IQ4_XS Please
#6
by
lingyezhixing
- opened
IQ4_XS may be more accurate than MXFP4_MOE under the same size or even smaller size
Great idea ! working on it
IQ4_XS is there now , it works best and the size is the smallest of the 4 bits quants.
Better, faster and smaller than MXFP4_MOE.
good call!
Using IQ4_XS , working very well thanks a lot Mr Fromage