default_stage:
  default_modifiers:
    AWQModifier:
      mappings:
      - smooth_layer: re:^model\.language_model\.layers\.\d+\.input_layernorm$
        balance_layers: ['re:^model\.language_model\.layers\.\d+\.self_attn\.q_proj$', 're:^model\.language_model\.layers\.\d+\.self_attn\.k_proj$',
          're:^model\.language_model\.layers\.\d+\.self_attn\.v_proj$']
        activation_hook_target: null
      - smooth_layer: re:^model\.language_model\.layers\.\d+\.self_attn\.v_proj$
        balance_layers: ['re:^model\.language_model\.layers\.\d+\.self_attn\.o_proj$']
        activation_hook_target: null
      - smooth_layer: re:^model\.language_model\.layers\.\d+\.post_attention_layernorm$
        balance_layers: ['re:^model\.language_model\.layers\.\d+\.mlp\.gate_proj$', 're:^model\.language_model\.layers\.\d+\.mlp\.up_proj$']
        activation_hook_target: null
      - smooth_layer: re:^model\.language_model\.layers\.\d+\.mlp\.up_proj$
        balance_layers: ['re:^model\.language_model\.layers\.\d+\.mlp\.down_proj$']
        activation_hook_target: null
      duo_scaling: true
      n_grid: 20
    QuantizationModifier:
      targets: ['re:^model\.language_model\.layers\.\d+\.self_attn\.(q_proj|k_proj|v_proj|o_proj)$',
        're:^model\.language_model\.layers\.\d+\.mlp\.(gate_proj|up_proj|down_proj)$']
      ignore: []
      scheme: W4A16_ASYM
      bypass_divisibility_checks: false