default_stage:
  default_modifiers:
    GPTQModifier:
      targets: [Linear]
      ignore: ['re:.*lm_head', 're:.*self_attn', 're:.*router', 're:vision_model.*', 're:multi_modal_projector.*',
        Llama4TextAttention]
      scheme: W4A16
      sequential_update: true
      block_size: 128
      dampening_frac: 0.01
      offload_hessians: false