I like this idea on expert streaming. I've been poking around fairly thoroughly at the same idea - can we fix a set of experts? when can we fix them? How long is the top-k selection "good" for in terms of number of forward passes?

One thing I've turned up in smaller models and I'm sort of winding my way toward verifying in larger ones is that if you train the MoE model from scratch with this kind of knockout / subset of experts baked in, then you get significantly better loss outcomes. In small models, it's actually better than training an MOE without conditioning on a reduced set of experts per pass.

Anyway, pretty cool. There's some Pareto-optimal curve based on memory bandwidth, amount of GPU / unified RAM and inference compute times for streaming stuff in.

[flagged]