def __init__(self,
scaling_init: Union[float, Tensor],
restrict_scaling_impl: Module,
scaling_min_val: Optional[float] = None) -> None:
super(ConstScaling, self).__init__()
self.restrict_clamp_scaling = _RestrictClampValue(
scaling_min_val, restrict_scaling_impl)
if isinstance(scaling_init, Tensor):
scaling_init = restrict_scaling_impl.restrict_init_tensor(
scaling_init)
self.value = StatelessBuffer(scaling_init.detach())
else:
scaling_init = restrict_scaling_impl.restrict_init_float(
scaling_init)
self.value = StatelessBuffer(torch.tensor(scaling_init))
def __init__(self,
stats_reduce_dim: int,
sigma: float,
std_dev_epsilon: float = DEFAULT_STD_DEV_EPSILON) -> None:
super(MeanSigmaStd, self).__init__()
self.impl = _MeanSigmaStdImpl(stats_reduce_dim, std_dev_epsilon)
self.sigma = StatelessBuffer(torch.tensor(sigma))
def __init__(self,
scaling_impl: Module,
threshold: float,
quant_delay_steps: int = None):
super(TernaryQuant, self).__init__()
self.scaling_impl = scaling_impl
self.threshold = threshold
self.bit_width = BitWidthConst(2)
self.zero_point = StatelessBuffer(torch.tensor(0.0))
self.delay_wrapper = DelayWrapper(quant_delay_steps)
def __init__(
self,
scaling_impl: Module,
tensor_clamp_impl: Module = TensorClamp(),
quant_delay_steps: int = 0):
super(ClampedBinaryQuant, self).__init__()
self.scaling_impl = scaling_impl
self.bit_width = BitWidthConst(1)
self.zero_point = StatelessBuffer(torch.tensor(0.0))
self.delay_wrapper = DelayWrapper(quant_delay_steps)
self.tensor_clamp_impl = tensor_clamp_impl
def __init__(self, quant_layer: nn.Module,
quant_injector: Injector) -> None:
super(QuantProxyFromInjector, self).__init__()
self.is_signed = _is_signed(quant_injector)
self.is_narrow_range = _is_narrow_range(quant_injector)
self.update_state_dict_impl = _update_state_dict_impl(quant_injector)
self.quant_injector = quant_injector
self._zero_hw_sentinel = StatelessBuffer(tensor(0.0))
self.tensor_quant = None
# Use a normal list and not a ModuleList since this is a pointer to parent modules
self.tracked_module_list = []
self.add_tracked_module(quant_layer)
def __init__(self,
negative_slope,
alpha_quant=qnn.QuantIdentity(Uint8ActPerTensorFixedPoint,
bit_width=16),
input_quant=qnn.QuantIdentity(Int8ActPerTensorFixedPoint,
bit_width=16,
scaling_stats_momentum=None),
output_quant=qnn.QuantIdentity(Int8ActPerTensorFixedPoint,
return_quant_tensor=True)):
super(FlexMLQuantLeakyReLU, self).__init__()
self.alpha_quant = alpha_quant
self.input_quant = input_quant
self.output_quant = output_quant
self.negative_slope = StatelessBuffer(torch.tensor(negative_slope))
def __init__(self,
quant_layer: nn.Module,
quant_injector: Injector,
export_mode: bool = False,
export_handler: Optional[nn.Module] = None) -> None:
super(QuantProxyFromInjector, self).__init__()
self.update_state_dict_impl = _update_state_dict_impl(quant_injector)
self.quant_injector = quant_injector
self._zero_hw_sentinel = StatelessBuffer(tensor(0.0))
self.tensor_quant = None
# Use a normal list and not a ModuleList since this is a pointer to parent modules
self.tracked_module_list = []
self.add_tracked_module(quant_layer)
self.export_handler = export_handler
self.export_mode = export_mode
self.export_debug_name = None
self.export_input_debug = False
self.export_output_debug = False
self.disable_quant = False
def __init__(self, stats_reduce_dim: Optional[int] = None) -> None:
super(NegativeMinOrZero, self).__init__()
self.stats_reduce_dim = stats_reduce_dim
self.zero = StatelessBuffer(torch.tensor(0.0))
def __init__(self):
super(QuantProxy, self).__init__()
self._zero_hw_sentinel = StatelessBuffer(tensor(0.0))
def __init__(self, bit_width: int) -> None:
super(BitWidthConst, self).__init__()
assert isinstance(bit_width, int)
self.bit_width = StatelessBuffer(torch.tensor(float(bit_width)))
def __init__(self):
super(IdentityQuant, self).__init__()
self.minus_one = StatelessBuffer(torch.tensor(-1))
def __init__(self, int_quant: Module, bit_width_impl: Module):
super(PrescaledRestrictIntQuant, self).__init__()
self.int_quant = int_quant
self.msb_clamp_bit_width_impl = bit_width_impl
self.int_scale = StatelessBuffer(torch.tensor(1.0))
def __init__(self, int_quant: Module, bit_width_impl: Module):
super(PrescaledRestrictIntQuantWithInputBitWidth, self).__init__()
self.int_quant = int_quant
self.msb_clamp_bit_width_impl = bit_width_impl
self.zero_point = StatelessBuffer(torch.tensor(0.0))
def __init__(self, parameter_stats_scaling: StatsFromParameterScaling):
self.parameter_stats_scaling = parameter_stats_scaling
self.ignored = StatelessBuffer(torch.tensor(0.0))
