recipe.yaml

quant_stage:
  quant_modifiers:
    QuantizationModifier:
      ignore: ['re:.*lm_head', 're:multi_modal_projector.*', 're:vision_tower.*']
      targets: [Linear]
      scheme: FP8_DYNAMIC
      kv_cache_scheme: {num_bits: 8, type: float, strategy: tensor, dynamic: false, symmetric: true}