nncf.common.quantization.structs#

Classes#

QuantizationScheme

Enumeration for specifying quantization schemes.
class nncf.common.quantization.structs.QuantizationScheme[source]#

Bases: nncf.parameters.StrEnum

Enumeration for specifying quantization schemes.

Parameters:

  • SYMMETRIC – Symmetric quantization where the range is defined by a single parameter - scale. This range can include both negative and positive values if signed, or only positive values if unsigned.

  • ASYMMETRIC – Asymmetric quantization where the range is defined by two parameters - input_low and input_high, representing the lower and upper boundaries of the range, respectively.

  • SYMMETRIC_LORA – Symmetric quantization with Low-Rank Adapters (LoRA), involving the sum of weights and the multiplication of low-rank adapters.

  • SYMMETRIC_LORA_NLS – Symmetric quantization with Low-Rank Adapters (LoRA) and Neural Low-Rank Adapter Search (NLS), involving the sum of weights and the multiplication of low-rank adapters.

  • ASYMMETRIC_LORA – Asymmetric quantization with Low-Rank Adapters (LoRA), involving the sum of weights and the multiplication of low-rank adapters.

  • ASYMMETRIC_LORA_NLS – Asymmetric quantization with Low-Rank Adapters (LoRA) and Neural Low-Rank Adapter Search (NLS), involving the sum of weights and the multiplication of low-rank adapters.