Why quarot algo R2 rotate needed online_rotate?

Open 54limiao opened this issue 6 months ago • 2 comments

Based on the original quarot method, the R2 rotate can be observed by weight, no need for online rotate. https://github.com/ModelTC/llmc/blob/867fb4f536073a2898048c39aa098979521a45a6/llmc/compression/quantization/quarot.py#L139

Jun 19 '25 07:06 54limiao

I implement the reshape R2 transform from spinquant code base, this issuse fixed

Jun 26 '25 08:06 54limiao

btw, there are missing qk online rotate which is R3

Aug 27 '25 06:08 54limiao