tf.keras.quantizers.Int4QuantizationConfig

Int4 quantization config.

tf.keras.quantizers.Int4QuantizationConfig(
    weight_quantizer=None,
    activation_quantizer='default',
    block_size=128
)

Args
`weight_quantizer`	Quantizer for weights.
`activation_quantizer`	Quantizer for activations. If "default", uses AbsMaxQuantizer with axis=-1.
`block_size`	Size of groups along the input dimension for sub-channel quantization. If a positive integer, uses sub-channel quantization with `ceil(input_dim / block_size)` groups. If `None` or `-1`, uses per-channel quantization (one scale per output channel). Default: `128` (sub-channel with 128-element groups).

Attributes
`mode`

Attributes

mode

@staticmethod
activation_quantizer_or_default(
    config, default
)

@classmethod
from_config(
    config
)

get_config()

@staticmethod
weight_quantizer_or_default(
    config, default
)

Except as otherwise noted, the content of this page is licensed under the Creative Commons Attribution 4.0 License, and code samples are licensed under the Apache 2.0 License. For details, see the Google Developers Site Policies. Java is a registered trademark of Oracle and/or its affiliates. Some content is licensed under the numpy license.

Last updated 2026-10-31 UTC.