get_quantizer_mode

torchtune.utils.get_quantizer_mode(quantizer: Optional[Callable]) → Optional[str]

Given a quantizer object, returns a string that specifies the type of quantization.

For example, in the case of int4 weight only quantization, we’ll return “4w”. If the quantizer is not recognized as a known quantizer, we’ll return None.

Currently supported:

• Int4WeightOnlyQuantizer: “4w”

• Int8WeightOnlyQuantizer: “8w”

• Int4WeightOnlyGPTQQuantizer: “4w-gptq”

• Int8DynActInt4WeightQuantizer: “8da4w” (requires torch>=2.3.0)

• Int8DynActInt4WeightQATQuantizer: “8da4w-qat” (requires torch>=2.4.0)

Parameters:

quantizer (Optional[Callable]) – A callable object that implements the quantize method.

Returns:

The quantization mode.

Return type:

Optional[str]