default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head, 're:.*embed_tokens$', 're:.*norm$', 're:.*embed_tokens_per_layer.*',
        're:.*per_layer_model_projection.*', 're:.*vision_tower.*', 're:.*audio_tower.*',
        're:.*multi_modal_projector.*']
      scheme: W8A16
      bypass_divisibility_checks: false