.input) {
auto pred = wrapper.CalcMulti(row.vector_float_features, row.vector_categorial_features);
// check that difference (pred - expected) is small enough
}

and the answer matches.
If I look at the python-exported file, then the error is immediately visible there:

current_tree_leaf_values_index += (1 << current_tree_depth) * model.dimension

* model.dimension shouldn't be there (as it is in cpp-exported version, for now it raises index out of bound
exception).
I fixed it in model.fixed.py file and add some code to predict a model:

if __name__ == "__main__":
    input_file = "input.1.json"
    data = open(input_file)
    item = json.load(data)[0]
    ans = item['ans']
    float_feats = item['float_features']
    cat_feats = item['cat_features']
    resp = apply_catboost_model_multi(float_feats, cat_feats)
    print("expected = {}".format(ans))
    print("real     = {}".format(resp))

# expected = [-0.4315705250918395, -0.07602514583990287, 0.5075956709317426]
# real     = [-0.012774193043495252, -0.048760865913935775, 0.38847943878388946]

but the answer doesn't match. I suppose it's a problem. Or am I doing something wrong?

Attachment

https://gist.github.com/mkornaukhov03/5c5d9e394f17141cac4fa63d2b09e026