Accelerating 2K-scale pre-training by 1.28× with TorchAO, MXFP8 and TorchTitan | Not Hacker News!