recipe.yaml

DEFAULT_stage:
  DEFAULT_modifiers:
    SmoothQuantModifier: {smoothing_strength: 0.8}
    QuantizationModifier:
      ignore: [lm_head]
      targets: Linear
      scheme: W4A16
      kv_cache_scheme: {num_bits: 8, type: float, symmetric: true, strategy: tensor, dynamic: false}