Float Loss Will expand to more formats If I can safely use them with pytorch, or maybe LLM.c supports it? current graph (tiny-shakesperae):
