Whenever you're ready, walk me through how you'd approach this.
Explain how mixed precision interacts with convergence and numerical stability for a large model at NVIDIA.
Describe a practical training design using FP16 and BF16, including precision selection, loss scaling, optimizer state handling, gradient accumulation, overflow detection, and monitoring. Explain how you would demonstrate that reduced precision has not changed optimization behavior or final model quality. Include a small implementation or experiment plan suitable for PyTorch on NVIDIA GPUs.