default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head, model.language_model.embed_tokens, model.visual, 're:^model\.visual\..*',
        're:.*delta_attn\.conv1d$', 're:.*delta_attn\.in_proj_a$', 're:.*delta_attn\.in_proj_b$',
        're:^mtp.*']
      scheme: FP8
      bypass_divisibility_checks: false
      requires_calibration_data: true