3 ms·Accelerating 2K-scale pre-training by 1.28× with TorchAO, MXFP8 and TorchTitan1 points by andrewstetsenko 1y ago