Explain how you would implement gradient checkpointing and mixed-precision (FP8/BF16) training to fit a massive video model into GPU memory at Genmo.
