coreai_opt.quantization.spec.MovingAverageQParamsCalculator

class coreai_opt.quantization.spec.MovingAverageQParamsCalculator(dtype, qscheme, granularity, target_dtype, quant_min, quant_max, range_calculator, float_range, averaging_constant=0.01, **kwargs)

Bases: RunningRangeMixin, QParamsCalculatorBase

Computes the scale and zero point using a moving average of the range.

Maintains running_min and running_max buffers that are updated each forward pass using exponential moving average (EMA):

a_{i} = c * x_{i} + (1 - c) * a_{i-1}

where c is the averaging_constant.

Parameters:

• dtype (torch.dtype)

• qscheme (QuantizationScheme)

• granularity (QuantizationGranularity)

• target_dtype (torch.dtype)

• quant_min (int)

• quant_max (int)

• range_calculator (RangeCalculatorBase)

• float_range (list[float | None])

• averaging_constant (float)

__init__(dtype, qscheme, granularity, target_dtype, quant_min, quant_max, range_calculator, float_range, averaging_constant=0.01, **kwargs)

Parameters:

• dtype (dtype)

• qscheme (QuantizationScheme)

• granularity (QuantizationGranularity)

• target_dtype (dtype)

• quant_min (int)

• quant_max (int)

• range_calculator (RangeCalculatorBase)

• float_range (list[float | None])

• averaging_constant (float)

Methods

compute_qparams(tensor, min_val, max_val)	Update running range, persist to buffers, then compute qparams.
extra_repr()	Return the extra representation of the module.
forward(tensor)	Compute scale, zero point, and minval from the input tensor.
get_class(key)
get_qparams()	Return the computed scale, zero point and minval.
list_registry_keys()
list_registry_values()
register(key)	Register a virtual subclass of an ABC.
set_export_mode([enabled])
update_running_range(min_val, max_val)	Return (updated_min, updated_max) using subclass-specific rule.

update_running_range(min_val, max_val)

Return (updated_min, updated_max) using subclass-specific rule.

Parameters:

• min_val (Tensor)

• max_val (Tensor)

Return type:

tuple[Tensor, Tensor]