How do you calculate the exact memory required to train a 7-billion parameter transformer model at Mistral AI using mixed-precision training?
