The same heterogeneous treatment effect for each sample

Try to use NonParamDML for task with binary target Y~{0,1} and continious treatment T.

Define and fit model as

est3 = NonParamDML(model_y=CatBoostClassifier(n_estimators=10, max_depth=4),
                   model_t=CatBoostRegressor(n_estimators=10, max_depth=4),
                   model_final=RandomForestRegressor(),
                   discrete_treatment=False, 
                   cv=3)
est3.fit(train.target.values, train.treatment_size.values, X=train.loc[:, features_econml].values)

Then apply model to valid dataset:

te_pred_3 = est3.effect(valid.loc[:, features_econml].values) 

and get only one unique value in prediction: 8.35220653631326⋅10−6 and it is near to zero..

Feature values for each sample are different.

Tell me please, have I made any mistake?

Start by reproducing the shown NonParamDML fit and effect call with the stated models and checking the unique predictions. Compare the result with the expected behavior for differing feature values; done means determining whether the constant near-zero effect is expected or establishing a reproducible defect.