How to parallelize functions in cuda?

Question

GPGPU

Hello, I am writing a neural network for digit recognition in cuda. I want to get maximum acceleration. On python, 10,000 workouts are completed in 20 seconds. In C++ in 10 seconds. Now it's cuda's turn. There is a neuralNet class and it has 3 functions: constructor, training and polling. How can I simultaneously call several training functions at once? I understand that control flows and blocked memory will be needed there, but I haven’t worked with this much and I can’t put it all together.
Thanks in advance for your reply)