Question on wrapped forward quantized module

[Jonxjdong]

https://github.com/neuralmagic/compressed-tensors/blob/main/src/compressed_tensors/quantization/lifecycle/forward.py#L257
In the forward function above, the Linear operation with quantization is done in the way:
Y = (Sx*[X/Sx]) @ (Sw*[W/Sw])
which X and W is quantized and dequantized separately and then the two fp16-format matrices multiplied.

Why not in the way like
Y = ([X/Sx] @ [W/Sw]) * (Sx @ Sw)
which [X/Sx] and [W/Sw] are in int format that the integer multiplication would be faster in CUDA