default_stage:
  quantization_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head, 're:.*gate$', 're:.*self_attn.*', 're:.*layers\.[0-2]\..*', 're:.*shared_experts.*']
      scheme: FP8_BLOCK
      bypass_divisibility_checks: false
      requires_calibration_data: false