recipe.yaml

default_stage:
  default_modifiers:
    QuantizationModifier:
      ignore: [lm_head, 're:.*mlp.gate$', 're:.*mlp.shared_expert_gate$']
      targets: [Linear]
      scheme: W4A16