default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head, model.embed_tokens, 're:.*block_sparse_moe\.gate$']
      scheme: NVFP4A16