Int8DynActInt4WeightQuantizer quantize() not propagating the scales_precision argument

[HaroldBenoit]

Hello,

In Int8DynActInt4WeightQuantizer, the scales_precision argument of replace_linear_8da4w is set to self.precision instead of self.scales_precision.

This introduces an inconsistency between calling _create_quantized_state_dict(model) and calling quantize(model).state_dict().

[danielvegamyhre]

@HaroldBenoit Thanks for reporting this!

cc @jerryzh168 who I think is the best PoC for this

[jerryzh168]

this is deprecated, please use quantize_(model, Int8DynamicActivationInt4Weight(...)) instead