quantization_recipe.yaml

quant_stage:
    quant_modifiers:
        QuantizationModifier:
            ignore: ["lm_head", "visual"]
            config_groups:
                group_0:
                    targets: ["Linear"]
                    input_activations: null
                    weights:
                        num_bits: 8
                        symmetric: true
                        strategy: channel
                        observer: minmax