Quantized feature modules

Quantized feature modules store packed affine weights for sparse-feature projections. Activations remain floating point and coordinate identity is preserved.

Related pages

• Packed weight container: Quantized weights

• Functional feature API: Feature operations

• Quantized backend routes: Quantization routes

• Floating feature modules: Feature modules

class mlx_lattice.nn.quantized_feature.QuantizedLinear(input_dims, output_dims, bias=True, group_size=None, bits=4, mode='affine')

Bases: Module

Affine weight-quantized sparse-feature linear module.

The module stores packed int4/int8 weight, scale, and bias arrays and uses floating-point sparse activations. Calling the module preserves coordinate identity and projects x.feats to output_dims channels.

Parameters:

• input_dims (int)

• output_dims (int)

• bias (bool)

• group_size (int | None)

• bits (int)

• mode (str)

classmethod from_linear(linear, group_size=None, bits=4, mode='affine')

Return type:

QuantizedLinear

Parameters:

• linear (Module)

• group_size (int | None)

• bits (int)

• mode (str)