p2pfl.learning.compression.quantization_strategy module¶

Post-Training Quantization (PTQ) compression strategy.

class p2pfl.learning.compression.quantization_strategy.PTQuantization[source]¶

Post-Training Quantization (PTQ) with proper scaling.

apply_strategy(params, dtype='float16', scheme='symmetric', granularity='per_tensor', channel_axis=0)[source]¶

Reduce the precision of model parameters with proper scaling.

Parameters:

params (list[ndarray]) – The parameters to compress.
dtype (str) – The desired precision (e.g., “float16”, “int8”).
scheme (Literal['symmetric', 'asymmetric']) – Quantization scheme - “symmetric” (centered around 0) or “asymmetric” (uses full range).
granularity (Literal['per_tensor', 'per_channel']) – “per_tensor” uses one scale for the whole tensor, “per_channel” uses separate scales for each channel.
channel_axis (int) – Axis to use for per-channel quantization.

Return type:

tuple[list[ndarray], dict]

Returns:

Tuple of quantized parameters and additional info for dequantization.

Raises:

ValueError – If an unsupported data type is provided or if parameters are invalid.

reverse_strategy(params, additional_info)[source]¶

Return model parameters to saved original precision.

Parameters:

Return type:

list[ndarray]

Returns:

Decompressed parameters.

Raises:

ValueError – If the parameters or additional info are invalid.