tf.keras.quantizers.abs_max_quantize_grouped_with_zero_point

Quantizes a 2D tensor using grouped asymmetric quantization with zero point.

tf.keras.quantizers.abs_max_quantize_grouped_with_zero_point(
    inputs,
    block_size,
    value_range=(-8, 7),
    dtype='int8',
    epsilon=backend.epsilon(),
    to_numpy=False
)

Groups are formed along axis 0 (the input/contracting dimension). Each group of block_size rows gets its own scale factor and zero point per column. This is useful for weight distributions that are not centered around zero.

inputs Input tensor to quantize. Shape: (input_dim, output_dim).
block_size Number of elements per group along axis 0.
value_range Tuple of (min, max) quantization range.
dtype Data type of quantized output.
epsilon Small value to avoid division by zero.
to_numpy Whether to perform computation in numpy for memory efficiency.

A tuple (quantized_tensor, scale, zero_point) where:

  • quantized_tensor: Same shape as inputs, dtype=dtype.
  • scale: Shape (n_groups, output_dim) where n_groups = ceil(input_dim / block_size).
  • zero_point: Shape (n_groups, output_dim), dtype=uint8.

Example:

import numpy as np from keras.quantizers import abs_max_quantize_grouped_with_zero_point kernel = np.random.randn(512, 256).astype("float32") quantized, scale, zero_point = abs_max_quantize_grouped_with_zero_point( ... kernel, block_size=128, value_range=(-8, 7) ... ) quantized.shape (512, 256) scale.shape # 512 / 128 = 4 groups (4, 256) zero_point.shape (4, 256)