quant_stage:
  quant_modifiers:
    QuantizationModifier:
      config_groups:
        group_0:
          input_activations: {dynamic: true, num_bits: 8, observer: memoryless, strategy: token,
            symmetric: true, type: float}
          targets: [Linear]
          weights: {num_bits: 8, observer: mse, strategy: channel, symmetric: true, type: float}
      ignore: [lm_head]