Artifact weights

class torch_lattice.artifact.weights.PackedWeight(weight, scales, biases)

Bases: object

Affine group-quantized artifact storage tensors.

Parameters:

• weight (Tensor)

• scales (Tensor)

• biases (Tensor)

weight: Tensor

scales: Tensor

biases: Tensor

torch_lattice.artifact.weights.dequantize_artifact_weight(tensor, *, bits, group_size, scale_dtype='f16')

Round-trip a logical weight through artifact quantization.

Return type:

Tensor

Parameters:

• tensor (Tensor)

• bits (int)

• group_size (int)

• scale_dtype (str)

torch_lattice.artifact.weights.pack_quantized_weight(tensor, *, bits, group_size, scale_dtype)

Pack a logical artifact weight using affine group quantization.

Return type:

PackedWeight

Parameters:

• tensor (Tensor)

• bits (int)

• group_size (int)

• scale_dtype (str)