input_quantizer with mx format like mxint8 was diabled when use awq_lite.

[bestzsq]

I want to test awq_lite with mx format input_quantizer, however input_quantizer was disabled after call awq_lite function, my quant_cfg and quant_summary:

quant_cfg = {
    "quant_cfg": {
        "*weight_quantizer": {
            "num_bits": (2, 1),
            "block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
            "axis": None,
            "enable": True,
        },
        "*input_quantizer": {
            "num_bits": 8,
            "block_sizes": {-1: 32, "type": "dynamic", "scale_bits": (8, 0)},
            "enable": True,
        },
        **_default_disabled_quantizer_cfg,
    },
    "algorithm": "awq_lite",
}

and with below quant summary(awq_lite/)

model.layers.23.mlp.up_proj.input_quantizer                                      TensorQuantizer(disabled)
model.layers.23.mlp.up_proj.output_quantizer                                     TensorQuantizer(disabled)
model.layers.23.mlp.up_proj.weight_quantizer                                     TensorQuantizer((2, 1) bit fake block_sizes：{-1: 16, 'type': 'dynamic', 'scale_bits': (4, 3)}, amax=2.4219 calibrator=MaxCalibrator quant)

when i try awq_clip with mx format input_quantizer, the input_quantizer was as expected：

{
    "quant_cfg": {
        "*weight_quantizer": {
            "num_bits": (2, 1),
            "block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
            "axis": None,
            "enable": True,
        },
        "*input_quantizer": {
            "num_bits": 8,
            "block_sizes": {-1: 32, "type": "dynamic", "scale_bits": (8, 0)},
            "enable": True,
        },
        **_default_disabled_quantizer_cfg,
    },
    "algorithm": {"method": "awq_clip"},
}

and with quant summary(awq_clip)

model.layers.23.mlp.down_proj.input_quantizer                                    TensorQuantizer(8 bit fake block_sizes={-1: 32, 'type': 'dynamic', 'scale_bits': (8, 0)}, amax=None calibrator=MaxCalibrator quant)
model.layers.23.mlp.down_proj.output_quantizer                                   TensorQuantizer(disabled)
model.layers.23.mlp.down_proj.weight_quantizer                                   TensorQuantizer((2, 1) bit fake block_sizes={-1: 16, 'type': 'dynamic', 'scale_bits': (4, 3)}, amax=0.5000 calibrator=MaxCalibrator quant)

[bestzsq]

I have replaced the code from

TensorRT-Model-Optimizer/modelopt/torch/quantization/model_calib.py

Line 651 in 7d0f7a9

if module.awq_lite.is_input_quantized and module.input_quantizer.amax is not None:

with below, it looks like it's working well.

        if module.awq_lite.is_input_quantized and module.input_quantizer._amax is not None:
            act_amax = module.input_quantizer._amax