default_stage:
  default_modifiers:
    AWQModifier:
      mappings:
      - smooth_layer: re:.*language_model\.layers\.([3-9]|[1-5][0-9])\.post_attention_layernorm$
        balance_layers: ['re:.*language_model\.layers\.\d+\.mlp\.gate$', 're:.*language_model\.layers\.\d+\.mlp\.experts\.\d+\.gate_proj$',
          're:.*language_model\.layers\.\d+\.mlp\.experts\.\d+\.up_proj$', 're:.*language_model\.layers\.\d+\.mlp\.shared_experts\.gate_up_proj$']
        activation_hook_target: null
      - smooth_layer: re:.*language_model\.layers\.\d+\.mlp\.experts\.\d+\.up_proj$
        balance_layers: ['re:.*language_model\.layers\.\d+\.mlp\.experts\.\d+\.down_proj$']
        activation_hook_target: null
      offload_device: cpu
      duo_scaling: true
      n_grid: 24
    QuantizationModifier:
      config_groups:
        group_0:
          targets: ['re:.*language_model\.layers\.\d+\.mlp\.experts\.\d+\.(gate_proj|up_proj|down_proj)$']
          weights:
            num_bits: 4
            type: int
            symmetric: true
            group_size: 128
            strategy: group
            block_structure: null
            dynamic: false
            actorder: null
            scale_dtype: null
            zp_dtype: null
            observer: mse
            observer_kwargs: {}
          input_activations: null
          output_activations: null
          format: null
      targets: [Linear]
      ignore: [lm_head]
      bypass_divisibility_checks: false