[vessenes]

I'll pipe in - a series of Mac optimized MOEs which can stream experts just in time would be really amazing. And popular; I'm guessing in the next year we'll be able to run a very able openclaw with a stack like that. You'll get a lot of installs there. If I were a PM at Gemma, I'd release a stack for each Mac mini memory size.

[zozbot234]

Expert streaming is something that has to be implemented by the inference engine/library, the model architecture itself has very little to do with it. It's a great idea (for local inference; it uses too much power at scale), but making it work really well is actually not that easy.

(I've mentioned this before but AIUI it would require some new feature definitions in GGUF, to allow for coalescing model data about any one expert-layer into a single extent, so that it can be accessed in bulk. That's what seems to make the new Flash-MoE work so well.)

[vessenes]

I’ve been doing some low-key testing on smaller models, and it looks to me like it’s possible to train an MOE model with characteristics that are helpful for streaming… For instance, you could add a loss function to penalize expert swapping both in a single forward, pass and across multiple forward passes. So I believe there is a place for thinking about this on the model training side.

[zozbot234]

Penalizing expert swaps doesn't seem like it would help much, because experts vary by layer and are picked layer-wise. There's no guarantee that expert X in layer Y that was used for the previous token will still be available for this token's load from layer Y. The optimum would vary depending on how much memory you have at any given moment, and such. It's not obviously worth optimizing for.

[vessenes]

Right. You need to predict a set of experts through the entire forward pass. Think of a vertical strip.