Max Balandat

Yeah - though the fact that things break even if you don't do this but instead adjust the import as in https://github.com/pytorch/botorch/pull/2854#issuecomment-2897994321 is disconcerting - we should understand what causes...

Hmm yeah that would make sense. Let's take a look and see if that's the case when you're back; we may want to think about separating out the registries in...

@gpleiss, @corwinjoy - speak now or forever hold your peace.

> Now as a next step - would it make sense to get this new MMVN implemented for non-lazy tensors, and focus on adding LTs later down the road? I...

Good call @jacobrgardner. I was an early adopter of NamedTensors, but upsterma development on them has been somewhat de-prioritized, so I didn't do much with them. But I think it...

Also a lot of stuff is currently not supported with NamedTensors (e.g. permute), so while we can align the naming interface, we can't really do much with them until there...

Added a lazy version for the task-independent case. Not sure what level of generality we want / need to support for the lazies, at some point things become really clunky....

@gpleiss, @jacobrgardner, @dme65 I finally got back to this. To recall, I drafted a `MultioutputMultivariateNormal` distribution (subclasses `MultivariateNormal`, but for now completely separate from `MultitaskMultivariateNormal`) . This is an abstract...

@neerajprad you might be interested in this, too

@TobyBoyne thanks for raising this - We'd be interested in incorporating this; I believe that it would be best to land the change in gpytorch and then use that (optionally)...