# @package _global_ parameterization: ar trainer: ar_shift: true model: full_attention: false use_flex_attention: false