# W4A16 metadata is deliberate. Explicit W4A8 metadata enters the fork's
# CPU-only W4A8 MoE adapter; Marlin's INT8 activation override is a runtime choice.
default_stage:
  default_modifiers:
    AWQModifier:
      mappings:
        - smooth_layer: re:model.*input_layernorm$
          balance_layers:
            - re:model.*self_attn[.]q_proj$
            - re:model.*self_attn[.]k_proj$
            - re:model.*self_attn[.]v_proj$
        - smooth_layer: re:model.*post_attention_layernorm$
          balance_layers:
            # The BF16 router must be balanced too, preserving its logits when
            # the shared input normalization is smoothed.
            - re:model.*mlp[.]gate$
            - re:model.*mlp[.]experts.*gate_proj$
            - re:model.*mlp[.]experts.*up_proj$
        - smooth_layer: re:model.*mlp[.]experts.*up_proj$
          balance_layers:
            - re:model.*mlp[.]experts.*down_proj$
      duo_scaling: both
      n_grid: 20
      offload_device: cpu
    QuantizationModifier:
      config_groups:
        group_0:
          targets: [Linear]
          weights:
            num_bits: 4
            type: int
            symmetric: true
            group_size: 32
            strategy: group
            dynamic: false
            observer: mse
          input_activations: null
          output_activations: null
      ignore:
        - lm_head
        - re:.*embed_tokens.*
        - re:.*mlp[.]gate$
        - re:.*mtp.*
      bypass_divisibility_checks: false