Compute-efficient pretraining and scaling to trillion-parameter models