Xgboost random seed setting does not ensures reproduce-ability

Question

In spite of setting a value for random_state and/or seed parameter, the performance is not reproducible in Xgboost Sklearn API wrapper

Here is the code

from xgboost.sklearn import XGBClassifier
from sklearn.metrics import roc_auc_score

(X_train,y_train),(X_eval,y_eval) = pickle.load(open('xxxx.pkl',"rb"))

hyperparams = {'eval_metric': 'auc', 'colsample_bylevel': 0.7, 'learning_rate': 0.125,  'random_state': 0}
GBM = XGBClassifier(**hyperparams)
GBM.fit(X_train,y_train,eval_metric="auc",verbose = True,eval_set=[(X_eval,y_eval)],early_stopping_rounds=2)

print(roc_auc_score(y_eval, GBM.predict_proba(X_eval)[:,1]))

Each time, when I run the above snippet the concordance value differs.

78.9451246
79.001542

Some references issues