Closed leiwen83 closed 1 month ago
Yeah, we only compared the accuracy to FP16 and the accuracy loss is already negligible. We did not further compare it to FP8 because FP8 would not be better than FP16.
For FP6, we can do post-training quantization. I think we do not need a calibration dataset. Please refer to https://arxiv.org/abs/2312.08583 for more algorithm level details.
And have fp6 compare the accuracy with fp8? I haven't find any related in the paper...