upvote
I you look at https://arxiv.org/pdf/2401.04088 table 5 on page 8, you'll see that expert(s) used can change from token to token. The experts aren't divided along predictable lines.
reply
For purely coding tasks, is every single expert required? 50%? 25%?
reply
Yes, every expert. It's not like different personas, more like very coarse dropout training.
reply
Right. So it's not like human experts or even different brain regions
reply
I have a similar question and I’m inferring the answer is no - look at the cache hit rate of 23% for the 128GB M5 Max. I had previously assumed that the 40B active meant that a set of layers was chosen as THE expert for a given prompt and generation was then limited to those layers until complete. But in that case you’d have expected the expert caching to have a super high hit rate once you had enough RAM to hold an entire expert’s worth of layers.
reply
Or could you parallelise your experts on different hardware?
reply
You could (e.g. by replacing residual-dependent expert routing with hardcoded logic), but quality will suffer dramatically. It’s far better to use a similar-sized dense model then.
reply
Ahh, I remember hearing that before, and it makes general sense too.
reply