Max Balandat

Yeah so it's an indexing issue with lazy tensors then, we'll need to investigate this on the gpytorch end. If you change https://github.com/cornellius-gp/gpytorch/blob/45e560c3417cb970c3a402f8a1a92f87b733e470/gpytorch/settings.py#L409 you're just modifying the defaults. Using the...

reopening this until the issue is fixed upstream

These kinds of constraints are currently not supported out of the box. If you have a very high cardinality discrete space that is constrained by both equality and nonlinear inequality...

> So maybe it's not a good idea to leverage optimize_acqf_discrete for huge discrete design space. Could you please give me some suggestion on this situation? Yes, if you have...

Hmm interesting, this clearly seems to be due to the constraint being tight. We'll have to look in more detail what exactly is happening, but my hunch is that while...

cc @jelena-markovic

> due to numerical imprecisions the value can be a small negative number Hmm it seems that this should be handled by [this piece of code](https://github.com/pytorch/botorch/blob/main/botorch/utils/sampling.py#L358-L359) just below? > I...

> But this results in that alpha_min = alpha_max = 0, so basically the only option is to stay where we are in the polytope, producing all equal points. Is...

Not quite, I think - This improved some cases in which this can happen, but if the (non-equality) constraints are such that they indeed result in a zero-volume polytope then...

I looked into this some more. It seems like this is caused by the closures that are used for setting/getting parameters that have a prior registered from gpytorch modules don't...