@benjitheblender https://t.co/1XoGsWt05s
Buy something like this, cut the PEX, put in new pipe, attach with the couplers, sinch with the provided straps. Will take less than 30 min
@dylan522p "but the key takeaway is those serving an unmodified model (FP16) required batch sizes of 64+ to turn a profit. We believe that Mistral, Together, and Fireworks are serving Mistral at breakeven to slight profit margins."
Im curious how they do this
@abacaj I haven't seen any finetuning harnesses do it correctly at least. If you look at the mistral implementation, the only lever you really have is the weights on the gate for routing. You then have to coordinate the updating of the appropriate expert in conjunction with the router.
@ericmitchellai@teknium This has been my experience as well. You are training for a preference objective that wasn't captured in your SFT training set. Ride the wave and perform evals every 5000 steps or so.