4 ms·
Btw, shouldn't it in theory be possible to run the Mixtral MoE loading next submodel sequentially and store outputs and then do the rest of the algorithm to mak
by TotalCrackpot 3y ago
Btw, shouldn't it in theory be possible to run the Mixtral MoE loading next submodel sequentially and store outputs and then do the rest of the algorithm to make it easier to run on machines that cannot fit whole model in the memory?
- wfhpw 3y agoYes but loading weights into memory takes time
- TotalCrackpot 3y agoYeah I imagine sequential inference would be slower. How long do you have to wait to load these weights on a personal PC? I have not tried using those systems so far.