recipe.yaml

default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head, 're:.*mlp.gate$', 're:.*mlp.shared_expert_gate$', 're:.*linear_attn.*',
        're:.*layers.92.shared_head.*', 're:.*layers.92.eh_proj.*']
      scheme: NVFP4