--model-type=transformer --distribution-strategy=llm-training --enable-saturate-infinity