recipe.yaml

quant_stage:
  quant_modifiers:
    QuantizationModifier:
      config_groups:
        fp8_attention:
          output_activations: {num_bits: 8, type: float, strategy: channel, dynamic: false,
            symmetric: true}
          targets: ['re:.*q_proj', 're:.*k_proj', 're:.*v_proj']