I am running an ensemble of catboost, lightgmb and xgboost, in kaggle and colab.
I have issues with reproducing the same result, while using cpu or gpu.
This is for the lightgbm code.

import os
def seed_everything(seed=0):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    os.environ['CUDA_LAUNCH_BLOCKING'] = str(1)
    np.random.seed(seed)

rmse_average = []
test_preds = []

seed_everything(seed=0)
kf = KFold(n_splits=10, shuffle=True, random_state=0)
for fold, (train_index, val_index) in enumerate(kf.split(X, y)):
  seed_everything(seed=0)
  print(50*'-')
  print(f'Fold {fold+1}')
  X_train, X_val = X.iloc[train_index], X.iloc[val_index]
  y_train, y_val = np.log(y.iloc[train_index]), np.log(y.iloc[val_index])

  seed_everything(seed=0)
  model = LGBMRegressor(device = 'gpu', verbose=0, random_seed=0)
  seed_everything(seed=0)
  model.fit(X_train, y_train, eval_set=[(X_val, y_val)], eval_metric='rmse')#, early_stopping_rounds=50, verbose=500)
  seed_everything(seed=0)
  val_preds = np.exp(model.predict(X_val))
  
  seed_everything(seed=0)
  rmse = np.sqrt(mean_squared_error(np.exp(y_val), val_preds))
  print("Validation RMSE:", rmse)

  rmse_average.append(rmse)

  # Append predictions from each fold to test_preds
  seed_everything(seed=0)
  test_pred_fold = np.exp(model.predict(test))
  test_preds.append(test_pred_fold)


print(f'nAverage RMSE: {mean(rmse_average)}')

# Aggregate predictions from all folds
test_preds_combined = np.mean(test_preds, axis=0)

sample_submission['pm2_5'] = test_preds_combined
sample_submission.to_csv('submission.csv', index=False)