Result of model with linear trees depends on the number of used during fitting CPU threads

[StrikerRUS]

Reproducible example:

import sys

import numpy as np
import lightgbm as lgb

sys.path.append(r'LightGBM\tests\python_package_test')

from test_consistency import FileLoader, EXAMPLES_DIR

fd = FileLoader(EXAMPLES_DIR / 'binary_classification', 'binary', 'train_linear.conf')
X_train, y_train, _ = fd.load_dataset('.train')
X_test, _, _ = fd.load_dataset('.test')

def fit_predict(n_jobs):
    gbm = lgb.LGBMClassifier(**fd.params, verbose=-1, n_jobs=n_jobs)
    gbm.fit(X_train, y_train)
    return gbm.predict_proba(X_test)[:, 1]

pred_1 = fit_predict(n_jobs=2)
pred_2 = fit_predict(n_jobs=4)

np.testing.assert_allclose(pred_1, pred_2)

AssertionError: 
Not equal to tolerance rtol=1e-07, atol=0

Mismatched elements: 500 / 500 (100%)
Max absolute difference: 0.33526452
Max relative difference: 2.5919644
 x: array([0.814095, 0.543692, 0.029034, 0.763285, 0.151306, 0.157106,
       0.237552, 0.146994, 0.936084, 0.366068, 0.790745, 0.871099,
       0.862789, 0.84715 , 0.621987, 0.866871, 0.276202, 0.76066 ,...
 y: array([0.824547, 0.466547, 0.060617, 0.598674, 0.05619 , 0.097133,
       0.215748, 0.109334, 0.926242, 0.378924, 0.871532, 0.839161,
       0.857007, 0.845114, 0.69853 , 0.683535, 0.201123, 0.697704,...

Bug was originally found in #5105 (comment).

Fix of this issue should include removing n_jobs=0 workaround here

LightGBM/tests/python_package_test/test_consistency.py

Line 87 in f715645

gbm = lgb.LGBMClassifier(**fd.params)

[StrikerRUS]

@btrotta Could you please help with this?

[btrotta]

I think I've fixed this; it was due to a numerical precision issue and was not specific to the scikit-learn interface (I've added a test that reproduces the issue in the normal python interface).

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.