default_stage:
  default_modifiers:
    AWQModifier:
      mappings:
      - smooth_layer: re:model\.layers\.[1-9][0-9]*\.post_attention_layernorm$
        balance_layers: ['re:.*mlp.gate$', 're:.*mlp.experts.*.gate_proj$', 're:.*mlp.experts.*.up_proj$']
        activation_hook_target: null
      - smooth_layer: re:.*mlp.experts.*.up_proj$
        balance_layers: ['re:.*mlp.experts.*.down_proj$']
        activation_hook_target: null
      offload_device: cpu
      duo_scaling: false
      n_grid: 4
    QuantizationModifier:
      targets: ['re:.*mlp\.experts\..*_proj$']
      ignore: [lm_head, 're:.*mlp\.gate$', 're:.*self_attn.*', 're:.*\.mtp\..*', 're:.*visual.*',
        're:.*audio.*', 're:.*speech.*', 're:.*\.layers\.41\.mlp\.experts\..*']
      scheme: W4A16
      bypass_divisibility_checks: false