default_stage: default_modifiers: QuantizationModifier: targets: [Linear] ignore: [embed_tokens, lm_head] scheme: W4A16