[query] is the int symmetric quantisation only for unsigned int?

[EricLiclair]

based on the implementations here - https://github.com/intel/auto-round/blob/9718d20333e4448b7cda96074ef481668d19c861/auto_round/data_type/int.py#L70

and my experiments at - https://colab.research.google.com/drive/1rjfaNl8B_9sQMMYXupk4DWtxwhR6LNpx?usp=sharing

it seems that the symmetric int quantization is intentionally only for unsigned int, am i correct on this? based on my understanding, the expected zero-point for symmetric int quantization is "0"; which in the current case is not so. currently, zp=8 if bits=4, and zp=128 if bits=8, etc. which reflects a case where the target data type could be unsigned int.

i wanted to know what int types (signed or unsigned) are generally preferred and recommend for inference time optimisations?

[wenhuach21]

Yes, you are correct, this is intentionally designed to align with GPTQ's logic, allowing us to leverage their CUDA kernel. Besides, the fixed zero point (zp) can be converted to zp=0 to utilize the Marlin kernel, which requires zp=0.

[wenhuach21]

feel free to reopen this if you have additional questions