default_stage:
  default_modifiers:
    AWQModifier:
      config_groups:
        group_0:
          targets: [Linear]
          weights:
            num_bits: 4
            type: int
            symmetric: true
            group_size: 32
            strategy: group
            block_structure: null
            dynamic: false
            actorder: null
            observer: mse
            observer_kwargs: {}
          input_activations: null
          output_activations: null
          format: null
      targets: [Linear]
      ignore: [model.embed_tokens, 're:model[.]layers[.]0[.].*', 're:.*input_layernorm$',
        're:.*norm.*', 're:.*shared_experts.*', 're:.*block_sparse_moe[.]gate$', 're:.*router.*',
        're:.*post_attention_layernorm$', 're:.*self_attn.*', lm_head]
      mappings:
      - smooth_layer: re:.*input_layernorm$
        balance_layers: ['re:.*q_proj$', 're:.*k_proj$', 're:.*v_proj$', 're:.*kv_a_proj_with_mqa$',
          're:.*f_a_proj$', 're:.*b_proj$', 're:.*g_a_proj$']
      - smooth_layer: re:.*f_a_proj$
        balance_layers: ['re:.*f_b_proj$']
      - smooth_layer: re:.*g_a_proj$
        balance_layers: ['re:.*g_b_proj$']
      - smooth_layer: re:.*kv_a_layernorm$
        balance_layers: ['re:.*kv_b_proj$']
      - smooth_layer: re:.*v_proj$
        balance_layers: ['re:.*o_proj$']
      - smooth_layer: re:.*post_attention_layernorm$
        balance_layers: ['re:.*gate_proj$', 're:.*up_proj$', 're:.*w1$', 're:.*w3$']
      - smooth_layer: re:.*up_proj$
        balance_layers: ['re:.*down_proj$']
      - smooth_layer: re:.*w3$
        balance_layers: ['re:.*w2$']
      - smooth_layer: model.norm
        balance_layers: [lm_head]
      offload_device: !!python/object/apply:torch.device [cpu]
      duo_scaling: true