VIDRAFT Trains a Language Model from Scratch on NPUs — No GPUs Required