Efficient approach to calculate test RMSE score in Time series forecasting

I have trained a Time series LSTM model of unique product id each having n days data. I used a keras Dataloader to send those product id in singularity to train. I'm calculating RMSE score using this code

start_time = time.time()

for i in range(10000):

ext=l[I]
'''ext here repserents unique product id'''
train,test=r[ext][:21],r[ext][20:31]


global trainX,trainY,testX,testY
trainX, trainY = create_dataset(train, look_back)
testX, testY = create_dataset(test, look_back)
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))


trainPredict = new_model.predict(trainX)
testPredict = new_model.predict(testX)
trainPredictinv = scalar.inverse_transform(trainPredict)

trainYinv = scalar.inverse_transform([trainY])

testPredictinv = scalar.inverse_transform(testPredict)
testYinv = scalar.inverse_transform([testY])

trainScore = np.sqrt(mean_squared_error(trainYinv[0], trainPredictinv[:,0]))

prodtrainScore.append(trainScore)
testScore = np.sqrt(mean_squared_error(testYinv[0], testPredictinv[:,0]))
prodtestScore.append(testScore)



end_time = time.time()
print('Execution time = %.6f seconds' % (end_time-start_time))  

this is taking a lot of time to calculate RMSE of 1000000 queries. How can I improvise this?

Create dataset function used above:

def create_dataset(dataset, look_back=1):
dataX, dataY = [], []
for p in range(len(dataset)-(look_back-1)):
    a = dataset[p:(p+look_back)]
    dataX.append(a)
    dataY.append(dataset[p])
return np.array(dataX), np.array(dataY)