default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head, model.language_model.embed_tokens, 're:.*visual.*', 're:.*vision.*',
        're:.*attn.*', 're:.*shared_expert.*', 're:.*mtp.*', 're:.*mlp\.gate$']
      scheme: NVFP4A16
      bypass_divisibility_checks: false