deepseek_recipe_w4a16.yaml

quant_stage:
  quant_modifiers:
    GPTQModifier:
      ignore: [lm_head, "re:.*mlp.gate$"]
      config_groups:
        group_0:
          weights: {num_bits: 4, type: int, symmetric: true, strategy: channel, dynamic: false}
          targets: [Linear]