Latest version training crashes

[ganlei333]

My environment is Python 3.9 established through Conda. The default installation for XGBoost is 1.7.0, so there is no problem training with this version. I can train normally, but when I upgraded XGBoost to 2.0.3 and used the same code and data for training, I encountered a crash. And 100% reproduce the crash. My hardware is 2697AV4, with 1TB of DDR4-2400T memory and SSD solid-state drive. CPU mode training. The ntthread parameter is set to -1, with a dataset size of 89M and 195 features for binary classification.

My code function is as follows:

def optimize_hyperparameters(X_train, y_train):
    param_dist = {
        'n_estimators': randint(100, 600),
        'learning_rate': uniform(0.01, 0.2),
        'subsample': uniform(0.5, 0.5),
        'max_depth': randint(20, 30),
        'colsample_bytree': uniform(0.5, 0.5),
        'min_child_weight': randint(1, 10)
    }

    scale_pos_weight = len(y_train[y_train == 0]) / len(y_train[y_train == 1])

    xgb_clf = xgb.XGBClassifier(use_label_encoder=False, eval_metric='logloss',scale_pos_weight=scale_pos_weight,nthread=-1)
    rscv = RandomizedSearchCV(xgb_clf, param_distributions=param_dist, n_iter=500, scoring='roc_auc', cv=3, verbose=2, random_state=42, n_jobs=1)
    rscv.fit(X_train, y_train)

    print(f"Best parameters found: {rscv.best_params_}")
    return rscv.best_estimator_

[ganlei333]

sklearn ：1.3.0

[trivialfis]

Hi, thank you for raising the issue. Could you please provide a reproducible example that I can run on my machine?

[ganlei333]

I have provided the training function, but the data is too large to upload successfully.

[trivialfis]

Could you please provide the code in a more complete form so that I can run it with synthesized data?

[trivialfis]

In addition, could you please share what you mean by crash? Is it running out of memory or is it running into segfault?

[ganlei333]

Uploading xgb.zip…
The crash was caused by the server automatically restarting

[trivialfis]

Feel free to close the issue if this is not related to XGBoost itself.

[ganlei333]

这个
I have tested that there is no problem with training XGBoost 2.0.3 in the py3.10 environment of the Windows 2019 operating system, while in the py3.11 environment of the debian12 operating system of Linux, as long as the training system is trained, it will immediately restart (the data cleaning part is normal, and once the training logic is entered, the system restart will be triggered). And the operating system environment is very clean. It's not a CPU temperature issue either.

[trivialfis]

By restart, do you mean the login session restart, or the whole OS got restarted?

[ganlei333]

Restart the entire operating system

[trivialfis]

Then that's beyond me. The OS has bugs, likely in the kernel, you may upgrade your OS, I don't think this is related to XGBoost.

[trivialfis]

Feel free to reopen it if you need further information.