Quantization method diff between fake quant and true quant

quantization
1

according to the code. The quantization formula for true quant is

int32_t res = round_half_even(value / scale) + zero_point;

and according to the document, the formula for fake quant is

int32_t res = round_half_even(value / scale + zero_point);

image
image950×405 25.3 KB

Why is there a difference between these two?

2

probably the document is a bit outdated, we are unifying the implementation in ao/torchao/quantization/quant_primitives.py at 04259ebd96a8f1d29db16a7fe73a50d1f21bddef · pytorch/ao · GitHub