default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head, 're:.*mlp\.gate$', 're:.*mlp\.shared_expert_gate$', 're:.*\.input_layernorm$',
        model.norm, 're:.*\.norm$', 're:.*\.q_norm$', 're:.*\.k_norm$', 're:.*\.rotary_emb.*',
        embed_tokens]
      scheme: NVFP4
      bypass_divisibility_checks: false