You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This is motivated mainly for operator decomposability purposes (came out of the wave 3 PR), to express quantizeLinear and dequantizeLinear generically rather than having a form of broadcasting that is unique to only those two operators, and it may be useful to express other future operators. This would move the blockwise broadcasting aspect into expand, leaving the rest of the decomposition being the respective mul/div/sub/add for Q and DQ. Alternately, we could extend resample with nearest neighbor to support multiple axes, which we may want to do anyway, but putting it on expand makes sense conceptually as it's always an integer upsampling (resample takes any ratio).