default_stage: default_modifiers: QuantizationModifier: targets: [Linear] ignore: [lm_head, 're:.*gate.*', 're:.*embed.*', 're:.*kv_a.*'] scheme: FP8 bypass_divisibility_checks: false